diff --git "a/axquant_plan.json" "b/axquant_plan.json"
new file mode 100644--- /dev/null
+++ "b/axquant_plan.json"
@@ -0,0 +1,81199 @@
+{
+  "analysis_sha256": "6668d94bd5a93dee2be86c88adec206a817b596cbca9562e42dfd35566bb578c",
+  "architecture_profile": {
+    "adapter_id": "qwen35-moe-v1",
+    "audio_present": false,
+    "config_model_type": "qwen3_5_moe",
+    "dense": false,
+    "mtp_declared": true,
+    "notes": [
+      "AXQuant optimizes the Qwen3.5-class MoE language path only (development convert; not the Qwen 3.6 certification track).",
+      "Vision tensors are preserved at BF16 and VLM quality is not claimed.",
+      "Validated sizes: 35B-A3B (Ornith-1.0-35B, Ornith-1.5-35B-A3B, Holo3-35B-A3B, Holo-3.1-35B-A3B) and 397B (Ornith-1.5-397B).",
+      "MoE experts quantize as fused switch modules with uniform per-group precision; label packs as development evidence until certified."
+    ],
+    "optimization_scope": "text-path",
+    "product_family": "qwen3.5-moe",
+    "support_level": "supported",
+    "support_tier": "convertible",
+    "text_layer_count": 40,
+    "vision_present": true
+  },
+  "assignments": [
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.lm_head.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-lm-head: Protection floor - LM head stays BF16",
+      "role": "lm_head",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.lm_head.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.lm_head.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-lm-head: Protection floor - LM head stays BF16",
+      "role": "lm_head",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.lm_head.scales"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.lm_head",
+      "outlier_strategy": "none",
+      "parameters": 508559360,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-lm-head: Protection floor - LM head stays BF16",
+      "role": "lm_head",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.lm_head.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.embed_tokens.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "embedding",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.embed_tokens.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.embed_tokens.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "embedding",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.embed_tokens.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.embed_tokens",
+      "outlier_strategy": "none",
+      "parameters": 508559360,
+      "predicted_loss": 0.0,
+      "reason": "manual rule embedding-8: Embedding floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "embedding",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.embed_tokens.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.0.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.linear_attn.A_log",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.0.linear_attn.A_log"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.linear_attn.conv1d",
+      "outlier_strategy": "none",
+      "parameters": 32768,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.0.linear_attn.conv1d.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.linear_attn.dt_bias",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.0.linear_attn.dt_bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.linear_attn.in_proj_a.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.0.linear_attn.in_proj_a.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.linear_attn.in_proj_a.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.0.linear_attn.in_proj_a.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.linear_attn.in_proj_a",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.0.linear_attn.in_proj_a.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.linear_attn.in_proj_b.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.0.linear_attn.in_proj_b.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.linear_attn.in_proj_b.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.0.linear_attn.in_proj_b.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.linear_attn.in_proj_b",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.0.linear_attn.in_proj_b.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.linear_attn.in_proj_qkv.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.0.linear_attn.in_proj_qkv.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.linear_attn.in_proj_qkv.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.0.linear_attn.in_proj_qkv.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.linear_attn.in_proj_qkv",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.0.linear_attn.in_proj_qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.linear_attn.in_proj_z.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.0.linear_attn.in_proj_z.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.linear_attn.in_proj_z.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.0.linear_attn.in_proj_z.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.linear_attn.in_proj_z",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.0.linear_attn.in_proj_z.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.linear_attn.norm",
+      "outlier_strategy": "none",
+      "parameters": 128,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.0.linear_attn.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.linear_attn.out_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.0.linear_attn.out_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.linear_attn.out_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.0.linear_attn.out_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.linear_attn.out_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.0.linear_attn.out_proj.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.0.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.0.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.0.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.0.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.0.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.0.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.0.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.0.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.0.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.0.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.0.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.0.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.0.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.0.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.0.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.0.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.0.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.0.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.0.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.0.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.0.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.0.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.0.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.0.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.1.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.linear_attn.A_log",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.1.linear_attn.A_log"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.linear_attn.conv1d",
+      "outlier_strategy": "none",
+      "parameters": 32768,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.1.linear_attn.conv1d.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.linear_attn.dt_bias",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.1.linear_attn.dt_bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.linear_attn.in_proj_a.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.1.linear_attn.in_proj_a.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.linear_attn.in_proj_a.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.1.linear_attn.in_proj_a.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.linear_attn.in_proj_a",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.1.linear_attn.in_proj_a.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.linear_attn.in_proj_b.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.1.linear_attn.in_proj_b.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.linear_attn.in_proj_b.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.1.linear_attn.in_proj_b.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.linear_attn.in_proj_b",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.1.linear_attn.in_proj_b.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.linear_attn.in_proj_qkv.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.1.linear_attn.in_proj_qkv.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.linear_attn.in_proj_qkv.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.1.linear_attn.in_proj_qkv.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.linear_attn.in_proj_qkv",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.1.linear_attn.in_proj_qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.linear_attn.in_proj_z.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.1.linear_attn.in_proj_z.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.linear_attn.in_proj_z.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.1.linear_attn.in_proj_z.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.linear_attn.in_proj_z",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.1.linear_attn.in_proj_z.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.linear_attn.norm",
+      "outlier_strategy": "none",
+      "parameters": 128,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.1.linear_attn.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.linear_attn.out_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.1.linear_attn.out_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.linear_attn.out_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.1.linear_attn.out_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.linear_attn.out_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.1.linear_attn.out_proj.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.1.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.1.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.1.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.1.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.1.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.1.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.1.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.1.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.1.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.1.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.1.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.1.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.1.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.1.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.1.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.1.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.1.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.1.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.1.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.1.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.1.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.1.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.1.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.1.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.10.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.linear_attn.A_log",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.10.linear_attn.A_log"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.linear_attn.conv1d",
+      "outlier_strategy": "none",
+      "parameters": 32768,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.10.linear_attn.conv1d.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.linear_attn.dt_bias",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.10.linear_attn.dt_bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.linear_attn.in_proj_a.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.10.linear_attn.in_proj_a.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.linear_attn.in_proj_a.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.10.linear_attn.in_proj_a.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.linear_attn.in_proj_a",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.10.linear_attn.in_proj_a.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.linear_attn.in_proj_b.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.10.linear_attn.in_proj_b.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.linear_attn.in_proj_b.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.10.linear_attn.in_proj_b.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.linear_attn.in_proj_b",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.10.linear_attn.in_proj_b.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.linear_attn.in_proj_qkv.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.10.linear_attn.in_proj_qkv.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.linear_attn.in_proj_qkv.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.10.linear_attn.in_proj_qkv.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.linear_attn.in_proj_qkv",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.10.linear_attn.in_proj_qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.linear_attn.in_proj_z.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.10.linear_attn.in_proj_z.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.linear_attn.in_proj_z.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.10.linear_attn.in_proj_z.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.linear_attn.in_proj_z",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.10.linear_attn.in_proj_z.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.linear_attn.norm",
+      "outlier_strategy": "none",
+      "parameters": 128,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.10.linear_attn.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.linear_attn.out_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.10.linear_attn.out_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.linear_attn.out_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.10.linear_attn.out_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.linear_attn.out_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.10.linear_attn.out_proj.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.10.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.10.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.10.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.10.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.10.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.10.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.10.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.10.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.10.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.10.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.10.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.10.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.10.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.10.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.11.input_layernorm.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.11.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.11.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.11.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.11.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.11.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.11.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.11.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.11.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.11.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.11.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.11.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.11.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.11.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.11.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.self_attn.k_norm",
+      "outlier_strategy": "none",
+      "parameters": 256,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.11.self_attn.k_norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.self_attn.k_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.11.self_attn.k_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.self_attn.k_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.11.self_attn.k_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.self_attn.k_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.11.self_attn.k_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.self_attn.o_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.11.self_attn.o_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.self_attn.o_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.11.self_attn.o_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.self_attn.o_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.11.self_attn.o_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.self_attn.q_norm",
+      "outlier_strategy": "none",
+      "parameters": 256,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.11.self_attn.q_norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.self_attn.q_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.11.self_attn.q_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.self_attn.q_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.11.self_attn.q_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.self_attn.q_proj",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.11.self_attn.q_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.self_attn.v_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.11.self_attn.v_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.self_attn.v_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.11.self_attn.v_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.self_attn.v_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.11.self_attn.v_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.12.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.linear_attn.A_log",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.12.linear_attn.A_log"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.linear_attn.conv1d",
+      "outlier_strategy": "none",
+      "parameters": 32768,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.12.linear_attn.conv1d.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.linear_attn.dt_bias",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.12.linear_attn.dt_bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.linear_attn.in_proj_a.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.12.linear_attn.in_proj_a.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.linear_attn.in_proj_a.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.12.linear_attn.in_proj_a.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.linear_attn.in_proj_a",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.12.linear_attn.in_proj_a.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.linear_attn.in_proj_b.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.12.linear_attn.in_proj_b.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.linear_attn.in_proj_b.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.12.linear_attn.in_proj_b.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.linear_attn.in_proj_b",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.12.linear_attn.in_proj_b.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.linear_attn.in_proj_qkv.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.12.linear_attn.in_proj_qkv.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.linear_attn.in_proj_qkv.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.12.linear_attn.in_proj_qkv.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.linear_attn.in_proj_qkv",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.12.linear_attn.in_proj_qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.linear_attn.in_proj_z.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.12.linear_attn.in_proj_z.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.linear_attn.in_proj_z.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.12.linear_attn.in_proj_z.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.linear_attn.in_proj_z",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.12.linear_attn.in_proj_z.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.linear_attn.norm",
+      "outlier_strategy": "none",
+      "parameters": 128,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.12.linear_attn.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.linear_attn.out_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.12.linear_attn.out_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.linear_attn.out_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.12.linear_attn.out_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.linear_attn.out_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.12.linear_attn.out_proj.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.12.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.12.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.12.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.12.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.12.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.12.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.12.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.12.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.12.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.12.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.12.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.12.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.12.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.12.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.13.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.linear_attn.A_log",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.13.linear_attn.A_log"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.linear_attn.conv1d",
+      "outlier_strategy": "none",
+      "parameters": 32768,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.13.linear_attn.conv1d.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.linear_attn.dt_bias",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.13.linear_attn.dt_bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.linear_attn.in_proj_a.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.13.linear_attn.in_proj_a.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.linear_attn.in_proj_a.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.13.linear_attn.in_proj_a.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.linear_attn.in_proj_a",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.13.linear_attn.in_proj_a.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.linear_attn.in_proj_b.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.13.linear_attn.in_proj_b.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.linear_attn.in_proj_b.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.13.linear_attn.in_proj_b.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.linear_attn.in_proj_b",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.13.linear_attn.in_proj_b.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.linear_attn.in_proj_qkv.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.13.linear_attn.in_proj_qkv.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.linear_attn.in_proj_qkv.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.13.linear_attn.in_proj_qkv.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.linear_attn.in_proj_qkv",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.13.linear_attn.in_proj_qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.linear_attn.in_proj_z.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.13.linear_attn.in_proj_z.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.linear_attn.in_proj_z.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.13.linear_attn.in_proj_z.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.linear_attn.in_proj_z",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.13.linear_attn.in_proj_z.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.linear_attn.norm",
+      "outlier_strategy": "none",
+      "parameters": 128,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.13.linear_attn.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.linear_attn.out_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.13.linear_attn.out_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.linear_attn.out_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.13.linear_attn.out_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.linear_attn.out_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.13.linear_attn.out_proj.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.13.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.13.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.13.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.13.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.13.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.13.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.13.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.13.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.13.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.13.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.13.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.13.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.13.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.13.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.14.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.linear_attn.A_log",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.14.linear_attn.A_log"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.linear_attn.conv1d",
+      "outlier_strategy": "none",
+      "parameters": 32768,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.14.linear_attn.conv1d.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.linear_attn.dt_bias",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.14.linear_attn.dt_bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.linear_attn.in_proj_a.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.14.linear_attn.in_proj_a.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.linear_attn.in_proj_a.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.14.linear_attn.in_proj_a.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.linear_attn.in_proj_a",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.14.linear_attn.in_proj_a.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.linear_attn.in_proj_b.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.14.linear_attn.in_proj_b.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.linear_attn.in_proj_b.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.14.linear_attn.in_proj_b.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.linear_attn.in_proj_b",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.14.linear_attn.in_proj_b.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.linear_attn.in_proj_qkv.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.14.linear_attn.in_proj_qkv.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.linear_attn.in_proj_qkv.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.14.linear_attn.in_proj_qkv.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.linear_attn.in_proj_qkv",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.14.linear_attn.in_proj_qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.linear_attn.in_proj_z.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.14.linear_attn.in_proj_z.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.linear_attn.in_proj_z.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.14.linear_attn.in_proj_z.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.linear_attn.in_proj_z",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.14.linear_attn.in_proj_z.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.linear_attn.norm",
+      "outlier_strategy": "none",
+      "parameters": 128,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.14.linear_attn.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.linear_attn.out_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.14.linear_attn.out_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.linear_attn.out_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.14.linear_attn.out_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.linear_attn.out_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.14.linear_attn.out_proj.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.14.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.14.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.14.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.14.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.14.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.14.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.14.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.14.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.14.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.14.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.14.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.14.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.14.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.14.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.15.input_layernorm.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.15.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.15.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.15.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.15.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.15.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.15.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.15.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.15.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.15.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.15.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.15.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.15.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.15.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.15.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.self_attn.k_norm",
+      "outlier_strategy": "none",
+      "parameters": 256,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.15.self_attn.k_norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.self_attn.k_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.15.self_attn.k_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.self_attn.k_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.15.self_attn.k_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.self_attn.k_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.15.self_attn.k_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.self_attn.o_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.15.self_attn.o_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.self_attn.o_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.15.self_attn.o_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.self_attn.o_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.15.self_attn.o_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.self_attn.q_norm",
+      "outlier_strategy": "none",
+      "parameters": 256,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.15.self_attn.q_norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.self_attn.q_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.15.self_attn.q_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.self_attn.q_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.15.self_attn.q_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.self_attn.q_proj",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.15.self_attn.q_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.self_attn.v_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.15.self_attn.v_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.self_attn.v_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.15.self_attn.v_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.self_attn.v_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.15.self_attn.v_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.16.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.linear_attn.A_log",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.16.linear_attn.A_log"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.linear_attn.conv1d",
+      "outlier_strategy": "none",
+      "parameters": 32768,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.16.linear_attn.conv1d.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.linear_attn.dt_bias",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.16.linear_attn.dt_bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.linear_attn.in_proj_a.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.16.linear_attn.in_proj_a.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.linear_attn.in_proj_a.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.16.linear_attn.in_proj_a.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.linear_attn.in_proj_a",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.16.linear_attn.in_proj_a.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.linear_attn.in_proj_b.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.16.linear_attn.in_proj_b.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.linear_attn.in_proj_b.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.16.linear_attn.in_proj_b.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.linear_attn.in_proj_b",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.16.linear_attn.in_proj_b.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.linear_attn.in_proj_qkv.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.16.linear_attn.in_proj_qkv.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.linear_attn.in_proj_qkv.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.16.linear_attn.in_proj_qkv.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.linear_attn.in_proj_qkv",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.16.linear_attn.in_proj_qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.linear_attn.in_proj_z.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.16.linear_attn.in_proj_z.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.linear_attn.in_proj_z.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.16.linear_attn.in_proj_z.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.linear_attn.in_proj_z",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.16.linear_attn.in_proj_z.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.linear_attn.norm",
+      "outlier_strategy": "none",
+      "parameters": 128,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.16.linear_attn.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.linear_attn.out_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.16.linear_attn.out_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.linear_attn.out_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.16.linear_attn.out_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.linear_attn.out_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.16.linear_attn.out_proj.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.16.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.16.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.16.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.16.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.16.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.16.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.16.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.16.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.16.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.16.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.16.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.16.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.16.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.16.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.17.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.linear_attn.A_log",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.17.linear_attn.A_log"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.linear_attn.conv1d",
+      "outlier_strategy": "none",
+      "parameters": 32768,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.17.linear_attn.conv1d.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.linear_attn.dt_bias",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.17.linear_attn.dt_bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.linear_attn.in_proj_a.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.17.linear_attn.in_proj_a.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.linear_attn.in_proj_a.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.17.linear_attn.in_proj_a.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.linear_attn.in_proj_a",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.17.linear_attn.in_proj_a.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.linear_attn.in_proj_b.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.17.linear_attn.in_proj_b.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.linear_attn.in_proj_b.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.17.linear_attn.in_proj_b.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.linear_attn.in_proj_b",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.17.linear_attn.in_proj_b.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.linear_attn.in_proj_qkv.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.17.linear_attn.in_proj_qkv.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.linear_attn.in_proj_qkv.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.17.linear_attn.in_proj_qkv.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.linear_attn.in_proj_qkv",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.17.linear_attn.in_proj_qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.linear_attn.in_proj_z.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.17.linear_attn.in_proj_z.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.linear_attn.in_proj_z.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.17.linear_attn.in_proj_z.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.linear_attn.in_proj_z",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.17.linear_attn.in_proj_z.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.linear_attn.norm",
+      "outlier_strategy": "none",
+      "parameters": 128,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.17.linear_attn.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.linear_attn.out_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.17.linear_attn.out_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.linear_attn.out_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.17.linear_attn.out_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.linear_attn.out_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.17.linear_attn.out_proj.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.17.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.17.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.17.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.17.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.17.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.17.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.17.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.17.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.17.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.17.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.17.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.17.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.17.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.17.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.18.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.linear_attn.A_log",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.18.linear_attn.A_log"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.linear_attn.conv1d",
+      "outlier_strategy": "none",
+      "parameters": 32768,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.18.linear_attn.conv1d.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.linear_attn.dt_bias",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.18.linear_attn.dt_bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.linear_attn.in_proj_a.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.18.linear_attn.in_proj_a.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.linear_attn.in_proj_a.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.18.linear_attn.in_proj_a.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.linear_attn.in_proj_a",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.18.linear_attn.in_proj_a.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.linear_attn.in_proj_b.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.18.linear_attn.in_proj_b.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.linear_attn.in_proj_b.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.18.linear_attn.in_proj_b.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.linear_attn.in_proj_b",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.18.linear_attn.in_proj_b.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.linear_attn.in_proj_qkv.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.18.linear_attn.in_proj_qkv.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.linear_attn.in_proj_qkv.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.18.linear_attn.in_proj_qkv.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.linear_attn.in_proj_qkv",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.18.linear_attn.in_proj_qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.linear_attn.in_proj_z.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.18.linear_attn.in_proj_z.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.linear_attn.in_proj_z.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.18.linear_attn.in_proj_z.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.linear_attn.in_proj_z",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.18.linear_attn.in_proj_z.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.linear_attn.norm",
+      "outlier_strategy": "none",
+      "parameters": 128,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.18.linear_attn.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.linear_attn.out_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.18.linear_attn.out_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.linear_attn.out_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.18.linear_attn.out_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.linear_attn.out_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.18.linear_attn.out_proj.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.18.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.18.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.18.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.18.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.18.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.18.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.18.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.18.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.18.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.18.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.18.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.18.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.18.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.18.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.19.input_layernorm.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.19.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.19.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.19.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.19.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.19.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.19.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.19.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.19.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.19.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.19.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.19.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.19.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.19.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.19.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.self_attn.k_norm",
+      "outlier_strategy": "none",
+      "parameters": 256,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.19.self_attn.k_norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.self_attn.k_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.19.self_attn.k_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.self_attn.k_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.19.self_attn.k_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.self_attn.k_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.19.self_attn.k_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.self_attn.o_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.19.self_attn.o_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.self_attn.o_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.19.self_attn.o_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.self_attn.o_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.19.self_attn.o_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.self_attn.q_norm",
+      "outlier_strategy": "none",
+      "parameters": 256,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.19.self_attn.q_norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.self_attn.q_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.19.self_attn.q_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.self_attn.q_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.19.self_attn.q_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.self_attn.q_proj",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.19.self_attn.q_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.self_attn.v_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.19.self_attn.v_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.self_attn.v_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.19.self_attn.v_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.self_attn.v_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.19.self_attn.v_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.2.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.linear_attn.A_log",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.2.linear_attn.A_log"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.linear_attn.conv1d",
+      "outlier_strategy": "none",
+      "parameters": 32768,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.2.linear_attn.conv1d.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.linear_attn.dt_bias",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.2.linear_attn.dt_bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.linear_attn.in_proj_a.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.2.linear_attn.in_proj_a.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.linear_attn.in_proj_a.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.2.linear_attn.in_proj_a.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.linear_attn.in_proj_a",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.2.linear_attn.in_proj_a.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.linear_attn.in_proj_b.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.2.linear_attn.in_proj_b.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.linear_attn.in_proj_b.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.2.linear_attn.in_proj_b.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.linear_attn.in_proj_b",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.2.linear_attn.in_proj_b.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.linear_attn.in_proj_qkv.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.2.linear_attn.in_proj_qkv.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.linear_attn.in_proj_qkv.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.2.linear_attn.in_proj_qkv.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.linear_attn.in_proj_qkv",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.2.linear_attn.in_proj_qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.linear_attn.in_proj_z.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.2.linear_attn.in_proj_z.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.linear_attn.in_proj_z.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.2.linear_attn.in_proj_z.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.linear_attn.in_proj_z",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.2.linear_attn.in_proj_z.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.linear_attn.norm",
+      "outlier_strategy": "none",
+      "parameters": 128,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.2.linear_attn.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.linear_attn.out_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.2.linear_attn.out_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.linear_attn.out_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.2.linear_attn.out_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.linear_attn.out_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.2.linear_attn.out_proj.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.2.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.2.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.2.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.2.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.2.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.2.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.2.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.2.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.2.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.2.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.2.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.2.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.2.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.2.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.2.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.2.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.2.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.2.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.2.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.2.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.20.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.linear_attn.A_log",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.20.linear_attn.A_log"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.linear_attn.conv1d",
+      "outlier_strategy": "none",
+      "parameters": 32768,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.20.linear_attn.conv1d.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.linear_attn.dt_bias",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.20.linear_attn.dt_bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.linear_attn.in_proj_a.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.20.linear_attn.in_proj_a.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.linear_attn.in_proj_a.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.20.linear_attn.in_proj_a.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.linear_attn.in_proj_a",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.20.linear_attn.in_proj_a.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.linear_attn.in_proj_b.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.20.linear_attn.in_proj_b.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.linear_attn.in_proj_b.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.20.linear_attn.in_proj_b.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.linear_attn.in_proj_b",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.20.linear_attn.in_proj_b.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.linear_attn.in_proj_qkv.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.20.linear_attn.in_proj_qkv.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.linear_attn.in_proj_qkv.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.20.linear_attn.in_proj_qkv.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.linear_attn.in_proj_qkv",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.20.linear_attn.in_proj_qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.linear_attn.in_proj_z.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.20.linear_attn.in_proj_z.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.linear_attn.in_proj_z.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.20.linear_attn.in_proj_z.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.linear_attn.in_proj_z",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.20.linear_attn.in_proj_z.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.linear_attn.norm",
+      "outlier_strategy": "none",
+      "parameters": 128,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.20.linear_attn.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.linear_attn.out_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.20.linear_attn.out_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.linear_attn.out_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.20.linear_attn.out_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.linear_attn.out_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.20.linear_attn.out_proj.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.20.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.20.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.20.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.20.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.20.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.20.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.20.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.20.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.20.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.20.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.20.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.20.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.20.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.20.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.21.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.linear_attn.A_log",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.21.linear_attn.A_log"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.linear_attn.conv1d",
+      "outlier_strategy": "none",
+      "parameters": 32768,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.21.linear_attn.conv1d.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.linear_attn.dt_bias",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.21.linear_attn.dt_bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.linear_attn.in_proj_a.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.21.linear_attn.in_proj_a.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.linear_attn.in_proj_a.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.21.linear_attn.in_proj_a.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.linear_attn.in_proj_a",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.21.linear_attn.in_proj_a.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.linear_attn.in_proj_b.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.21.linear_attn.in_proj_b.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.linear_attn.in_proj_b.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.21.linear_attn.in_proj_b.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.linear_attn.in_proj_b",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.21.linear_attn.in_proj_b.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.linear_attn.in_proj_qkv.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.21.linear_attn.in_proj_qkv.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.linear_attn.in_proj_qkv.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.21.linear_attn.in_proj_qkv.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.linear_attn.in_proj_qkv",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.21.linear_attn.in_proj_qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.linear_attn.in_proj_z.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.21.linear_attn.in_proj_z.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.linear_attn.in_proj_z.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.21.linear_attn.in_proj_z.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.linear_attn.in_proj_z",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.21.linear_attn.in_proj_z.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.linear_attn.norm",
+      "outlier_strategy": "none",
+      "parameters": 128,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.21.linear_attn.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.linear_attn.out_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.21.linear_attn.out_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.linear_attn.out_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.21.linear_attn.out_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.linear_attn.out_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.21.linear_attn.out_proj.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.21.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.21.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.21.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.21.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.21.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.21.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.21.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.21.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.21.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.21.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.21.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.21.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.21.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.21.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.22.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.linear_attn.A_log",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.22.linear_attn.A_log"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.linear_attn.conv1d",
+      "outlier_strategy": "none",
+      "parameters": 32768,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.22.linear_attn.conv1d.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.linear_attn.dt_bias",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.22.linear_attn.dt_bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.linear_attn.in_proj_a.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.22.linear_attn.in_proj_a.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.linear_attn.in_proj_a.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.22.linear_attn.in_proj_a.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.linear_attn.in_proj_a",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.22.linear_attn.in_proj_a.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.linear_attn.in_proj_b.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.22.linear_attn.in_proj_b.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.linear_attn.in_proj_b.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.22.linear_attn.in_proj_b.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.linear_attn.in_proj_b",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.22.linear_attn.in_proj_b.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.linear_attn.in_proj_qkv.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.22.linear_attn.in_proj_qkv.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.linear_attn.in_proj_qkv.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.22.linear_attn.in_proj_qkv.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.linear_attn.in_proj_qkv",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.22.linear_attn.in_proj_qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.linear_attn.in_proj_z.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.22.linear_attn.in_proj_z.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.linear_attn.in_proj_z.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.22.linear_attn.in_proj_z.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.linear_attn.in_proj_z",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.22.linear_attn.in_proj_z.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.linear_attn.norm",
+      "outlier_strategy": "none",
+      "parameters": 128,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.22.linear_attn.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.linear_attn.out_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.22.linear_attn.out_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.linear_attn.out_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.22.linear_attn.out_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.linear_attn.out_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.22.linear_attn.out_proj.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.22.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.22.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.22.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.22.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.22.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.22.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.22.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.22.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.22.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.22.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.22.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.22.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.22.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.22.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.23.input_layernorm.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.23.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.23.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.23.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.23.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.23.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.23.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.23.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.23.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.23.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.23.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.23.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.23.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.23.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.23.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.self_attn.k_norm",
+      "outlier_strategy": "none",
+      "parameters": 256,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.23.self_attn.k_norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.self_attn.k_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.23.self_attn.k_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.self_attn.k_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.23.self_attn.k_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.self_attn.k_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.23.self_attn.k_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.self_attn.o_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.23.self_attn.o_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.self_attn.o_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.23.self_attn.o_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.self_attn.o_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.23.self_attn.o_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.self_attn.q_norm",
+      "outlier_strategy": "none",
+      "parameters": 256,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.23.self_attn.q_norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.self_attn.q_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.23.self_attn.q_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.self_attn.q_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.23.self_attn.q_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.self_attn.q_proj",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.23.self_attn.q_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.self_attn.v_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.23.self_attn.v_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.self_attn.v_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.23.self_attn.v_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.self_attn.v_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.23.self_attn.v_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.24.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.linear_attn.A_log",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.24.linear_attn.A_log"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.linear_attn.conv1d",
+      "outlier_strategy": "none",
+      "parameters": 32768,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.24.linear_attn.conv1d.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.linear_attn.dt_bias",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.24.linear_attn.dt_bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.linear_attn.in_proj_a.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.24.linear_attn.in_proj_a.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.linear_attn.in_proj_a.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.24.linear_attn.in_proj_a.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.linear_attn.in_proj_a",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.24.linear_attn.in_proj_a.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.linear_attn.in_proj_b.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.24.linear_attn.in_proj_b.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.linear_attn.in_proj_b.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.24.linear_attn.in_proj_b.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.linear_attn.in_proj_b",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.24.linear_attn.in_proj_b.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.linear_attn.in_proj_qkv.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.24.linear_attn.in_proj_qkv.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.linear_attn.in_proj_qkv.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.24.linear_attn.in_proj_qkv.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.linear_attn.in_proj_qkv",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.24.linear_attn.in_proj_qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.linear_attn.in_proj_z.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.24.linear_attn.in_proj_z.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.linear_attn.in_proj_z.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.24.linear_attn.in_proj_z.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.linear_attn.in_proj_z",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.24.linear_attn.in_proj_z.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.linear_attn.norm",
+      "outlier_strategy": "none",
+      "parameters": 128,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.24.linear_attn.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.linear_attn.out_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.24.linear_attn.out_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.linear_attn.out_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.24.linear_attn.out_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.linear_attn.out_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.24.linear_attn.out_proj.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.24.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.24.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.24.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.24.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.24.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.24.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.24.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.24.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.24.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.24.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.24.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.24.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.24.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.24.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.25.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.linear_attn.A_log",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.25.linear_attn.A_log"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.linear_attn.conv1d",
+      "outlier_strategy": "none",
+      "parameters": 32768,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.25.linear_attn.conv1d.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.linear_attn.dt_bias",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.25.linear_attn.dt_bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.linear_attn.in_proj_a.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.25.linear_attn.in_proj_a.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.linear_attn.in_proj_a.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.25.linear_attn.in_proj_a.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.linear_attn.in_proj_a",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.25.linear_attn.in_proj_a.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.linear_attn.in_proj_b.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.25.linear_attn.in_proj_b.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.linear_attn.in_proj_b.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.25.linear_attn.in_proj_b.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.linear_attn.in_proj_b",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.25.linear_attn.in_proj_b.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.linear_attn.in_proj_qkv.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.25.linear_attn.in_proj_qkv.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.linear_attn.in_proj_qkv.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.25.linear_attn.in_proj_qkv.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.linear_attn.in_proj_qkv",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.25.linear_attn.in_proj_qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.linear_attn.in_proj_z.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.25.linear_attn.in_proj_z.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.linear_attn.in_proj_z.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.25.linear_attn.in_proj_z.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.linear_attn.in_proj_z",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.25.linear_attn.in_proj_z.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.linear_attn.norm",
+      "outlier_strategy": "none",
+      "parameters": 128,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.25.linear_attn.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.linear_attn.out_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.25.linear_attn.out_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.linear_attn.out_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.25.linear_attn.out_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.linear_attn.out_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.25.linear_attn.out_proj.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.25.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.25.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.25.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.25.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.25.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.25.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.25.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.25.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.25.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.25.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.25.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.25.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.25.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.25.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.26.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.linear_attn.A_log",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.26.linear_attn.A_log"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.linear_attn.conv1d",
+      "outlier_strategy": "none",
+      "parameters": 32768,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.26.linear_attn.conv1d.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.linear_attn.dt_bias",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.26.linear_attn.dt_bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.linear_attn.in_proj_a.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.26.linear_attn.in_proj_a.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.linear_attn.in_proj_a.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.26.linear_attn.in_proj_a.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.linear_attn.in_proj_a",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.26.linear_attn.in_proj_a.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.linear_attn.in_proj_b.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.26.linear_attn.in_proj_b.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.linear_attn.in_proj_b.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.26.linear_attn.in_proj_b.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.linear_attn.in_proj_b",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.26.linear_attn.in_proj_b.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.linear_attn.in_proj_qkv.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.26.linear_attn.in_proj_qkv.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.linear_attn.in_proj_qkv.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.26.linear_attn.in_proj_qkv.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.linear_attn.in_proj_qkv",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.26.linear_attn.in_proj_qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.linear_attn.in_proj_z.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.26.linear_attn.in_proj_z.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.linear_attn.in_proj_z.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.26.linear_attn.in_proj_z.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.linear_attn.in_proj_z",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.26.linear_attn.in_proj_z.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.linear_attn.norm",
+      "outlier_strategy": "none",
+      "parameters": 128,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.26.linear_attn.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.linear_attn.out_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.26.linear_attn.out_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.linear_attn.out_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.26.linear_attn.out_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.linear_attn.out_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.26.linear_attn.out_proj.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.26.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.26.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.26.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.26.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.26.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.26.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.26.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.26.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.26.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.26.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.26.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.26.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.26.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.26.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.27.input_layernorm.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.27.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.27.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.27.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.27.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.27.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.27.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.27.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.27.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.27.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.27.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.27.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.27.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.27.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.27.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.self_attn.k_norm",
+      "outlier_strategy": "none",
+      "parameters": 256,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.27.self_attn.k_norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.self_attn.k_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.27.self_attn.k_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.self_attn.k_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.27.self_attn.k_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.self_attn.k_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.27.self_attn.k_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.self_attn.o_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.27.self_attn.o_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.self_attn.o_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.27.self_attn.o_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.self_attn.o_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.27.self_attn.o_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.self_attn.q_norm",
+      "outlier_strategy": "none",
+      "parameters": 256,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.27.self_attn.q_norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.self_attn.q_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.27.self_attn.q_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.self_attn.q_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.27.self_attn.q_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.self_attn.q_proj",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.27.self_attn.q_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.self_attn.v_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.27.self_attn.v_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.self_attn.v_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.27.self_attn.v_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.self_attn.v_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.27.self_attn.v_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.28.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.linear_attn.A_log",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.28.linear_attn.A_log"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.linear_attn.conv1d",
+      "outlier_strategy": "none",
+      "parameters": 32768,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.28.linear_attn.conv1d.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.linear_attn.dt_bias",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.28.linear_attn.dt_bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.linear_attn.in_proj_a.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.28.linear_attn.in_proj_a.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.linear_attn.in_proj_a.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.28.linear_attn.in_proj_a.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.linear_attn.in_proj_a",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.28.linear_attn.in_proj_a.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.linear_attn.in_proj_b.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.28.linear_attn.in_proj_b.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.linear_attn.in_proj_b.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.28.linear_attn.in_proj_b.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.linear_attn.in_proj_b",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.28.linear_attn.in_proj_b.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.linear_attn.in_proj_qkv.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.28.linear_attn.in_proj_qkv.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.linear_attn.in_proj_qkv.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.28.linear_attn.in_proj_qkv.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.linear_attn.in_proj_qkv",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.28.linear_attn.in_proj_qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.linear_attn.in_proj_z.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.28.linear_attn.in_proj_z.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.linear_attn.in_proj_z.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.28.linear_attn.in_proj_z.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.linear_attn.in_proj_z",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.28.linear_attn.in_proj_z.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.linear_attn.norm",
+      "outlier_strategy": "none",
+      "parameters": 128,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.28.linear_attn.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.linear_attn.out_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.28.linear_attn.out_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.linear_attn.out_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.28.linear_attn.out_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.linear_attn.out_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.28.linear_attn.out_proj.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.28.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.28.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.28.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.28.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.28.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.28.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.28.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.28.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.28.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.28.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.28.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.28.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.28.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.28.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.29.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.linear_attn.A_log",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.29.linear_attn.A_log"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.linear_attn.conv1d",
+      "outlier_strategy": "none",
+      "parameters": 32768,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.29.linear_attn.conv1d.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.linear_attn.dt_bias",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.29.linear_attn.dt_bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.linear_attn.in_proj_a.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.29.linear_attn.in_proj_a.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.linear_attn.in_proj_a.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.29.linear_attn.in_proj_a.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.linear_attn.in_proj_a",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.29.linear_attn.in_proj_a.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.linear_attn.in_proj_b.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.29.linear_attn.in_proj_b.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.linear_attn.in_proj_b.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.29.linear_attn.in_proj_b.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.linear_attn.in_proj_b",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.29.linear_attn.in_proj_b.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.linear_attn.in_proj_qkv.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.29.linear_attn.in_proj_qkv.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.linear_attn.in_proj_qkv.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.29.linear_attn.in_proj_qkv.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.linear_attn.in_proj_qkv",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.29.linear_attn.in_proj_qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.linear_attn.in_proj_z.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.29.linear_attn.in_proj_z.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.linear_attn.in_proj_z.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.29.linear_attn.in_proj_z.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.linear_attn.in_proj_z",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.29.linear_attn.in_proj_z.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.linear_attn.norm",
+      "outlier_strategy": "none",
+      "parameters": 128,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.29.linear_attn.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.linear_attn.out_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.29.linear_attn.out_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.linear_attn.out_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.29.linear_attn.out_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.linear_attn.out_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.29.linear_attn.out_proj.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.29.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.29.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.29.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.29.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.29.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.29.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.29.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.29.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.29.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.29.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.29.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.29.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.29.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.29.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.3.input_layernorm.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.3.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.3.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.3.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.3.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.3.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.3.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.3.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.3.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.3.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.3.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.3.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.3.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.3.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.3.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.self_attn.k_norm",
+      "outlier_strategy": "none",
+      "parameters": 256,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.3.self_attn.k_norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.self_attn.k_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.3.self_attn.k_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.self_attn.k_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.3.self_attn.k_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.self_attn.k_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.3.self_attn.k_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.self_attn.o_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.3.self_attn.o_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.self_attn.o_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.3.self_attn.o_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.self_attn.o_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.3.self_attn.o_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.self_attn.q_norm",
+      "outlier_strategy": "none",
+      "parameters": 256,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.3.self_attn.q_norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.self_attn.q_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.3.self_attn.q_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.self_attn.q_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.3.self_attn.q_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.self_attn.q_proj",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.3.self_attn.q_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.self_attn.v_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.3.self_attn.v_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.self_attn.v_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.3.self_attn.v_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.self_attn.v_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.3.self_attn.v_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.30.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.linear_attn.A_log",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.30.linear_attn.A_log"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.linear_attn.conv1d",
+      "outlier_strategy": "none",
+      "parameters": 32768,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.30.linear_attn.conv1d.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.linear_attn.dt_bias",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.30.linear_attn.dt_bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.linear_attn.in_proj_a.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.30.linear_attn.in_proj_a.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.linear_attn.in_proj_a.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.30.linear_attn.in_proj_a.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.linear_attn.in_proj_a",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.30.linear_attn.in_proj_a.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.linear_attn.in_proj_b.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.30.linear_attn.in_proj_b.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.linear_attn.in_proj_b.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.30.linear_attn.in_proj_b.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.linear_attn.in_proj_b",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.30.linear_attn.in_proj_b.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.linear_attn.in_proj_qkv.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.30.linear_attn.in_proj_qkv.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.linear_attn.in_proj_qkv.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.30.linear_attn.in_proj_qkv.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.linear_attn.in_proj_qkv",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.30.linear_attn.in_proj_qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.linear_attn.in_proj_z.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.30.linear_attn.in_proj_z.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.linear_attn.in_proj_z.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.30.linear_attn.in_proj_z.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.linear_attn.in_proj_z",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.30.linear_attn.in_proj_z.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.linear_attn.norm",
+      "outlier_strategy": "none",
+      "parameters": 128,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.30.linear_attn.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.linear_attn.out_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.30.linear_attn.out_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.linear_attn.out_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.30.linear_attn.out_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.linear_attn.out_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.30.linear_attn.out_proj.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.30.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.30.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.30.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.30.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.30.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.30.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.30.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.30.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.30.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.30.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.30.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.30.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.30.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.30.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.31.input_layernorm.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.31.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.31.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.31.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.31.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.31.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.31.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.31.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.31.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.31.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.31.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.31.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.31.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.31.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.31.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.self_attn.k_norm",
+      "outlier_strategy": "none",
+      "parameters": 256,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.31.self_attn.k_norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.self_attn.k_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.31.self_attn.k_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.self_attn.k_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.31.self_attn.k_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.self_attn.k_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.31.self_attn.k_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.self_attn.o_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.31.self_attn.o_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.self_attn.o_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.31.self_attn.o_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.self_attn.o_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.31.self_attn.o_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.self_attn.q_norm",
+      "outlier_strategy": "none",
+      "parameters": 256,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.31.self_attn.q_norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.self_attn.q_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.31.self_attn.q_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.self_attn.q_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.31.self_attn.q_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.self_attn.q_proj",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.31.self_attn.q_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.self_attn.v_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.31.self_attn.v_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.self_attn.v_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.31.self_attn.v_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.self_attn.v_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.31.self_attn.v_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.32.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.linear_attn.A_log",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.32.linear_attn.A_log"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.linear_attn.conv1d",
+      "outlier_strategy": "none",
+      "parameters": 32768,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.32.linear_attn.conv1d.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.linear_attn.dt_bias",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.32.linear_attn.dt_bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.linear_attn.in_proj_a.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.32.linear_attn.in_proj_a.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.linear_attn.in_proj_a.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.32.linear_attn.in_proj_a.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.linear_attn.in_proj_a",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.32.linear_attn.in_proj_a.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.linear_attn.in_proj_b.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.32.linear_attn.in_proj_b.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.linear_attn.in_proj_b.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.32.linear_attn.in_proj_b.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.linear_attn.in_proj_b",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.32.linear_attn.in_proj_b.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.linear_attn.in_proj_qkv.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.32.linear_attn.in_proj_qkv.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.linear_attn.in_proj_qkv.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.32.linear_attn.in_proj_qkv.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.linear_attn.in_proj_qkv",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.32.linear_attn.in_proj_qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.linear_attn.in_proj_z.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.32.linear_attn.in_proj_z.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.linear_attn.in_proj_z.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.32.linear_attn.in_proj_z.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.linear_attn.in_proj_z",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.32.linear_attn.in_proj_z.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.linear_attn.norm",
+      "outlier_strategy": "none",
+      "parameters": 128,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.32.linear_attn.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.linear_attn.out_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.32.linear_attn.out_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.linear_attn.out_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.32.linear_attn.out_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.linear_attn.out_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.32.linear_attn.out_proj.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.32.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.32.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.32.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.32.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.32.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.32.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.32.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.32.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.32.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.32.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.32.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.32.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.32.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.32.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.33.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.linear_attn.A_log",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.33.linear_attn.A_log"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.linear_attn.conv1d",
+      "outlier_strategy": "none",
+      "parameters": 32768,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.33.linear_attn.conv1d.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.linear_attn.dt_bias",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.33.linear_attn.dt_bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.linear_attn.in_proj_a.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.33.linear_attn.in_proj_a.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.linear_attn.in_proj_a.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.33.linear_attn.in_proj_a.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.linear_attn.in_proj_a",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.33.linear_attn.in_proj_a.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.linear_attn.in_proj_b.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.33.linear_attn.in_proj_b.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.linear_attn.in_proj_b.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.33.linear_attn.in_proj_b.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.linear_attn.in_proj_b",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.33.linear_attn.in_proj_b.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.linear_attn.in_proj_qkv.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.33.linear_attn.in_proj_qkv.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.linear_attn.in_proj_qkv.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.33.linear_attn.in_proj_qkv.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.linear_attn.in_proj_qkv",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.33.linear_attn.in_proj_qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.linear_attn.in_proj_z.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.33.linear_attn.in_proj_z.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.linear_attn.in_proj_z.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.33.linear_attn.in_proj_z.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.linear_attn.in_proj_z",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.33.linear_attn.in_proj_z.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.linear_attn.norm",
+      "outlier_strategy": "none",
+      "parameters": 128,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.33.linear_attn.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.linear_attn.out_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.33.linear_attn.out_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.linear_attn.out_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.33.linear_attn.out_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.linear_attn.out_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.33.linear_attn.out_proj.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.33.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.33.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.33.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.33.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.33.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.33.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.33.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.33.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.33.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.33.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.33.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.33.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.33.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.33.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.34.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.linear_attn.A_log",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.34.linear_attn.A_log"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.linear_attn.conv1d",
+      "outlier_strategy": "none",
+      "parameters": 32768,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.34.linear_attn.conv1d.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.linear_attn.dt_bias",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.34.linear_attn.dt_bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.linear_attn.in_proj_a.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.34.linear_attn.in_proj_a.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.linear_attn.in_proj_a.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.34.linear_attn.in_proj_a.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.linear_attn.in_proj_a",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.34.linear_attn.in_proj_a.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.linear_attn.in_proj_b.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.34.linear_attn.in_proj_b.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.linear_attn.in_proj_b.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.34.linear_attn.in_proj_b.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.linear_attn.in_proj_b",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.34.linear_attn.in_proj_b.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.linear_attn.in_proj_qkv.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.34.linear_attn.in_proj_qkv.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.linear_attn.in_proj_qkv.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.34.linear_attn.in_proj_qkv.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.linear_attn.in_proj_qkv",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.34.linear_attn.in_proj_qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.linear_attn.in_proj_z.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.34.linear_attn.in_proj_z.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.linear_attn.in_proj_z.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.34.linear_attn.in_proj_z.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.linear_attn.in_proj_z",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.34.linear_attn.in_proj_z.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.linear_attn.norm",
+      "outlier_strategy": "none",
+      "parameters": 128,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.34.linear_attn.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.linear_attn.out_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.34.linear_attn.out_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.linear_attn.out_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.34.linear_attn.out_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.linear_attn.out_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.34.linear_attn.out_proj.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.34.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.34.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.34.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.34.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.34.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.34.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.34.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.34.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.34.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.34.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.34.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.34.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.34.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.34.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.35.input_layernorm.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.35.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.35.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.35.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.35.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.35.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.35.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.35.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.35.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.35.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.35.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.35.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.35.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.35.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.35.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.self_attn.k_norm",
+      "outlier_strategy": "none",
+      "parameters": 256,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.35.self_attn.k_norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.self_attn.k_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.35.self_attn.k_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.self_attn.k_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.35.self_attn.k_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.self_attn.k_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.35.self_attn.k_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.self_attn.o_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.35.self_attn.o_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.self_attn.o_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.35.self_attn.o_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.self_attn.o_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.35.self_attn.o_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.self_attn.q_norm",
+      "outlier_strategy": "none",
+      "parameters": 256,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.35.self_attn.q_norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.self_attn.q_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.35.self_attn.q_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.self_attn.q_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.35.self_attn.q_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.self_attn.q_proj",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.35.self_attn.q_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.self_attn.v_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.35.self_attn.v_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.self_attn.v_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.35.self_attn.v_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.self_attn.v_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.35.self_attn.v_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.36.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.linear_attn.A_log",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.36.linear_attn.A_log"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.linear_attn.conv1d",
+      "outlier_strategy": "none",
+      "parameters": 32768,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.36.linear_attn.conv1d.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.linear_attn.dt_bias",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.36.linear_attn.dt_bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.linear_attn.in_proj_a.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.36.linear_attn.in_proj_a.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.linear_attn.in_proj_a.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.36.linear_attn.in_proj_a.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.linear_attn.in_proj_a",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.36.linear_attn.in_proj_a.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.linear_attn.in_proj_b.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.36.linear_attn.in_proj_b.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.linear_attn.in_proj_b.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.36.linear_attn.in_proj_b.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.linear_attn.in_proj_b",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.36.linear_attn.in_proj_b.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.linear_attn.in_proj_qkv.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.36.linear_attn.in_proj_qkv.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.linear_attn.in_proj_qkv.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.36.linear_attn.in_proj_qkv.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.linear_attn.in_proj_qkv",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.36.linear_attn.in_proj_qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.linear_attn.in_proj_z.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.36.linear_attn.in_proj_z.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.linear_attn.in_proj_z.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.36.linear_attn.in_proj_z.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.linear_attn.in_proj_z",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.36.linear_attn.in_proj_z.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.linear_attn.norm",
+      "outlier_strategy": "none",
+      "parameters": 128,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.36.linear_attn.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.linear_attn.out_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.36.linear_attn.out_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.linear_attn.out_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.36.linear_attn.out_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.linear_attn.out_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.36.linear_attn.out_proj.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.36.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.36.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.36.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.36.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.36.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.36.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.36.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.36.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.36.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.36.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.36.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.36.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.36.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.36.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.37.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.linear_attn.A_log",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.37.linear_attn.A_log"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.linear_attn.conv1d",
+      "outlier_strategy": "none",
+      "parameters": 32768,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.37.linear_attn.conv1d.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.linear_attn.dt_bias",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.37.linear_attn.dt_bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.linear_attn.in_proj_a.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.37.linear_attn.in_proj_a.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.linear_attn.in_proj_a.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.37.linear_attn.in_proj_a.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.linear_attn.in_proj_a",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.37.linear_attn.in_proj_a.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.linear_attn.in_proj_b.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.37.linear_attn.in_proj_b.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.linear_attn.in_proj_b.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.37.linear_attn.in_proj_b.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.linear_attn.in_proj_b",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.37.linear_attn.in_proj_b.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.linear_attn.in_proj_qkv.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.37.linear_attn.in_proj_qkv.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.linear_attn.in_proj_qkv.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.37.linear_attn.in_proj_qkv.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.linear_attn.in_proj_qkv",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.37.linear_attn.in_proj_qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.linear_attn.in_proj_z.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.37.linear_attn.in_proj_z.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.linear_attn.in_proj_z.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.37.linear_attn.in_proj_z.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.linear_attn.in_proj_z",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.37.linear_attn.in_proj_z.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.linear_attn.norm",
+      "outlier_strategy": "none",
+      "parameters": 128,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.37.linear_attn.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.linear_attn.out_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.37.linear_attn.out_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.linear_attn.out_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.37.linear_attn.out_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.linear_attn.out_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.37.linear_attn.out_proj.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.37.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.37.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.37.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.37.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.37.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.37.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.37.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.37.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.37.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.37.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.37.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.37.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.37.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.37.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.38.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.linear_attn.A_log",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.38.linear_attn.A_log"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.linear_attn.conv1d",
+      "outlier_strategy": "none",
+      "parameters": 32768,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.38.linear_attn.conv1d.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.linear_attn.dt_bias",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.38.linear_attn.dt_bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.linear_attn.in_proj_a.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.38.linear_attn.in_proj_a.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.linear_attn.in_proj_a.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.38.linear_attn.in_proj_a.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.linear_attn.in_proj_a",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.38.linear_attn.in_proj_a.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.linear_attn.in_proj_b.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.38.linear_attn.in_proj_b.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.linear_attn.in_proj_b.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.38.linear_attn.in_proj_b.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.linear_attn.in_proj_b",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.38.linear_attn.in_proj_b.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.linear_attn.in_proj_qkv.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.38.linear_attn.in_proj_qkv.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.linear_attn.in_proj_qkv.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.38.linear_attn.in_proj_qkv.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.linear_attn.in_proj_qkv",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.38.linear_attn.in_proj_qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.linear_attn.in_proj_z.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.38.linear_attn.in_proj_z.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.linear_attn.in_proj_z.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.38.linear_attn.in_proj_z.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.linear_attn.in_proj_z",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.38.linear_attn.in_proj_z.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.linear_attn.norm",
+      "outlier_strategy": "none",
+      "parameters": 128,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.38.linear_attn.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.linear_attn.out_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.38.linear_attn.out_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.linear_attn.out_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.38.linear_attn.out_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.linear_attn.out_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.38.linear_attn.out_proj.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.38.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.38.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.38.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.38.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.38.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.38.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.38.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.38.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.38.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.38.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.38.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.38.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.38.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.38.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.39.input_layernorm.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.39.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.39.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.39.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.39.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.39.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.39.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.39.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.39.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.39.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.39.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.39.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.39.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.39.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.39.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.self_attn.k_norm",
+      "outlier_strategy": "none",
+      "parameters": 256,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.39.self_attn.k_norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.self_attn.k_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.39.self_attn.k_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.self_attn.k_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.39.self_attn.k_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.self_attn.k_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.39.self_attn.k_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.self_attn.o_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.39.self_attn.o_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.self_attn.o_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.39.self_attn.o_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.self_attn.o_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.39.self_attn.o_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.self_attn.q_norm",
+      "outlier_strategy": "none",
+      "parameters": 256,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.39.self_attn.q_norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.self_attn.q_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.39.self_attn.q_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.self_attn.q_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.39.self_attn.q_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.self_attn.q_proj",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.39.self_attn.q_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.self_attn.v_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.39.self_attn.v_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.self_attn.v_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.39.self_attn.v_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.self_attn.v_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.39.self_attn.v_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.4.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.linear_attn.A_log",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.4.linear_attn.A_log"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.linear_attn.conv1d",
+      "outlier_strategy": "none",
+      "parameters": 32768,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.4.linear_attn.conv1d.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.linear_attn.dt_bias",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.4.linear_attn.dt_bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.linear_attn.in_proj_a.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.4.linear_attn.in_proj_a.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.linear_attn.in_proj_a.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.4.linear_attn.in_proj_a.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.linear_attn.in_proj_a",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.4.linear_attn.in_proj_a.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.linear_attn.in_proj_b.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.4.linear_attn.in_proj_b.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.linear_attn.in_proj_b.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.4.linear_attn.in_proj_b.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.linear_attn.in_proj_b",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.4.linear_attn.in_proj_b.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.linear_attn.in_proj_qkv.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.4.linear_attn.in_proj_qkv.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.linear_attn.in_proj_qkv.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.4.linear_attn.in_proj_qkv.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.linear_attn.in_proj_qkv",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.4.linear_attn.in_proj_qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.linear_attn.in_proj_z.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.4.linear_attn.in_proj_z.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.linear_attn.in_proj_z.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.4.linear_attn.in_proj_z.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.linear_attn.in_proj_z",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.4.linear_attn.in_proj_z.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.linear_attn.norm",
+      "outlier_strategy": "none",
+      "parameters": 128,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.4.linear_attn.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.linear_attn.out_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.4.linear_attn.out_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.linear_attn.out_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.4.linear_attn.out_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.linear_attn.out_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.4.linear_attn.out_proj.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.4.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.4.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.4.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.4.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.4.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.4.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.4.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.4.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.4.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.4.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.4.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.4.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.4.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.4.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.5.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.linear_attn.A_log",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.5.linear_attn.A_log"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.linear_attn.conv1d",
+      "outlier_strategy": "none",
+      "parameters": 32768,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.5.linear_attn.conv1d.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.linear_attn.dt_bias",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.5.linear_attn.dt_bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.linear_attn.in_proj_a.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.5.linear_attn.in_proj_a.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.linear_attn.in_proj_a.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.5.linear_attn.in_proj_a.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.linear_attn.in_proj_a",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.5.linear_attn.in_proj_a.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.linear_attn.in_proj_b.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.5.linear_attn.in_proj_b.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.linear_attn.in_proj_b.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.5.linear_attn.in_proj_b.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.linear_attn.in_proj_b",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.5.linear_attn.in_proj_b.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.linear_attn.in_proj_qkv.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.5.linear_attn.in_proj_qkv.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.linear_attn.in_proj_qkv.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.5.linear_attn.in_proj_qkv.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.linear_attn.in_proj_qkv",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.5.linear_attn.in_proj_qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.linear_attn.in_proj_z.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.5.linear_attn.in_proj_z.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.linear_attn.in_proj_z.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.5.linear_attn.in_proj_z.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.linear_attn.in_proj_z",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.5.linear_attn.in_proj_z.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.linear_attn.norm",
+      "outlier_strategy": "none",
+      "parameters": 128,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.5.linear_attn.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.linear_attn.out_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.5.linear_attn.out_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.linear_attn.out_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.5.linear_attn.out_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.linear_attn.out_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.5.linear_attn.out_proj.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.5.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.5.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.5.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.5.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.5.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.5.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.5.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.5.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.5.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.5.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.5.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.5.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.5.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.5.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.6.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.linear_attn.A_log",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.6.linear_attn.A_log"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.linear_attn.conv1d",
+      "outlier_strategy": "none",
+      "parameters": 32768,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.6.linear_attn.conv1d.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.linear_attn.dt_bias",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.6.linear_attn.dt_bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.linear_attn.in_proj_a.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.6.linear_attn.in_proj_a.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.linear_attn.in_proj_a.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.6.linear_attn.in_proj_a.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.linear_attn.in_proj_a",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.6.linear_attn.in_proj_a.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.linear_attn.in_proj_b.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.6.linear_attn.in_proj_b.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.linear_attn.in_proj_b.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.6.linear_attn.in_proj_b.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.linear_attn.in_proj_b",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.6.linear_attn.in_proj_b.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.linear_attn.in_proj_qkv.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.6.linear_attn.in_proj_qkv.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.linear_attn.in_proj_qkv.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.6.linear_attn.in_proj_qkv.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.linear_attn.in_proj_qkv",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.6.linear_attn.in_proj_qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.linear_attn.in_proj_z.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.6.linear_attn.in_proj_z.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.linear_attn.in_proj_z.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.6.linear_attn.in_proj_z.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.linear_attn.in_proj_z",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.6.linear_attn.in_proj_z.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.linear_attn.norm",
+      "outlier_strategy": "none",
+      "parameters": 128,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.6.linear_attn.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.linear_attn.out_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.6.linear_attn.out_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.linear_attn.out_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.6.linear_attn.out_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.linear_attn.out_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.6.linear_attn.out_proj.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.6.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.6.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.6.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.6.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.6.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.6.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.6.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.6.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.6.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.6.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.6.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.6.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.6.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.6.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.7.input_layernorm.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.7.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.7.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.7.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.7.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.7.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.7.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.7.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.7.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.7.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.7.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.7.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.7.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.7.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.7.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.self_attn.k_norm",
+      "outlier_strategy": "none",
+      "parameters": 256,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.7.self_attn.k_norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.self_attn.k_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.7.self_attn.k_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.self_attn.k_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.7.self_attn.k_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.self_attn.k_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.7.self_attn.k_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.self_attn.o_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.7.self_attn.o_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.self_attn.o_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.7.self_attn.o_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.self_attn.o_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.7.self_attn.o_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.self_attn.q_norm",
+      "outlier_strategy": "none",
+      "parameters": 256,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.7.self_attn.q_norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.self_attn.q_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.7.self_attn.q_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.self_attn.q_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.7.self_attn.q_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.self_attn.q_proj",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.7.self_attn.q_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.self_attn.v_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.7.self_attn.v_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.self_attn.v_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.7.self_attn.v_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.self_attn.v_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.7.self_attn.v_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.8.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.linear_attn.A_log",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.8.linear_attn.A_log"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.linear_attn.conv1d",
+      "outlier_strategy": "none",
+      "parameters": 32768,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.8.linear_attn.conv1d.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.linear_attn.dt_bias",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.8.linear_attn.dt_bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.linear_attn.in_proj_a.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.8.linear_attn.in_proj_a.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.linear_attn.in_proj_a.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.8.linear_attn.in_proj_a.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.linear_attn.in_proj_a",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.8.linear_attn.in_proj_a.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.linear_attn.in_proj_b.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.8.linear_attn.in_proj_b.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.linear_attn.in_proj_b.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.8.linear_attn.in_proj_b.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.linear_attn.in_proj_b",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.8.linear_attn.in_proj_b.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.linear_attn.in_proj_qkv.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.8.linear_attn.in_proj_qkv.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.linear_attn.in_proj_qkv.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.8.linear_attn.in_proj_qkv.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.linear_attn.in_proj_qkv",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.8.linear_attn.in_proj_qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.linear_attn.in_proj_z.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.8.linear_attn.in_proj_z.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.linear_attn.in_proj_z.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.8.linear_attn.in_proj_z.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.linear_attn.in_proj_z",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.8.linear_attn.in_proj_z.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.linear_attn.norm",
+      "outlier_strategy": "none",
+      "parameters": 128,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.8.linear_attn.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.linear_attn.out_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.8.linear_attn.out_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.linear_attn.out_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.8.linear_attn.out_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.linear_attn.out_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.8.linear_attn.out_proj.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.8.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.8.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.8.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.8.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.8.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.8.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.8.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.8.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.8.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.8.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.8.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.8.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.8.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.8.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.9.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.linear_attn.A_log",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.9.linear_attn.A_log"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.linear_attn.conv1d",
+      "outlier_strategy": "none",
+      "parameters": 32768,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.9.linear_attn.conv1d.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.linear_attn.dt_bias",
+      "outlier_strategy": "none",
+      "parameters": 32,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.9.linear_attn.dt_bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.linear_attn.in_proj_a.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.9.linear_attn.in_proj_a.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.linear_attn.in_proj_a.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.9.linear_attn.in_proj_a.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.linear_attn.in_proj_a",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.9.linear_attn.in_proj_a.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.linear_attn.in_proj_b.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.9.linear_attn.in_proj_b.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.linear_attn.in_proj_b.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.9.linear_attn.in_proj_b.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.linear_attn.in_proj_b",
+      "outlier_strategy": "none",
+      "parameters": 65536,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.9.linear_attn.in_proj_b.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.linear_attn.in_proj_qkv.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.9.linear_attn.in_proj_qkv.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.linear_attn.in_proj_qkv.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.9.linear_attn.in_proj_qkv.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.linear_attn.in_proj_qkv",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.9.linear_attn.in_proj_qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.linear_attn.in_proj_z.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.9.linear_attn.in_proj_z.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.linear_attn.in_proj_z.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.9.linear_attn.in_proj_z.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.linear_attn.in_proj_z",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.9.linear_attn.in_proj_z.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.linear_attn.norm",
+      "outlier_strategy": "none",
+      "parameters": 128,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.9.linear_attn.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.linear_attn.out_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.9.linear_attn.out_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.linear_attn.out_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "attention",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.9.linear_attn.out_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.linear_attn.out_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule attention-mxfp8: AXQ-MXFP8 attention trunk",
+      "role": "attention",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.9.linear_attn.out_proj.weight"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.9.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.mlp.shared_expert.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.9.mlp.shared_expert.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.mlp.shared_expert.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.9.mlp.shared_expert.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.9.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.mlp.shared_expert.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.9.mlp.shared_expert.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.mlp.shared_expert.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.9.mlp.shared_expert.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.9.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.mlp.shared_expert.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.9.mlp.shared_expert.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.mlp.shared_expert.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.9.mlp.shared_expert.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.9.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.mlp.shared_expert_gate.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.9.mlp.shared_expert_gate.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.mlp.shared_expert_gate.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "router",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.9.mlp.shared_expert_gate.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule routers-8: Router floor 8-bit affine group 32 for MXFP8 remap",
+      "role": "router",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.9.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.9.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.norm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-norms: Norms non-quantizable / protected",
+      "role": "norm",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.0.attn.proj.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.0.attn.proj.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.0.attn.proj",
+      "outlier_strategy": "none",
+      "parameters": 1327104,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.0.attn.proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.0.attn.qkv.bias",
+      "outlier_strategy": "none",
+      "parameters": 3456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.0.attn.qkv.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.0.attn.qkv",
+      "outlier_strategy": "none",
+      "parameters": 3981312,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.0.attn.qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.0.mlp.linear_fc1.bias",
+      "outlier_strategy": "none",
+      "parameters": 4304,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.0.mlp.linear_fc1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.0.mlp.linear_fc1",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.0.mlp.linear_fc1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.0.mlp.linear_fc2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.0.mlp.linear_fc2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.0.mlp.linear_fc2",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.0.mlp.linear_fc2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.0.norm1.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.0.norm1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.0.norm1",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.0.norm1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.0.norm2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.0.norm2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.0.norm2",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.0.norm2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.1.attn.proj.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.1.attn.proj.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.1.attn.proj",
+      "outlier_strategy": "none",
+      "parameters": 1327104,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.1.attn.proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.1.attn.qkv.bias",
+      "outlier_strategy": "none",
+      "parameters": 3456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.1.attn.qkv.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.1.attn.qkv",
+      "outlier_strategy": "none",
+      "parameters": 3981312,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.1.attn.qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.1.mlp.linear_fc1.bias",
+      "outlier_strategy": "none",
+      "parameters": 4304,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.1.mlp.linear_fc1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.1.mlp.linear_fc1",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.1.mlp.linear_fc1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.1.mlp.linear_fc2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.1.mlp.linear_fc2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.1.mlp.linear_fc2",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.1.mlp.linear_fc2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.1.norm1.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.1.norm1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.1.norm1",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.1.norm1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.1.norm2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.1.norm2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.1.norm2",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.1.norm2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.10.attn.proj.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.10.attn.proj.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.10.attn.proj",
+      "outlier_strategy": "none",
+      "parameters": 1327104,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.10.attn.proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.10.attn.qkv.bias",
+      "outlier_strategy": "none",
+      "parameters": 3456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.10.attn.qkv.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.10.attn.qkv",
+      "outlier_strategy": "none",
+      "parameters": 3981312,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.10.attn.qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.10.mlp.linear_fc1.bias",
+      "outlier_strategy": "none",
+      "parameters": 4304,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.10.mlp.linear_fc1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.10.mlp.linear_fc1",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.10.mlp.linear_fc1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.10.mlp.linear_fc2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.10.mlp.linear_fc2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.10.mlp.linear_fc2",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.10.mlp.linear_fc2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.10.norm1.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.10.norm1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.10.norm1",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.10.norm1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.10.norm2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.10.norm2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.10.norm2",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.10.norm2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.11.attn.proj.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.11.attn.proj.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.11.attn.proj",
+      "outlier_strategy": "none",
+      "parameters": 1327104,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.11.attn.proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.11.attn.qkv.bias",
+      "outlier_strategy": "none",
+      "parameters": 3456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.11.attn.qkv.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.11.attn.qkv",
+      "outlier_strategy": "none",
+      "parameters": 3981312,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.11.attn.qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.11.mlp.linear_fc1.bias",
+      "outlier_strategy": "none",
+      "parameters": 4304,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.11.mlp.linear_fc1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.11.mlp.linear_fc1",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.11.mlp.linear_fc1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.11.mlp.linear_fc2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.11.mlp.linear_fc2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.11.mlp.linear_fc2",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.11.mlp.linear_fc2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.11.norm1.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.11.norm1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.11.norm1",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.11.norm1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.11.norm2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.11.norm2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.11.norm2",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.11.norm2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.12.attn.proj.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.12.attn.proj.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.12.attn.proj",
+      "outlier_strategy": "none",
+      "parameters": 1327104,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.12.attn.proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.12.attn.qkv.bias",
+      "outlier_strategy": "none",
+      "parameters": 3456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.12.attn.qkv.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.12.attn.qkv",
+      "outlier_strategy": "none",
+      "parameters": 3981312,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.12.attn.qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.12.mlp.linear_fc1.bias",
+      "outlier_strategy": "none",
+      "parameters": 4304,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.12.mlp.linear_fc1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.12.mlp.linear_fc1",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.12.mlp.linear_fc1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.12.mlp.linear_fc2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.12.mlp.linear_fc2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.12.mlp.linear_fc2",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.12.mlp.linear_fc2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.12.norm1.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.12.norm1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.12.norm1",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.12.norm1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.12.norm2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.12.norm2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.12.norm2",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.12.norm2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.13.attn.proj.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.13.attn.proj.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.13.attn.proj",
+      "outlier_strategy": "none",
+      "parameters": 1327104,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.13.attn.proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.13.attn.qkv.bias",
+      "outlier_strategy": "none",
+      "parameters": 3456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.13.attn.qkv.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.13.attn.qkv",
+      "outlier_strategy": "none",
+      "parameters": 3981312,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.13.attn.qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.13.mlp.linear_fc1.bias",
+      "outlier_strategy": "none",
+      "parameters": 4304,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.13.mlp.linear_fc1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.13.mlp.linear_fc1",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.13.mlp.linear_fc1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.13.mlp.linear_fc2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.13.mlp.linear_fc2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.13.mlp.linear_fc2",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.13.mlp.linear_fc2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.13.norm1.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.13.norm1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.13.norm1",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.13.norm1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.13.norm2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.13.norm2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.13.norm2",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.13.norm2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.14.attn.proj.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.14.attn.proj.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.14.attn.proj",
+      "outlier_strategy": "none",
+      "parameters": 1327104,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.14.attn.proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.14.attn.qkv.bias",
+      "outlier_strategy": "none",
+      "parameters": 3456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.14.attn.qkv.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.14.attn.qkv",
+      "outlier_strategy": "none",
+      "parameters": 3981312,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.14.attn.qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.14.mlp.linear_fc1.bias",
+      "outlier_strategy": "none",
+      "parameters": 4304,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.14.mlp.linear_fc1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.14.mlp.linear_fc1",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.14.mlp.linear_fc1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.14.mlp.linear_fc2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.14.mlp.linear_fc2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.14.mlp.linear_fc2",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.14.mlp.linear_fc2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.14.norm1.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.14.norm1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.14.norm1",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.14.norm1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.14.norm2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.14.norm2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.14.norm2",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.14.norm2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.15.attn.proj.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.15.attn.proj.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.15.attn.proj",
+      "outlier_strategy": "none",
+      "parameters": 1327104,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.15.attn.proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.15.attn.qkv.bias",
+      "outlier_strategy": "none",
+      "parameters": 3456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.15.attn.qkv.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.15.attn.qkv",
+      "outlier_strategy": "none",
+      "parameters": 3981312,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.15.attn.qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.15.mlp.linear_fc1.bias",
+      "outlier_strategy": "none",
+      "parameters": 4304,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.15.mlp.linear_fc1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.15.mlp.linear_fc1",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.15.mlp.linear_fc1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.15.mlp.linear_fc2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.15.mlp.linear_fc2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.15.mlp.linear_fc2",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.15.mlp.linear_fc2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.15.norm1.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.15.norm1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.15.norm1",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.15.norm1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.15.norm2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.15.norm2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.15.norm2",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.15.norm2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.16.attn.proj.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.16.attn.proj.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.16.attn.proj",
+      "outlier_strategy": "none",
+      "parameters": 1327104,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.16.attn.proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.16.attn.qkv.bias",
+      "outlier_strategy": "none",
+      "parameters": 3456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.16.attn.qkv.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.16.attn.qkv",
+      "outlier_strategy": "none",
+      "parameters": 3981312,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.16.attn.qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.16.mlp.linear_fc1.bias",
+      "outlier_strategy": "none",
+      "parameters": 4304,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.16.mlp.linear_fc1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.16.mlp.linear_fc1",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.16.mlp.linear_fc1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.16.mlp.linear_fc2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.16.mlp.linear_fc2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.16.mlp.linear_fc2",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.16.mlp.linear_fc2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.16.norm1.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.16.norm1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.16.norm1",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.16.norm1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.16.norm2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.16.norm2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.16.norm2",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.16.norm2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.17.attn.proj.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.17.attn.proj.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.17.attn.proj",
+      "outlier_strategy": "none",
+      "parameters": 1327104,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.17.attn.proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.17.attn.qkv.bias",
+      "outlier_strategy": "none",
+      "parameters": 3456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.17.attn.qkv.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.17.attn.qkv",
+      "outlier_strategy": "none",
+      "parameters": 3981312,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.17.attn.qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.17.mlp.linear_fc1.bias",
+      "outlier_strategy": "none",
+      "parameters": 4304,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.17.mlp.linear_fc1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.17.mlp.linear_fc1",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.17.mlp.linear_fc1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.17.mlp.linear_fc2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.17.mlp.linear_fc2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.17.mlp.linear_fc2",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.17.mlp.linear_fc2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.17.norm1.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.17.norm1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.17.norm1",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.17.norm1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.17.norm2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.17.norm2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.17.norm2",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.17.norm2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.18.attn.proj.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.18.attn.proj.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.18.attn.proj",
+      "outlier_strategy": "none",
+      "parameters": 1327104,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.18.attn.proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.18.attn.qkv.bias",
+      "outlier_strategy": "none",
+      "parameters": 3456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.18.attn.qkv.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.18.attn.qkv",
+      "outlier_strategy": "none",
+      "parameters": 3981312,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.18.attn.qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.18.mlp.linear_fc1.bias",
+      "outlier_strategy": "none",
+      "parameters": 4304,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.18.mlp.linear_fc1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.18.mlp.linear_fc1",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.18.mlp.linear_fc1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.18.mlp.linear_fc2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.18.mlp.linear_fc2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.18.mlp.linear_fc2",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.18.mlp.linear_fc2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.18.norm1.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.18.norm1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.18.norm1",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.18.norm1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.18.norm2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.18.norm2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.18.norm2",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.18.norm2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.19.attn.proj.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.19.attn.proj.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.19.attn.proj",
+      "outlier_strategy": "none",
+      "parameters": 1327104,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.19.attn.proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.19.attn.qkv.bias",
+      "outlier_strategy": "none",
+      "parameters": 3456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.19.attn.qkv.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.19.attn.qkv",
+      "outlier_strategy": "none",
+      "parameters": 3981312,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.19.attn.qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.19.mlp.linear_fc1.bias",
+      "outlier_strategy": "none",
+      "parameters": 4304,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.19.mlp.linear_fc1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.19.mlp.linear_fc1",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.19.mlp.linear_fc1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.19.mlp.linear_fc2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.19.mlp.linear_fc2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.19.mlp.linear_fc2",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.19.mlp.linear_fc2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.19.norm1.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.19.norm1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.19.norm1",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.19.norm1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.19.norm2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.19.norm2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.19.norm2",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.19.norm2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.2.attn.proj.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.2.attn.proj.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.2.attn.proj",
+      "outlier_strategy": "none",
+      "parameters": 1327104,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.2.attn.proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.2.attn.qkv.bias",
+      "outlier_strategy": "none",
+      "parameters": 3456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.2.attn.qkv.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.2.attn.qkv",
+      "outlier_strategy": "none",
+      "parameters": 3981312,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.2.attn.qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.2.mlp.linear_fc1.bias",
+      "outlier_strategy": "none",
+      "parameters": 4304,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.2.mlp.linear_fc1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.2.mlp.linear_fc1",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.2.mlp.linear_fc1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.2.mlp.linear_fc2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.2.mlp.linear_fc2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.2.mlp.linear_fc2",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.2.mlp.linear_fc2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.2.norm1.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.2.norm1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.2.norm1",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.2.norm1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.2.norm2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.2.norm2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.2.norm2",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.2.norm2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.20.attn.proj.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.20.attn.proj.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.20.attn.proj",
+      "outlier_strategy": "none",
+      "parameters": 1327104,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.20.attn.proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.20.attn.qkv.bias",
+      "outlier_strategy": "none",
+      "parameters": 3456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.20.attn.qkv.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.20.attn.qkv",
+      "outlier_strategy": "none",
+      "parameters": 3981312,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.20.attn.qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.20.mlp.linear_fc1.bias",
+      "outlier_strategy": "none",
+      "parameters": 4304,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.20.mlp.linear_fc1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.20.mlp.linear_fc1",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.20.mlp.linear_fc1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.20.mlp.linear_fc2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.20.mlp.linear_fc2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.20.mlp.linear_fc2",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.20.mlp.linear_fc2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.20.norm1.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.20.norm1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.20.norm1",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.20.norm1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.20.norm2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.20.norm2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.20.norm2",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.20.norm2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.21.attn.proj.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.21.attn.proj.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.21.attn.proj",
+      "outlier_strategy": "none",
+      "parameters": 1327104,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.21.attn.proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.21.attn.qkv.bias",
+      "outlier_strategy": "none",
+      "parameters": 3456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.21.attn.qkv.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.21.attn.qkv",
+      "outlier_strategy": "none",
+      "parameters": 3981312,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.21.attn.qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.21.mlp.linear_fc1.bias",
+      "outlier_strategy": "none",
+      "parameters": 4304,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.21.mlp.linear_fc1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.21.mlp.linear_fc1",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.21.mlp.linear_fc1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.21.mlp.linear_fc2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.21.mlp.linear_fc2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.21.mlp.linear_fc2",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.21.mlp.linear_fc2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.21.norm1.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.21.norm1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.21.norm1",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.21.norm1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.21.norm2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.21.norm2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.21.norm2",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.21.norm2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.22.attn.proj.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.22.attn.proj.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.22.attn.proj",
+      "outlier_strategy": "none",
+      "parameters": 1327104,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.22.attn.proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.22.attn.qkv.bias",
+      "outlier_strategy": "none",
+      "parameters": 3456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.22.attn.qkv.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.22.attn.qkv",
+      "outlier_strategy": "none",
+      "parameters": 3981312,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.22.attn.qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.22.mlp.linear_fc1.bias",
+      "outlier_strategy": "none",
+      "parameters": 4304,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.22.mlp.linear_fc1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.22.mlp.linear_fc1",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.22.mlp.linear_fc1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.22.mlp.linear_fc2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.22.mlp.linear_fc2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.22.mlp.linear_fc2",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.22.mlp.linear_fc2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.22.norm1.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.22.norm1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.22.norm1",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.22.norm1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.22.norm2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.22.norm2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.22.norm2",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.22.norm2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.23.attn.proj.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.23.attn.proj.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.23.attn.proj",
+      "outlier_strategy": "none",
+      "parameters": 1327104,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.23.attn.proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.23.attn.qkv.bias",
+      "outlier_strategy": "none",
+      "parameters": 3456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.23.attn.qkv.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.23.attn.qkv",
+      "outlier_strategy": "none",
+      "parameters": 3981312,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.23.attn.qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.23.mlp.linear_fc1.bias",
+      "outlier_strategy": "none",
+      "parameters": 4304,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.23.mlp.linear_fc1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.23.mlp.linear_fc1",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.23.mlp.linear_fc1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.23.mlp.linear_fc2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.23.mlp.linear_fc2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.23.mlp.linear_fc2",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.23.mlp.linear_fc2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.23.norm1.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.23.norm1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.23.norm1",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.23.norm1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.23.norm2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.23.norm2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.23.norm2",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.23.norm2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.24.attn.proj.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.24.attn.proj.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.24.attn.proj",
+      "outlier_strategy": "none",
+      "parameters": 1327104,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.24.attn.proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.24.attn.qkv.bias",
+      "outlier_strategy": "none",
+      "parameters": 3456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.24.attn.qkv.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.24.attn.qkv",
+      "outlier_strategy": "none",
+      "parameters": 3981312,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.24.attn.qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.24.mlp.linear_fc1.bias",
+      "outlier_strategy": "none",
+      "parameters": 4304,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.24.mlp.linear_fc1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.24.mlp.linear_fc1",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.24.mlp.linear_fc1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.24.mlp.linear_fc2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.24.mlp.linear_fc2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.24.mlp.linear_fc2",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.24.mlp.linear_fc2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.24.norm1.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.24.norm1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.24.norm1",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.24.norm1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.24.norm2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.24.norm2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.24.norm2",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.24.norm2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.25.attn.proj.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.25.attn.proj.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.25.attn.proj",
+      "outlier_strategy": "none",
+      "parameters": 1327104,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.25.attn.proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.25.attn.qkv.bias",
+      "outlier_strategy": "none",
+      "parameters": 3456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.25.attn.qkv.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.25.attn.qkv",
+      "outlier_strategy": "none",
+      "parameters": 3981312,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.25.attn.qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.25.mlp.linear_fc1.bias",
+      "outlier_strategy": "none",
+      "parameters": 4304,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.25.mlp.linear_fc1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.25.mlp.linear_fc1",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.25.mlp.linear_fc1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.25.mlp.linear_fc2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.25.mlp.linear_fc2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.25.mlp.linear_fc2",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.25.mlp.linear_fc2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.25.norm1.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.25.norm1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.25.norm1",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.25.norm1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.25.norm2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.25.norm2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.25.norm2",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.25.norm2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.26.attn.proj.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.26.attn.proj.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.26.attn.proj",
+      "outlier_strategy": "none",
+      "parameters": 1327104,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.26.attn.proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.26.attn.qkv.bias",
+      "outlier_strategy": "none",
+      "parameters": 3456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.26.attn.qkv.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.26.attn.qkv",
+      "outlier_strategy": "none",
+      "parameters": 3981312,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.26.attn.qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.26.mlp.linear_fc1.bias",
+      "outlier_strategy": "none",
+      "parameters": 4304,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.26.mlp.linear_fc1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.26.mlp.linear_fc1",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.26.mlp.linear_fc1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.26.mlp.linear_fc2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.26.mlp.linear_fc2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.26.mlp.linear_fc2",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.26.mlp.linear_fc2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.26.norm1.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.26.norm1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.26.norm1",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.26.norm1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.26.norm2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.26.norm2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.26.norm2",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.26.norm2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.3.attn.proj.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.3.attn.proj.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.3.attn.proj",
+      "outlier_strategy": "none",
+      "parameters": 1327104,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.3.attn.proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.3.attn.qkv.bias",
+      "outlier_strategy": "none",
+      "parameters": 3456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.3.attn.qkv.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.3.attn.qkv",
+      "outlier_strategy": "none",
+      "parameters": 3981312,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.3.attn.qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.3.mlp.linear_fc1.bias",
+      "outlier_strategy": "none",
+      "parameters": 4304,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.3.mlp.linear_fc1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.3.mlp.linear_fc1",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.3.mlp.linear_fc1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.3.mlp.linear_fc2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.3.mlp.linear_fc2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.3.mlp.linear_fc2",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.3.mlp.linear_fc2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.3.norm1.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.3.norm1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.3.norm1",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.3.norm1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.3.norm2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.3.norm2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.3.norm2",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.3.norm2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.4.attn.proj.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.4.attn.proj.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.4.attn.proj",
+      "outlier_strategy": "none",
+      "parameters": 1327104,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.4.attn.proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.4.attn.qkv.bias",
+      "outlier_strategy": "none",
+      "parameters": 3456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.4.attn.qkv.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.4.attn.qkv",
+      "outlier_strategy": "none",
+      "parameters": 3981312,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.4.attn.qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.4.mlp.linear_fc1.bias",
+      "outlier_strategy": "none",
+      "parameters": 4304,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.4.mlp.linear_fc1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.4.mlp.linear_fc1",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.4.mlp.linear_fc1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.4.mlp.linear_fc2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.4.mlp.linear_fc2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.4.mlp.linear_fc2",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.4.mlp.linear_fc2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.4.norm1.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.4.norm1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.4.norm1",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.4.norm1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.4.norm2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.4.norm2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.4.norm2",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.4.norm2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.5.attn.proj.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.5.attn.proj.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.5.attn.proj",
+      "outlier_strategy": "none",
+      "parameters": 1327104,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.5.attn.proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.5.attn.qkv.bias",
+      "outlier_strategy": "none",
+      "parameters": 3456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.5.attn.qkv.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.5.attn.qkv",
+      "outlier_strategy": "none",
+      "parameters": 3981312,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.5.attn.qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.5.mlp.linear_fc1.bias",
+      "outlier_strategy": "none",
+      "parameters": 4304,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.5.mlp.linear_fc1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.5.mlp.linear_fc1",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.5.mlp.linear_fc1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.5.mlp.linear_fc2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.5.mlp.linear_fc2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.5.mlp.linear_fc2",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.5.mlp.linear_fc2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.5.norm1.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.5.norm1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.5.norm1",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.5.norm1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.5.norm2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.5.norm2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.5.norm2",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.5.norm2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.6.attn.proj.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.6.attn.proj.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.6.attn.proj",
+      "outlier_strategy": "none",
+      "parameters": 1327104,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.6.attn.proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.6.attn.qkv.bias",
+      "outlier_strategy": "none",
+      "parameters": 3456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.6.attn.qkv.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.6.attn.qkv",
+      "outlier_strategy": "none",
+      "parameters": 3981312,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.6.attn.qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.6.mlp.linear_fc1.bias",
+      "outlier_strategy": "none",
+      "parameters": 4304,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.6.mlp.linear_fc1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.6.mlp.linear_fc1",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.6.mlp.linear_fc1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.6.mlp.linear_fc2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.6.mlp.linear_fc2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.6.mlp.linear_fc2",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.6.mlp.linear_fc2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.6.norm1.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.6.norm1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.6.norm1",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.6.norm1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.6.norm2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.6.norm2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.6.norm2",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.6.norm2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.7.attn.proj.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.7.attn.proj.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.7.attn.proj",
+      "outlier_strategy": "none",
+      "parameters": 1327104,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.7.attn.proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.7.attn.qkv.bias",
+      "outlier_strategy": "none",
+      "parameters": 3456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.7.attn.qkv.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.7.attn.qkv",
+      "outlier_strategy": "none",
+      "parameters": 3981312,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.7.attn.qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.7.mlp.linear_fc1.bias",
+      "outlier_strategy": "none",
+      "parameters": 4304,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.7.mlp.linear_fc1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.7.mlp.linear_fc1",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.7.mlp.linear_fc1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.7.mlp.linear_fc2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.7.mlp.linear_fc2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.7.mlp.linear_fc2",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.7.mlp.linear_fc2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.7.norm1.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.7.norm1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.7.norm1",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.7.norm1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.7.norm2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.7.norm2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.7.norm2",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.7.norm2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.8.attn.proj.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.8.attn.proj.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.8.attn.proj",
+      "outlier_strategy": "none",
+      "parameters": 1327104,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.8.attn.proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.8.attn.qkv.bias",
+      "outlier_strategy": "none",
+      "parameters": 3456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.8.attn.qkv.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.8.attn.qkv",
+      "outlier_strategy": "none",
+      "parameters": 3981312,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.8.attn.qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.8.mlp.linear_fc1.bias",
+      "outlier_strategy": "none",
+      "parameters": 4304,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.8.mlp.linear_fc1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.8.mlp.linear_fc1",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.8.mlp.linear_fc1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.8.mlp.linear_fc2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.8.mlp.linear_fc2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.8.mlp.linear_fc2",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.8.mlp.linear_fc2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.8.norm1.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.8.norm1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.8.norm1",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.8.norm1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.8.norm2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.8.norm2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.8.norm2",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.8.norm2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.9.attn.proj.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.9.attn.proj.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.9.attn.proj",
+      "outlier_strategy": "none",
+      "parameters": 1327104,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.9.attn.proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.9.attn.qkv.bias",
+      "outlier_strategy": "none",
+      "parameters": 3456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.9.attn.qkv.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.9.attn.qkv",
+      "outlier_strategy": "none",
+      "parameters": 3981312,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.9.attn.qkv.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.9.mlp.linear_fc1.bias",
+      "outlier_strategy": "none",
+      "parameters": 4304,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.9.mlp.linear_fc1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.9.mlp.linear_fc1",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.9.mlp.linear_fc1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.9.mlp.linear_fc2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.9.mlp.linear_fc2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.9.mlp.linear_fc2",
+      "outlier_strategy": "none",
+      "parameters": 4958208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.9.mlp.linear_fc2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.9.norm1.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.9.norm1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.9.norm1",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.9.norm1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.9.norm2.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.9.norm2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.blocks.9.norm2",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.blocks.9.norm2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.merger.linear_fc1.bias",
+      "outlier_strategy": "none",
+      "parameters": 4608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.merger.linear_fc1.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.merger.linear_fc1",
+      "outlier_strategy": "none",
+      "parameters": 21233664,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.merger.linear_fc1.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.merger.linear_fc2.bias",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.merger.linear_fc2.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.merger.linear_fc2",
+      "outlier_strategy": "none",
+      "parameters": 9437184,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.merger.linear_fc2.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.merger.norm.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.merger.norm.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.merger.norm",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.merger.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.patch_embed.proj.bias",
+      "outlier_strategy": "none",
+      "parameters": 1152,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.patch_embed.proj.bias"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.patch_embed.proj",
+      "outlier_strategy": "none",
+      "parameters": 1769472,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.patch_embed.proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "vision_tower.pos_embed",
+      "outlier_strategy": "none",
+      "parameters": 2654208,
+      "predicted_loss": 0.0,
+      "reason": "manual rule protect-vision: Vision tower BF16-protected",
+      "role": "vision",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "vision_tower.pos_embed.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.10.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.10.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.10.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.2.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.2.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.2.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.2.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.3.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.3.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.3.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.3.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.3.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.3.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.3.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.3.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.3.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.3.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.4.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.4.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.4.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.4.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.4.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.4.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.4.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.4.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.4.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.4.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.5.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.5.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.5.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.5.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.5.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.5.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.5.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.5.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.5.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.5.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.6.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.6.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.6.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.6.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.6.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.6.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.6.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.6.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.6.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.6.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.7.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.7.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.7.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.7.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.7.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.7.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.7.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.7.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.7.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.7.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.8.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.8.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.8.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.8.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.8.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.8.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.8.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.8.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.8.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.8.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.9.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.9.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.9.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.9.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.9.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.9.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.9.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.9.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.9.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.9.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.10.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.10.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.10.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.10.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.10.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.10.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.10.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.11.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.11.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.11.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.11.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.11.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.11.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.11.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.11.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.11.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.11.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.12.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.12.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.12.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.12.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.12.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.12.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.12.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.12.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.12.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.12.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.13.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.13.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.13.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.13.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.13.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.13.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.13.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.13.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.13.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.13.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.14.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.14.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.14.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.14.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.14.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.14.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.14.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.14.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.14.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.14.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.15.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.15.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.15.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.15.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.15.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.15.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.15.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.15.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.15.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.15.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.16.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.16.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.16.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.16.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.16.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.16.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.16.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.16.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.16.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.16.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.17.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.17.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.17.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.17.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.17.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.17.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.17.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.17.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.17.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.17.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.18.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.18.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.18.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.18.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.18.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.18.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.18.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.18.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.18.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.18.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.19.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.19.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.19.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.19.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.19.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.19.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.19.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.19.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.19.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.19.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.20.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.20.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.20.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.20.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.20.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.20.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.20.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.20.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.20.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.20.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.21.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.21.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.21.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.21.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.21.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.21.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.21.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.21.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.21.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.21.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.22.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.22.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.22.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.22.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.22.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.22.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.22.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.22.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.22.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.22.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.23.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.23.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.23.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.23.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.23.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.23.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.23.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.23.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.23.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.23.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.24.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.24.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.24.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.24.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.24.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.24.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.24.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.24.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.24.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.24.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.25.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.25.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.25.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.25.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.25.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.25.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.25.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.25.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.25.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.25.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.26.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.26.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.26.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.26.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.26.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.26.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.26.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.26.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.26.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.26.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.27.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.27.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.27.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.27.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.27.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.27.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.27.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.27.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.27.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.27.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.28.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.28.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.28.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.28.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.28.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.28.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.28.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.28.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.28.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.28.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.29.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.29.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.29.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.29.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.29.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.29.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.29.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.29.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.29.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.29.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.30.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.30.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.30.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.30.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.30.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.30.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.30.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.30.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.30.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.30.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.31.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.31.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.31.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.31.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.31.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.31.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.31.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.31.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.31.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.31.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.32.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.32.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.32.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.32.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.32.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.32.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.32.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.32.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.32.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.32.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.33.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.33.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.33.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.33.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.33.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.33.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.33.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.33.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.33.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.33.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.34.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.34.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.34.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.34.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.34.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.34.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.34.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.34.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.34.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.34.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.35.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.35.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.35.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.35.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.35.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.35.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.35.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.35.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.35.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.35.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.36.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.36.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.36.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.36.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.36.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.36.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.36.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.36.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.36.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.36.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.37.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.37.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.37.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.37.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.37.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.37.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.37.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.37.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.37.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.37.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.38.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.38.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.38.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.38.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.38.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.38.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.38.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.38.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.38.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.38.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.mlp.switch_mlp.down_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.39.mlp.switch_mlp.down_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.mlp.switch_mlp.down_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.39.mlp.switch_mlp.down_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.mlp.switch_mlp.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.39.mlp.switch_mlp.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.mlp.switch_mlp.gate_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.39.mlp.switch_mlp.gate_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.mlp.switch_mlp.gate_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.39.mlp.switch_mlp.gate_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.mlp.switch_mlp.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.39.mlp.switch_mlp.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.mlp.switch_mlp.up_proj.biases",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.39.mlp.switch_mlp.up_proj.biases"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.mlp.switch_mlp.up_proj.scales",
+      "outlier_strategy": "none",
+      "parameters": 0,
+      "predicted_loss": 0.0,
+      "reason": "non-quantizable tensor preserved",
+      "role": "expert",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.model.layers.39.mlp.switch_mlp.up_proj.scales"
+    },
+    {
+      "bits": 8,
+      "group_size": 32,
+      "method": "affine",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.model.layers.39.mlp.switch_mlp.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 268435456,
+      "predicted_loss": 0.0,
+      "reason": "manual rule experts-mxfp8: AXQ-MXFP8 expert trunk",
+      "role": "expert",
+      "scale_strategy": "group-affine",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 9.0
+      },
+      "tensor": "language_model.model.layers.39.mlp.switch_mlp.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.fc",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.fc.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.input_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_block",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.input_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.0.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.0.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.0.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.0.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.0.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.0.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.1.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.1.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.1.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.1.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.1.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.1.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.10.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.10.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.10.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.10.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.10.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.10.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.100.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.100.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.100.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.100.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.100.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.100.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.101.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.101.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.101.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.101.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.101.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.101.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.102.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.102.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.102.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.102.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.102.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.102.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.103.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.103.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.103.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.103.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.103.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.103.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.104.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.104.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.104.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.104.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.104.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.104.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.105.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.105.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.105.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.105.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.105.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.105.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.106.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.106.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.106.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.106.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.106.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.106.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.107.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.107.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.107.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.107.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.107.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.107.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.108.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.108.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.108.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.108.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.108.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.108.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.109.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.109.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.109.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.109.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.109.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.109.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.11.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.11.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.11.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.11.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.11.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.11.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.110.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.110.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.110.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.110.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.110.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.110.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.111.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.111.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.111.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.111.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.111.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.111.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.112.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.112.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.112.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.112.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.112.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.112.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.113.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.113.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.113.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.113.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.113.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.113.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.114.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.114.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.114.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.114.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.114.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.114.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.115.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.115.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.115.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.115.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.115.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.115.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.116.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.116.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.116.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.116.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.116.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.116.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.117.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.117.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.117.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.117.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.117.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.117.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.118.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.118.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.118.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.118.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.118.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.118.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.119.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.119.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.119.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.119.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.119.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.119.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.12.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.12.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.12.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.12.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.12.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.12.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.120.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.120.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.120.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.120.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.120.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.120.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.121.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.121.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.121.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.121.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.121.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.121.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.122.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.122.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.122.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.122.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.122.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.122.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.123.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.123.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.123.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.123.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.123.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.123.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.124.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.124.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.124.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.124.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.124.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.124.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.125.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.125.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.125.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.125.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.125.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.125.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.126.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.126.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.126.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.126.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.126.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.126.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.127.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.127.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.127.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.127.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.127.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.127.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.128.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.128.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.128.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.128.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.128.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.128.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.129.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.129.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.129.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.129.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.129.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.129.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.13.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.13.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.13.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.13.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.13.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.13.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.130.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.130.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.130.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.130.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.130.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.130.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.131.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.131.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.131.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.131.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.131.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.131.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.132.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.132.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.132.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.132.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.132.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.132.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.133.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.133.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.133.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.133.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.133.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.133.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.134.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.134.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.134.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.134.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.134.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.134.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.135.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.135.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.135.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.135.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.135.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.135.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.136.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.136.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.136.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.136.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.136.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.136.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.137.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.137.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.137.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.137.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.137.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.137.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.138.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.138.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.138.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.138.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.138.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.138.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.139.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.139.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.139.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.139.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.139.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.139.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.14.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.14.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.14.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.14.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.14.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.14.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.140.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.140.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.140.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.140.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.140.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.140.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.141.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.141.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.141.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.141.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.141.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.141.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.142.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.142.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.142.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.142.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.142.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.142.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.143.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.143.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.143.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.143.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.143.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.143.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.144.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.144.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.144.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.144.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.144.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.144.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.145.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.145.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.145.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.145.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.145.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.145.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.146.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.146.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.146.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.146.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.146.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.146.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.147.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.147.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.147.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.147.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.147.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.147.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.148.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.148.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.148.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.148.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.148.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.148.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.149.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.149.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.149.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.149.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.149.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.149.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.15.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.15.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.15.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.15.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.15.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.15.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.150.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.150.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.150.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.150.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.150.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.150.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.151.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.151.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.151.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.151.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.151.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.151.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.152.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.152.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.152.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.152.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.152.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.152.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.153.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.153.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.153.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.153.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.153.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.153.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.154.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.154.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.154.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.154.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.154.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.154.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.155.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.155.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.155.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.155.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.155.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.155.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.156.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.156.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.156.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.156.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.156.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.156.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.157.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.157.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.157.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.157.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.157.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.157.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.158.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.158.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.158.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.158.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.158.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.158.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.159.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.159.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.159.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.159.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.159.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.159.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.16.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.16.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.16.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.16.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.16.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.16.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.160.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.160.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.160.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.160.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.160.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.160.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.161.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.161.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.161.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.161.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.161.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.161.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.162.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.162.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.162.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.162.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.162.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.162.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.163.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.163.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.163.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.163.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.163.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.163.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.164.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.164.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.164.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.164.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.164.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.164.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.165.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.165.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.165.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.165.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.165.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.165.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.166.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.166.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.166.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.166.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.166.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.166.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.167.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.167.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.167.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.167.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.167.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.167.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.168.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.168.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.168.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.168.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.168.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.168.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.169.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.169.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.169.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.169.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.169.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.169.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.17.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.17.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.17.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.17.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.17.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.17.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.170.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.170.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.170.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.170.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.170.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.170.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.171.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.171.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.171.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.171.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.171.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.171.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.172.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.172.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.172.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.172.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.172.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.172.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.173.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.173.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.173.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.173.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.173.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.173.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.174.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.174.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.174.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.174.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.174.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.174.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.175.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.175.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.175.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.175.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.175.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.175.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.176.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.176.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.176.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.176.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.176.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.176.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.177.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.177.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.177.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.177.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.177.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.177.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.178.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.178.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.178.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.178.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.178.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.178.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.179.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.179.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.179.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.179.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.179.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.179.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.18.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.18.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.18.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.18.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.18.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.18.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.180.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.180.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.180.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.180.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.180.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.180.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.181.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.181.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.181.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.181.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.181.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.181.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.182.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.182.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.182.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.182.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.182.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.182.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.183.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.183.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.183.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.183.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.183.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.183.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.184.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.184.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.184.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.184.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.184.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.184.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.185.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.185.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.185.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.185.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.185.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.185.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.186.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.186.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.186.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.186.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.186.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.186.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.187.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.187.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.187.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.187.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.187.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.187.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.188.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.188.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.188.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.188.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.188.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.188.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.189.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.189.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.189.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.189.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.189.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.189.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.19.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.19.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.19.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.19.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.19.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.19.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.190.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.190.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.190.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.190.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.190.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.190.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.191.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.191.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.191.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.191.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.191.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.191.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.192.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.192.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.192.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.192.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.192.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.192.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.193.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.193.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.193.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.193.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.193.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.193.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.194.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.194.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.194.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.194.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.194.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.194.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.195.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.195.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.195.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.195.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.195.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.195.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.196.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.196.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.196.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.196.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.196.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.196.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.197.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.197.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.197.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.197.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.197.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.197.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.198.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.198.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.198.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.198.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.198.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.198.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.199.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.199.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.199.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.199.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.199.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.199.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.2.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.2.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.2.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.2.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.2.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.2.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.20.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.20.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.20.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.20.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.20.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.20.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.200.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.200.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.200.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.200.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.200.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.200.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.201.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.201.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.201.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.201.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.201.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.201.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.202.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.202.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.202.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.202.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.202.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.202.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.203.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.203.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.203.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.203.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.203.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.203.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.204.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.204.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.204.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.204.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.204.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.204.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.205.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.205.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.205.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.205.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.205.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.205.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.206.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.206.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.206.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.206.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.206.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.206.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.207.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.207.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.207.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.207.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.207.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.207.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.208.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.208.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.208.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.208.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.208.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.208.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.209.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.209.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.209.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.209.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.209.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.209.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.21.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.21.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.21.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.21.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.21.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.21.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.210.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.210.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.210.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.210.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.210.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.210.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.211.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.211.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.211.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.211.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.211.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.211.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.212.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.212.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.212.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.212.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.212.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.212.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.213.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.213.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.213.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.213.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.213.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.213.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.214.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.214.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.214.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.214.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.214.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.214.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.215.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.215.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.215.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.215.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.215.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.215.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.216.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.216.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.216.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.216.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.216.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.216.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.217.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.217.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.217.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.217.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.217.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.217.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.218.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.218.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.218.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.218.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.218.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.218.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.219.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.219.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.219.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.219.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.219.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.219.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.22.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.22.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.22.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.22.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.22.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.22.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.220.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.220.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.220.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.220.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.220.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.220.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.221.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.221.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.221.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.221.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.221.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.221.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.222.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.222.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.222.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.222.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.222.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.222.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.223.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.223.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.223.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.223.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.223.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.223.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.224.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.224.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.224.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.224.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.224.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.224.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.225.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.225.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.225.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.225.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.225.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.225.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.226.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.226.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.226.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.226.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.226.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.226.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.227.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.227.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.227.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.227.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.227.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.227.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.228.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.228.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.228.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.228.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.228.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.228.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.229.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.229.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.229.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.229.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.229.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.229.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.23.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.23.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.23.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.23.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.23.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.23.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.230.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.230.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.230.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.230.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.230.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.230.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.231.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.231.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.231.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.231.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.231.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.231.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.232.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.232.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.232.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.232.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.232.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.232.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.233.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.233.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.233.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.233.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.233.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.233.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.234.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.234.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.234.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.234.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.234.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.234.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.235.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.235.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.235.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.235.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.235.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.235.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.236.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.236.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.236.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.236.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.236.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.236.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.237.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.237.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.237.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.237.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.237.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.237.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.238.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.238.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.238.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.238.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.238.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.238.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.239.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.239.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.239.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.239.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.239.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.239.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.24.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.24.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.24.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.24.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.24.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.24.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.240.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.240.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.240.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.240.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.240.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.240.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.241.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.241.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.241.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.241.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.241.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.241.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.242.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.242.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.242.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.242.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.242.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.242.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.243.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.243.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.243.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.243.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.243.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.243.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.244.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.244.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.244.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.244.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.244.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.244.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.245.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.245.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.245.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.245.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.245.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.245.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.246.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.246.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.246.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.246.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.246.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.246.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.247.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.247.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.247.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.247.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.247.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.247.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.248.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.248.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.248.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.248.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.248.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.248.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.249.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.249.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.249.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.249.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.249.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.249.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.25.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.25.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.25.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.25.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.25.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.25.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.250.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.250.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.250.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.250.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.250.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.250.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.251.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.251.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.251.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.251.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.251.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.251.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.252.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.252.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.252.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.252.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.252.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.252.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.253.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.253.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.253.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.253.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.253.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.253.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.254.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.254.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.254.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.254.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.254.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.254.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.255.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.255.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.255.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.255.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.255.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.255.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.26.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.26.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.26.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.26.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.26.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.26.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.27.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.27.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.27.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.27.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.27.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.27.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.28.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.28.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.28.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.28.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.28.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.28.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.29.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.29.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.29.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.29.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.29.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.29.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.3.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.3.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.3.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.3.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.3.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.3.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.30.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.30.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.30.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.30.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.30.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.30.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.31.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.31.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.31.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.31.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.31.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.31.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.32.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.32.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.32.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.32.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.32.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.32.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.33.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.33.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.33.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.33.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.33.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.33.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.34.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.34.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.34.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.34.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.34.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.34.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.35.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.35.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.35.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.35.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.35.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.35.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.36.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.36.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.36.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.36.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.36.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.36.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.37.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.37.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.37.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.37.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.37.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.37.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.38.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.38.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.38.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.38.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.38.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.38.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.39.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.39.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.39.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.39.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.39.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.39.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.4.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.4.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.4.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.4.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.4.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.4.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.40.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.40.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.40.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.40.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.40.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.40.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.41.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.41.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.41.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.41.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.41.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.41.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.42.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.42.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.42.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.42.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.42.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.42.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.43.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.43.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.43.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.43.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.43.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.43.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.44.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.44.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.44.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.44.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.44.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.44.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.45.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.45.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.45.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.45.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.45.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.45.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.46.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.46.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.46.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.46.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.46.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.46.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.47.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.47.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.47.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.47.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.47.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.47.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.48.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.48.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.48.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.48.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.48.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.48.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.49.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.49.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.49.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.49.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.49.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.49.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.5.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.5.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.5.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.5.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.5.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.5.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.50.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.50.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.50.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.50.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.50.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.50.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.51.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.51.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.51.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.51.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.51.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.51.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.52.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.52.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.52.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.52.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.52.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.52.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.53.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.53.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.53.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.53.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.53.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.53.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.54.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.54.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.54.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.54.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.54.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.54.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.55.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.55.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.55.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.55.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.55.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.55.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.56.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.56.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.56.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.56.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.56.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.56.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.57.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.57.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.57.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.57.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.57.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.57.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.58.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.58.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.58.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.58.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.58.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.58.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.59.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.59.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.59.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.59.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.59.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.59.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.6.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.6.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.6.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.6.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.6.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.6.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.60.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.60.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.60.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.60.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.60.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.60.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.61.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.61.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.61.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.61.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.61.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.61.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.62.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.62.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.62.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.62.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.62.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.62.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.63.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.63.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.63.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.63.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.63.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.63.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.64.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.64.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.64.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.64.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.64.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.64.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.65.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.65.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.65.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.65.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.65.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.65.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.66.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.66.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.66.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.66.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.66.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.66.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.67.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.67.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.67.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.67.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.67.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.67.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.68.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.68.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.68.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.68.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.68.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.68.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.69.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.69.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.69.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.69.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.69.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.69.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.7.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.7.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.7.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.7.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.7.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.7.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.70.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.70.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.70.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.70.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.70.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.70.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.71.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.71.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.71.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.71.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.71.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.71.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.72.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.72.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.72.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.72.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.72.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.72.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.73.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.73.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.73.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.73.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.73.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.73.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.74.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.74.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.74.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.74.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.74.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.74.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.75.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.75.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.75.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.75.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.75.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.75.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.76.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.76.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.76.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.76.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.76.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.76.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.77.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.77.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.77.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.77.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.77.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.77.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.78.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.78.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.78.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.78.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.78.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.78.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.79.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.79.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.79.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.79.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.79.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.79.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.8.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.8.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.8.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.8.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.8.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.8.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.80.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.80.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.80.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.80.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.80.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.80.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.81.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.81.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.81.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.81.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.81.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.81.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.82.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.82.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.82.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.82.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.82.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.82.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.83.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.83.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.83.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.83.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.83.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.83.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.84.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.84.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.84.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.84.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.84.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.84.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.85.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.85.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.85.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.85.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.85.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.85.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.86.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.86.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.86.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.86.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.86.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.86.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.87.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.87.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.87.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.87.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.87.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.87.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.88.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.88.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.88.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.88.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.88.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.88.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.89.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.89.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.89.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.89.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.89.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.89.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.9.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.9.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.9.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.9.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.9.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.9.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.90.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.90.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.90.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.90.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.90.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.90.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.91.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.91.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.91.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.91.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.91.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.91.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.92.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.92.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.92.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.92.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.92.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.92.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.93.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.93.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.93.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.93.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.93.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.93.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.94.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.94.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.94.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.94.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.94.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.94.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.95.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.95.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.95.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.95.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.95.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.95.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.96.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.96.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.96.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.96.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.96.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.96.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.97.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.97.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.97.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.97.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.97.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.97.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.98.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.98.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.98.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.98.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.98.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.98.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.99.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.99.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.99.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.99.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.experts.99.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.experts.99.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.gate",
+      "outlier_strategy": "none",
+      "parameters": 524288,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_block",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.shared_expert.down_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.shared_expert.down_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.shared_expert.gate_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.shared_expert.gate_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.shared_expert.up_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.shared_expert.up_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.mlp.shared_expert_gate",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_block",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.mlp.shared_expert_gate.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.post_attention_layernorm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_block",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.post_attention_layernorm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.self_attn.k_norm",
+      "outlier_strategy": "none",
+      "parameters": 256,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_block",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.self_attn.k_norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.self_attn.k_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.self_attn.k_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.self_attn.o_proj",
+      "outlier_strategy": "none",
+      "parameters": 8388608,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.self_attn.o_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.self_attn.q_norm",
+      "outlier_strategy": "none",
+      "parameters": 256,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_block",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.self_attn.q_norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.self_attn.q_proj",
+      "outlier_strategy": "none",
+      "parameters": 16777216,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.self_attn.q_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.layers.0.self_attn.v_proj",
+      "outlier_strategy": "none",
+      "parameters": 1048576,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_projection",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.layers.0.self_attn.v_proj.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.norm",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_block",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.norm.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.pre_fc_norm_embedding",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_block",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.pre_fc_norm_embedding.weight"
+    },
+    {
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
+      "metrics": {
+        "cosine_distance": 0.0,
+        "decode_latency_cost": 0.0,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
+        "peak_memory_cost": 0.0,
+        "prefill_latency_cost": 0.0,
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
+      },
+      "module_path": "language_model.mtp.pre_fc_norm_hidden",
+      "outlier_strategy": "none",
+      "parameters": 2048,
+      "predicted_loss": 0.0,
+      "reason": "manual rule mtp-bf16: External MTP sidecar preserved byte-for-byte",
+      "role": "mtp_block",
+      "scale_strategy": "none",
+      "strategy_metadata": {
+        "selected_from_candidates": 1,
+        "storage_bpw": 16.0
+      },
+      "tensor": "language_model.mtp.pre_fc_norm_hidden.weight"
+    }
+  ],
+  "calibration": null,
+  "candidate_bits": [
+    8,
+    16
+  ],
+  "candidate_group_sizes": [],
+  "constraints": {
+    "effective_bpw_limit": 9.4,
+    "lm_head_min_bits": 16,
+    "max_model_size_ratio_to_uniform4": 1.1,
+    "minimum_mtp_acceptance_retention": 0.95,
+    "minimum_mtp_speedup": 1.2,
+    "minimum_quality_retention": 0.98
+  },
+  "cost_model": "abstract-bpw",
+  "created_at": "2026-10-04T22:11:51.746606Z",
+  "effective_bpw": 9.350650324235088,
+  "evidence_kind": "architecture_prior",
+  "global_validation_required": true,
+  "group_size": 32,
+  "hardware": {
+    "name": "ax-engine-apple-silicon-affine-dwq-v3",
+    "runtime": "ax-engine",
+    "supported_bits": [
+      2,
+      3,
+      4,
+      6,
+      8,
+      16
+    ],
+    "supported_group_sizes": [
+      32,
+      64,
+      128
+    ],
+    "supported_methods": [
+      "affine",
+      "awq",
+      "dwq",
+      "gptq",
+      "gptq-act",
+      "bf16"
+    ]
+  },
+  "kernel_latency_host_id": null,
+  "kernel_latency_sha256": null,
+  "kv_cache": null,
+  "method_near_ties": [],
+  "method_near_ties_omitted": 0,
+  "mtp": {
+    "candidate_bits": [
+      8,
+      16
+    ],
+    "min_bits": 8,
+    "mode": "protected",
+    "optimize_for_acceptance": true,
+    "preserve_external_sidecar": true,
+    "protect_norms": true,
+    "protect_output_head": true
+  },
+  "mtp_distribution": {
+    "bf16": {
+      "fraction": 1.0,
+      "parameters": 844640768
+    }
+  },
+  "nominal_bpw": 8.400743227697243,
+  "objective": {
+    "cosine_distance": 0.03,
+    "decode_latency_cost": 0.07,
+    "hidden_state_error": 0.1,
+    "long_context_loss": 0.05,
+    "mtp_acceptance_loss": 0.22,
+    "output_kl": 0.15,
+    "peak_memory_cost": 0.03,
+    "prefill_latency_cost": 0.03,
+    "task_loss_delta": 0.2,
+    "token_disagreement": 0.12
+  },
+  "primary_runtime": "ax-engine",
+  "profile": "agent-coding",
+  "quantizer": "axquant",
+  "random_seed": 20260728,
+  "schema_version": "axquant.plan.v2",
+  "software_versions": {
+    "ax_engine": "7.5.7",
+    "axquant": "1.9.0",
+    "mlx": "0.32.1",
+    "mlx_lm": "0.31.3",
+    "pydantic": "2.13.4",
+    "python": "3.12.13",
+    "safetensors": "0.8.0"
+  },
+  "source_model": {
+    "architecture": "Qwen3_5MoeForConditionalGeneration",
+    "format": "mlx",
+    "local_path": null,
+    "model_id": "peculiar-ragdoll/Cyber-Tiel-Coder-35B-A3B-MLX-oQ6e-MTP",
+    "revision": "a443d4e30fd5228942cb7695f916f4b14a88fae9"
+  },
+  "status": "planned",
+  "target_bpw": 9.4,
+  "target_class": "9p4bpw",
+  "target_mode": "low-memory",
+  "warnings": [
+    "Manual assignments are unmeasured development evidence.",
+    "Conversion requires --allow-unmeasured and cannot pass publication gates."
+  ],
+  "weight_distribution": {
+    "8bit": {
+      "fraction": 0.9499070965378446,
+      "parameters": 34150891520
+    },
+    "bf16": {
+      "fraction": 0.05009290346215543,
+      "parameters": 1800931184
+    }
+  }
+}