Publish Ropedia Xperience-10M derived artifacts

PROJECT_README.md

@@ -118,6 +118,15 @@ included for readers who want to inspect the implementation, but they are
 supporting materials rather than the main reading path. Use
 [`ARTIFACT_GUIDE.md`](ARTIFACT_GUIDE.md) when you want the full file map.
 
+Source alignment is mirrored through [`SOURCE_ALIGNMENT_AUDIT.md`](SOURCE_ALIGNMENT_AUDIT.md)
+and `source_alignment_audit.json`. The official gated dataset card records
+`31.9 TB` on the live HF surface and an `about-1PB` full-scale storage
+statement; the committed API-listing snapshot records `12,103 episode folders`
+as upstream listing metadata only, not local raw-data possession. The public
+sample remains under `cc-by-nc-4.0`, with the `HOMIE Toolkit` and
+`Rerun 0.29.0` noted as source tooling. The official note that the data is
+`limited in diversity` is preserved.
+
 ## Project Status
 
 If you only have one minute, use

docs/data/omni_model_comparison.json

@@ -1,6 +1,6 @@
 {
   "title": "Ropedia Xperience-10M Current Result Versions and Model Groups",
-  "generated_at_utc": "2026-06-11T03:22:03+00:00",
+  "generated_at_utc": "2026-06-11T04:42:46+00:00",
   "status": "pass",
   "version_count": 3,
   "model_group_count": 5,
@@ -1626,6 +1626,6 @@
     "Cosmos3-Super now has both a 128-episode base-weight Reasoner evaluation on the JSON task and a fine-tuned forward-dynamics LoRA branch over camera-pose proxy targets."
   ],
   "pending": [
-    "Use the verified Qwen3 v4 4-epoch full-eval package as the current Qwen row; older Qwen package rows remain historical diagnostics for comparison."
+    "Use the verified Qwen3 v5 dense multiscale full-eval package as the current Qwen row; older Qwen package rows remain historical diagnostics for comparison."
   ]
 }

docs/data/publication_audit.json

@@ -1,6 +1,6 @@
 {
   "status": "pass",
-  "generated_at_utc": "2026-06-11T03:36:31+00:00",
+  "generated_at_utc": "2026-06-11T04:43:37+00:00",
   "checks": [
     {
       "name": "required_publication_assets_present",

docs/data/website_integrity.json

@@ -1,6 +1,6 @@
 {
   "status": "pass",
-  "generated_at_utc": "2026-06-11T03:40:10+00:00",
+  "generated_at_utc": "2026-06-11T04:43:29+00:00",
   "docs_root": "docs",
   "site_base": "/ropedia-xperience-10m-task-suite/",
   "summary": {
@@ -288,7 +288,7 @@
     },
     {
       "path": "data/live_publication_status.json",
-      "bytes": 89685,
+      "bytes": 97776,
       "top_level_type": "dict"
     },
     {
@@ -308,7 +308,7 @@
     },
     {
       "path": "data/omni_model_comparison.json",
-      "bytes": 75698,
+      "bytes": 75707,
       "top_level_type": "dict"
     },
     {

results/omni_finetune/OMNI_MODEL_COMPARISON.md

@@ -1,6 +1,6 @@
 # Omni Model Comparison
 
-Generated: `2026-06-11T03:22:03+00:00`
+Generated: `2026-06-11T04:42:46+00:00`
 
 Compare only rows with the same scope and target. Single-episode raw-feature metrics, 128-episode metadata baselines, Qwen3 structured JSON metrics, and the two Cosmos3 targets answer different questions: Nano future-window retrieval versus Super structured JSON Reasoner evaluation.
 
@@ -129,4 +129,4 @@ This is the first verified Cosmos3-Super fine-tuned adapter branch. Its metric i
 
 ## Pending
 
-- Use the verified Qwen3 v4 4-epoch full-eval package as the current Qwen row; older Qwen package rows remain historical diagnostics for comparison.
+- Use the verified Qwen3 v5 dense multiscale full-eval package as the current Qwen row; older Qwen package rows remain historical diagnostics for comparison.

results/omni_finetune/xperience10m_cosmos3_super_training_contract_audit_camera_pose_20260608/RUN_REPORT.md

@@ -0,0 +1,35 @@
+# Cosmos3-Super Training Contract Audit
+
+- Run id: `xperience10m_cosmos3_super_training_contract_audit_camera_pose_20260608`
+- Dataset: `/home/cy/Ropedia/ropedia-episode-task-suite/results/omni_finetune/xperience10m_cosmos3_camera_pose_targets_20260608/dataset_with_cosmos_actions.jsonl`
+- Rows: `3808`
+- Rows with Cosmos action targets: `3808`
+- Valid Cosmos action targets: `3808`
+- Status: `ready_for_cosmos3_super_forward_dynamics_lora`
+- Weights updated: `False`
+
+## Blockers
+
+- None
+
+## Required Target Schema
+
+```json
+{
+  "cosmos_action_target": {
+    "mode": "policy|forward_dynamics|inverse_dynamics",
+    "domain_name": "one Cosmos3 embodiment domain supported by CosmosActionCondition",
+    "chunk_size": "positive integer action transition count",
+    "raw_actions": "required for forward_dynamics; list[list[float]] with shape [T, raw_action_dim]",
+    "video": "required for inverse_dynamics, or image/video conditioning for policy and forward_dynamics",
+    "resolution_tier": "optional; one of 256, 480, 704, 720",
+    "view_point": "optional; ego_view|third_person_view|wrist_view|concat_view"
+  }
+}
+```
+
+## Next Steps
+
+- Run the one-sample action batch packer that calls Cosmos3OmniPipeline.prepare_latents and the static segment helpers, then records whether the current target supervises vision or action tokens.
+- For the current camera_pose forward_dynamics target, implement a one-sample overfit with vision velocity/rectified-flow loss under action conditioning; add a policy/inverse target export before claiming supervised action-token prediction.
+- Run a one-episode overfit before scheduling a 96/16/16 Super LoRA run; only publish a Cosmos model repo after new adapter/checkpoint weights exist.

results/omni_finetune/xperience10m_cosmos3_super_training_contract_audit_camera_pose_20260608/training_contract_audit.json

@@ -0,0 +1,122 @@
+{
+  "run_id": "xperience10m_cosmos3_super_training_contract_audit_camera_pose_20260608",
+  "run_kind": "cosmos3_super_training_contract_audit",
+  "started_at_unix": 1780853191.2988842,
+  "finished_at_unix": 1780853193.8374674,
+  "elapsed_seconds": 2.538583517074585,
+  "workspace": "/home/cy/Ropedia/ropedia-episode-task-suite",
+  "dataset_jsonl": "/home/cy/Ropedia/ropedia-episode-task-suite/results/omni_finetune/xperience10m_cosmos3_camera_pose_targets_20260608/dataset_with_cosmos_actions.jsonl",
+  "sample_limit": 0,
+  "backbone_config": "/home/cy/Ropedia/ropedia-episode-task-suite/configs/omni_backbones/cosmos3_super_reasoner.json",
+  "backbone": {
+    "id": "cosmos3_super_reasoner",
+    "display_name": "Cosmos3-Super Reasoner",
+    "training_objective": "zero_shot_structured_episode_understanding_json_qa_via_vllm_reasoner"
+  },
+  "model": {
+    "provided": true,
+    "path": "/home/cy/Ropedia/cosmos3_models/nv-community__Cosmos3-Super",
+    "exists": true,
+    "model_type": "cosmos3_omni",
+    "architectures": [
+      "Cosmos3ForConditionalGeneration"
+    ],
+    "pipeline_class": "Cosmos3OmniDiffusersPipeline",
+    "transformer_class": "Cosmos3OmniTransformer",
+    "action_gen": true,
+    "action_dim": 64,
+    "lora_enabled_default": false,
+    "lora_rank_default": 16,
+    "lora_alpha_default": 32,
+    "lora_target_modules_default": "q_proj_moe_gen,k_proj_moe_gen,v_proj_moe_gen,o_proj_moe_gen",
+    "rectified_flow_training_config_keys": [
+      "_type",
+      "action_loss_weight",
+      "high_sigma_ratio",
+      "high_sigma_timesteps_max",
+      "high_sigma_timesteps_min",
+      "image_loss_scale",
+      "independent_action_schedule",
+      "independent_sound_schedule",
+      "loss_scale",
+      "normalize_loss_by_active",
+      "shift",
+      "shift_action",
+      "shift_sound",
+      "sound_loss_scale",
+      "train_time_action_distribution",
+      "train_time_image_distribution",
+      "train_time_sound_distribution",
+      "train_time_video_distribution",
+      "train_time_weight",
+      "use_discrete_rf",
+      "use_dynamic_shift",
+      "use_high_sigma_strategy",
+      "use_high_sigma_strategy_action",
+      "use_high_sigma_strategy_sound"
+    ]
+  },
+  "dataset": {
+    "num_rows": 3808,
+    "split_counts": {
+      "train": 2848,
+      "val": 512,
+      "test": 448
+    },
+    "episode_split_counts": {
+      "test": 14,
+      "train": 89,
+      "val": 16
+    },
+    "rows_with_video": 3808,
+    "missing_json_answer": 0,
+    "missing_json_fields": {},
+    "rows_with_action_target": 3808,
+    "valid_action_targets": 3808,
+    "target_key_counts": {
+      "cosmos_action_target": 3808
+    },
+    "target_mode_counts": {
+      "forward_dynamics": 3808
+    },
+    "target_issue_counts": {},
+    "target_issue_examples": []
+  },
+  "decision": {
+    "status": "ready_for_cosmos3_super_forward_dynamics_lora",
+    "weights_updated": false,
+    "blockers": [],
+    "warnings": [],
+    "required_target_schema": {
+      "cosmos_action_target": {
+        "mode": "policy|forward_dynamics|inverse_dynamics",
+        "domain_name": "one Cosmos3 embodiment domain supported by CosmosActionCondition",
+        "chunk_size": "positive integer action transition count",
+        "raw_actions": "required for forward_dynamics; list[list[float]] with shape [T, raw_action_dim]",
+        "video": "required for inverse_dynamics, or image/video conditioning for policy and forward_dynamics",
+        "resolution_tier": "optional; one of 256, 480, 704, 720",
+        "view_point": "optional; ego_view|third_person_view|wrist_view|concat_view"
+      }
+    },
+    "trainer_contract": {
+      "diffusers_classes": [
+        "Cosmos3OmniPipeline",
+        "Cosmos3OmniTransformer",
+        "CosmosActionCondition"
+      ],
+      "packing_helpers": [
+        "Cosmos3OmniPipeline.prepare_latents",
+        "Cosmos3OmniPipeline._prepare_text_segment",
+        "Cosmos3OmniPipeline._prepare_vision_segment",
+        "Cosmos3OmniPipeline._prepare_action_segment"
+      ],
+      "forward_outputs": "Cosmos3OmniTransformer.forward returns (preds_vision, preds_sound, preds_action). The current camera_pose forward_dynamics target uses raw actions as conditioning and should supervise preds_vision; supervised preds_action needs policy or inverse_dynamics targets.",
+      "lora_targets": "use checkpoint-declared q_proj_moe_gen,k_proj_moe_gen,v_proj_moe_gen,o_proj_moe_gen unless a new audited config overrides them"
+    },
+    "next_steps": [
+      "Run the one-sample action batch packer that calls Cosmos3OmniPipeline.prepare_latents and the static segment helpers, then records whether the current target supervises vision or action tokens.",
+      "For the current camera_pose forward_dynamics target, implement a one-sample overfit with vision velocity/rectified-flow loss under action conditioning; add a policy/inverse target export before claiming supervised action-token prediction.",
+      "Run a one-episode overfit before scheduling a 96/16/16 Super LoRA run; only publish a Cosmos model repo after new adapter/checkpoint weights exist."
+    ]
+  }
+}

scripts/omni/build_omni_model_comparison.py

@@ -721,10 +721,14 @@ def build_report() -> dict[str, Any]:
         for branch in versions[2].get("branches", [])
         if branch.get("backbone") == "qwen3_omni_lora"
     }
-    pending = [
-        "Use the verified Qwen3 v4 4-epoch full-eval package as the current Qwen row; older Qwen package rows remain historical diagnostics for comparison.",
-    ]
-    if QWEN_V5_EVAL_RUN_ID not in qwen_branch_ids:
+    if QWEN_V5_EVAL_RUN_ID in qwen_branch_ids:
+        pending = [
+            "Use the verified Qwen3 v5 dense multiscale full-eval package as the current Qwen row; older Qwen package rows remain historical diagnostics for comparison.",
+        ]
+    else:
+        pending = [
+            "Use the verified Qwen3 v4 4-epoch full-eval package as the current Qwen row; older Qwen package rows remain historical diagnostics for comparison.",
+        ]
         pending.append(
             "Complete the Qwen3-Omni v5 dense multiscale raw-media export, all-GPU LoRA train, held-out eval, and public package before promoting it over the current Qwen v4 row."
         )