{
  "decision": "select mlp-only profile",
  "reason": "Preserving attention output projections costs approximately 105 MB but improves proposal KL, top-1 agreement, and hidden-state fidelity.",
  "profiles": {
    "mlp_plus_attention_output_fp8": {
      "tensor_payload_bytes": 2407368960,
      "quantized_matrices": 20,
      "proposal_kl_mean": 0.0026533013419426316,
      "proposal_top1_agreement": 0.9285714030265808,
      "top16_candidate_overlap": 0.9754464285714286,
      "draft_hidden_relative_rmse": 0.03659051767294803,
      "draft_hidden_cosine_similarity": 0.9993306398391724,
      "selector_path_token_agreement": 0.8571429252624512
    },
    "mlp_only_fp8": {
      "tensor_payload_bytes": 2512200960,
      "quantized_matrices": 15,
      "proposal_kl_mean": 0.0018929995239503794,
      "proposal_top1_agreement": 0.9642857313156128,
      "top16_candidate_overlap": 0.9765625,
      "draft_hidden_relative_rmse": 0.03063499420217177,
      "draft_hidden_cosine_similarity": 0.9995341897010803,
      "selector_path_token_agreement": 0.8571429252624512
    }
  }
}