{
  "relay-opd": {
    "baseline": {
      "accuracy": 0.171875,
      "mean_reward": 0.31242003408632096,
      "entropy": 1.7917594692220566,
      "kl_from_reference": 0.0
    },
    "final": {
      "accuracy": 0.890625,
      "mean_reward": 0.8867383809173551,
      "entropy": 1.2903922276456172,
      "kl_from_reference": 0.501367241576438
    },
    "relative_accuracy": 4.181818181818182,
    "training": {
      "steps": 120,
      "learning_rate": 0.08,
      "group_size": 4,
      "train_examples": 256,
      "validation_examples": 64,
      "data_source": "OpenAI GSM8K official train/test JSONL",
      "teacher_cache_entries": 256,
      "teacher_prefill_calls": 256,
      "online_teacher_calls": 0,
      "drift_events": 0,
      "critic_updates": 0,
      "rollout_policy_refreshes": 0,
      "last_diagnostics": {
        "prefix_failure_detected": 0.0,
        "teacher_handoff_triggered": 0.0,
        "relay_budget": 0.25,
        "student_resumes_after_teacher_leg": 1.0,
        "estimated_trajectory_reduction": 0.0,
        "loss": 0.3980543464137949,
        "policy_entropy": 0.9138031807106658
      },
      "fidelity": "mechanism reproduction on a candidate-policy model"
    }
  },
  "cort": {
    "baseline": {
      "accuracy": 0.171875,
      "mean_reward": 0.31242003408632096,
      "entropy": 1.7917594692220566,
      "kl_from_reference": 0.0
    },
    "final": {
      "accuracy": 0.890625,
      "mean_reward": 0.8907658968443455,
      "entropy": 1.7716829650176418,
      "kl_from_reference": 0.02007650420441317
    },
    "relative_accuracy": 4.181818181818182,
    "training": {
      "steps": 120,
      "learning_rate": 0.08,
      "group_size": 4,
      "train_examples": 256,
      "validation_examples": 64,
      "data_source": "OpenAI GSM8K official train/test JSONL",
      "teacher_cache_entries": 0,
      "teacher_prefill_calls": 0,
      "online_teacher_calls": 0,
      "drift_events": 0,
      "critic_updates": 0,
      "rollout_policy_refreshes": 0,
      "last_diagnostics": {
        "counterfactual_replays": 8.0,
        "rubric_conditioned_contrast": 0.3071592186438187,
        "token_weight_min": 0.8969930871364946,
        "token_weight_max": 1.145535997130842,
        "auxiliary_token_scorer_parameters": 0.0,
        "loss": -0.06456758400972576,
        "policy_entropy": 1.7834416597759666
      },
      "fidelity": "mechanism reproduction on a candidate-policy model"
    }
  },
  "schema_version": 2,
  "manifest_ref": "experiments:post-training-20260729-seed42",
  "evaluation_protocol": {
    "tier": "l1_mechanism",
    "seeds": [
      42
    ],
    "formal_comparison": false,
    "claim_policy": "single/few-seed smoke result; do not claim a stable improvement"
  },
  "provenance": {
    "artifact_path": "docs/experiments/post-training-20260729-seed42.json",
    "historical_migration": "historical-metrics-v2-2026-08-09",
    "original_code_commit": "not recorded",
    "dataset_fingerprint": "not recorded in historical artifact"
  }
}