{
  "algorithm": "causal-opd",
  "dataset": "arithmetic-smoke",
  "baseline": {
    "accuracy": 0.1953125,
    "mean_reward": 0.359689058826365,
    "entropy": 1.791759469222057,
    "kl_from_reference": 0.0
  },
  "final": {
    "accuracy": 0.6328125,
    "mean_reward": 0.714514301150916,
    "entropy": 1.432505672344747,
    "kl_from_reference": 0.35925379687730924
  },
  "relative_accuracy": 2.24,
  "training": {
    "steps": 120,
    "learning_rate": 0.08,
    "group_size": 4,
    "train_examples": 512,
    "validation_examples": 128,
    "data_source": "deterministic arithmetic smoke suite",
    "teacher_cache_entries": 512,
    "teacher_prefill_calls": 512,
    "online_teacher_calls": 0,
    "drift_events": 0,
    "critic_updates": 0,
    "rollout_policy_refreshes": 7,
    "last_diagnostics": {
      "mechanism_id": 1.0,
      "first_wrong_step_mean": 0.0,
      "advantage_std": 2.645302114366813,
      "loss": -0.47215003467370287,
      "policy_entropy": 1.5071988346753773
    },
    "fidelity": "mechanism reproduction on a candidate-policy model"
  },
  "history": [
    {
      "step": 1.0,
      "loss": 2.7755575615628914e-16,
      "accuracy": 0.5390625,
      "mean_reward": 0.6292284645371297,
      "entropy": 1.791495829997134,
      "kl_from_reference": 0.0002636392249208903
    },
    {
      "step": 12.0,
      "loss": -0.06178420436338561,
      "accuracy": 0.671875,
      "mean_reward": 0.7331497514580014,
      "entropy": 1.7829987994639065,
      "kl_from_reference": 0.008760669758148973
    },
    {
      "step": 24.0,
      "loss": -0.13479228082011385,
      "accuracy": 0.6796875,
      "mean_reward": 0.7446219052963703,
      "entropy": 1.752038976561229,
      "kl_from_reference": 0.03972049266082609
    },
    {
      "step": 36.0,
      "loss": -0.5671059391504258,
      "accuracy": 0.640625,
      "mean_reward": 0.7190951240303389,
      "entropy": 1.7110914981792527,
      "kl_from_reference": 0.08066797104280243
    },
    {
      "step": 48.0,
      "loss": -0.5264468918866162,
      "accuracy": 0.6328125,
      "mean_reward": 0.7150288058446362,
      "entropy": 1.6742464480394834,
      "kl_from_reference": 0.11751302118257174
    },
    {
      "step": 60.0,
      "loss": -0.5794040698694642,
      "accuracy": 0.625,
      "mean_reward": 0.7089493224956547,
      "entropy": 1.642485940329619,
      "kl_from_reference": 0.1492735288924356
    },
    {
      "step": 72.0,
      "loss": -0.8555745526478781,
      "accuracy": 0.6875,
      "mean_reward": 0.7463476449700682,
      "entropy": 1.595501106714302,
      "kl_from_reference": 0.1962583625077528
    },
    {
      "step": 84.0,
      "loss": -0.12043582291303045,
      "accuracy": 0.671875,
      "mean_reward": 0.7373911415677437,
      "entropy": 1.5577399258920785,
      "kl_from_reference": 0.2340195433299763
    },
    {
      "step": 96.0,
      "loss": -1.2418730727183114,
      "accuracy": 0.6328125,
      "mean_reward": 0.7150288058446362,
      "entropy": 1.512176707561539,
      "kl_from_reference": 0.27958276166051593
    },
    {
      "step": 108.0,
      "loss": -1.021919197390515,
      "accuracy": 0.6640625,
      "mean_reward": 0.733397346178058,
      "entropy": 1.4749870193180699,
      "kl_from_reference": 0.31677244990398495
    },
    {
      "step": 120.0,
      "loss": -0.47215003467370287,
      "accuracy": 0.6328125,
      "mean_reward": 0.714514301150916,
      "entropy": 1.432505672344747,
      "kl_from_reference": 0.35925379687730924
    }
  ],
  "schema_version": 2,
  "manifest_ref": "post-training:arithmetic-smoke-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/post-training/2608.03673-causal-opd/metrics/arithmetic-smoke-seed42.json",
    "dataset_fingerprint": "not recorded in historical artifact",
    "historical_migration": "historical-metrics-v2-2026-08-09",
    "original_code_commit": "not recorded"
  }
}