{
  "algorithm": "normalized-dpo",
  "dataset": "arithmetic-smoke",
  "baseline": {
    "accuracy": 0.1953125,
    "mean_reward": 0.359689058826365,
    "entropy": 1.791759469222057,
    "kl_from_reference": 0.0
  },
  "final": {
    "accuracy": 0.4765625,
    "mean_reward": 0.5776915157132447,
    "entropy": 1.3720206293987933,
    "kl_from_reference": 0.4197388398232624
  },
  "relative_accuracy": 1.44,
  "training": {
    "steps": 100,
    "learning_rate": 0.08,
    "group_size": 4,
    "train_examples": 512,
    "validation_examples": 128,
    "data_source": "deterministic arithmetic smoke suite",
    "teacher_cache_entries": 0,
    "teacher_prefill_calls": 0,
    "online_teacher_calls": 0,
    "drift_events": 0,
    "critic_updates": 0,
    "rollout_policy_refreshes": 6,
    "last_diagnostics": {
      "preference_margin": 4.639901611230503,
      "centered_softplus_loss": -2.053222109463773,
      "beta_gradient_prefactor": 1.0,
      "preference_force": 0.38603967464816663,
      "loss": -2.053222109463773,
      "policy_entropy": 1.3245922738894989
    },
    "fidelity": "mechanism reproduction on a candidate-policy model"
  },
  "history": [
    {
      "step": 1.0,
      "loss": 0.0,
      "accuracy": 0.03125,
      "mean_reward": 0.16727615568695806,
      "entropy": 1.791316196851267,
      "kl_from_reference": 0.0004432723707871453
    },
    {
      "step": 10.0,
      "loss": -0.20745173416335638,
      "accuracy": 0.3984375,
      "mean_reward": 0.5165101555845149,
      "entropy": 1.7724794949519476,
      "kl_from_reference": 0.019279974270107785
    },
    {
      "step": 20.0,
      "loss": -0.49232192162135924,
      "accuracy": 0.3671875,
      "mean_reward": 0.4897437065280847,
      "entropy": 1.7039704250733758,
      "kl_from_reference": 0.08778904414867915
    },
    {
      "step": 30.0,
      "loss": -0.7929021677237225,
      "accuracy": 0.375,
      "mean_reward": 0.4967870798835876,
      "entropy": 1.646777517739796,
      "kl_from_reference": 0.14498195148225929
    },
    {
      "step": 40.0,
      "loss": -0.943235981343642,
      "accuracy": 0.4375,
      "mean_reward": 0.5471050615065127,
      "entropy": 1.6046475234542514,
      "kl_from_reference": 0.18711194576780352
    },
    {
      "step": 50.0,
      "loss": -0.9208501363378929,
      "accuracy": 0.4765625,
      "mean_reward": 0.5776915157132447,
      "entropy": 1.5762799461500587,
      "kl_from_reference": 0.21547952307199653
    },
    {
      "step": 60.0,
      "loss": -0.1216605451933006,
      "accuracy": 0.4765625,
      "mean_reward": 0.5776915157132447,
      "entropy": 1.545688894171097,
      "kl_from_reference": 0.24607057505095833
    },
    {
      "step": 70.0,
      "loss": -1.0474072596953066,
      "accuracy": 0.4765625,
      "mean_reward": 0.5776915157132447,
      "entropy": 1.4850635779826455,
      "kl_from_reference": 0.3066958912394095
    },
    {
      "step": 80.0,
      "loss": -0.06147698379158073,
      "accuracy": 0.4765625,
      "mean_reward": 0.5776915157132447,
      "entropy": 1.4446292122693694,
      "kl_from_reference": 0.34713025695268585
    },
    {
      "step": 90.0,
      "loss": 0.7153826939141206,
      "accuracy": 0.484375,
      "mean_reward": 0.5836511096219167,
      "entropy": 1.417630063281959,
      "kl_from_reference": 0.3741294059400955
    },
    {
      "step": 100.0,
      "loss": -2.053222109463773,
      "accuracy": 0.4765625,
      "mean_reward": 0.5776915157132447,
      "entropy": 1.3720206293987933,
      "kl_from_reference": 0.4197388398232624
    }
  ],
  "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/2608.27032-normalized-dpo/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"
  }
}