{
  "algorithm": "rubric-dropout",
  "dataset": "arithmetic-smoke",
  "baseline": {
    "accuracy": 0.1953125,
    "mean_reward": 0.359689058826365,
    "entropy": 1.791759469222057,
    "kl_from_reference": 0.0
  },
  "final": {
    "accuracy": 0.28125,
    "mean_reward": 0.45564911291653953,
    "entropy": 1.27859848597865,
    "kl_from_reference": 0.5131609832434061
  },
  "relative_accuracy": 0.44,
  "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": 0,
    "teacher_prefill_calls": 0,
    "online_teacher_calls": 0,
    "drift_events": 0,
    "critic_updates": 0,
    "rollout_policy_refreshes": 7,
    "last_diagnostics": {
      "mechanism_id": 8.0,
      "rubrics_retained": 2.0,
      "advantage_std": 0.9999966393390927,
      "loss": -0.9379133640410813,
      "policy_entropy": 1.3375252928066517
    },
    "fidelity": "mechanism reproduction on a candidate-policy model"
  },
  "history": [
    {
      "step": 1.0,
      "loss": 1.6653345369377348e-16,
      "accuracy": 0.625,
      "mean_reward": 0.707737045084079,
      "entropy": 1.791722585206532,
      "kl_from_reference": 3.6884015523247646e-05
    },
    {
      "step": 12.0,
      "loss": -0.038161955286957094,
      "accuracy": 0.28125,
      "mean_reward": 0.45564911291653953,
      "entropy": 1.77934474385804,
      "kl_from_reference": 0.012414725364015316
    },
    {
      "step": 24.0,
      "loss": -0.06314212637497962,
      "accuracy": 0.28125,
      "mean_reward": 0.45564911291653953,
      "entropy": 1.7576443962146804,
      "kl_from_reference": 0.03411507300737475
    },
    {
      "step": 36.0,
      "loss": -0.43913118454505745,
      "accuracy": 0.28125,
      "mean_reward": 0.45564911291653953,
      "entropy": 1.7164098644817336,
      "kl_from_reference": 0.07534960474032062
    },
    {
      "step": 48.0,
      "loss": -0.09089889136716922,
      "accuracy": 0.28125,
      "mean_reward": 0.45564911291653953,
      "entropy": 1.6594325594181836,
      "kl_from_reference": 0.13232690980387193
    },
    {
      "step": 60.0,
      "loss": 0.043225278210572426,
      "accuracy": 0.28125,
      "mean_reward": 0.45564911291653953,
      "entropy": 1.6123042740074236,
      "kl_from_reference": 0.1794551952146315
    },
    {
      "step": 72.0,
      "loss": -0.064625551410194,
      "accuracy": 0.28125,
      "mean_reward": 0.45564911291653953,
      "entropy": 1.5584724650987893,
      "kl_from_reference": 0.23328700412326542
    },
    {
      "step": 84.0,
      "loss": -0.4959905060994379,
      "accuracy": 0.28125,
      "mean_reward": 0.45564911291653953,
      "entropy": 1.5026595913540817,
      "kl_from_reference": 0.28909987786797275
    },
    {
      "step": 96.0,
      "loss": -0.517043252402965,
      "accuracy": 0.28125,
      "mean_reward": 0.45564911291653953,
      "entropy": 1.4154445299100094,
      "kl_from_reference": 0.376314939312046
    },
    {
      "step": 108.0,
      "loss": -0.23964662033950046,
      "accuracy": 0.28125,
      "mean_reward": 0.45564911291653953,
      "entropy": 1.3565262551734256,
      "kl_from_reference": 0.4352332140486299
    },
    {
      "step": 120.0,
      "loss": -0.9379133640410813,
      "accuracy": 0.28125,
      "mean_reward": 0.45564911291653953,
      "entropy": 1.27859848597865,
      "kl_from_reference": 0.5131609832434061
    }
  ],
  "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/2608.11669-rubric-dropout/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"
  }
}