{
  "algorithm": "sr-opsd",
  "dataset": "arithmetic-smoke",
  "baseline": {
    "accuracy": 0.1953125,
    "mean_reward": 0.359689058826365,
    "entropy": 1.791759469222057,
    "kl_from_reference": 0.0
  },
  "final": {
    "accuracy": 0.6796875,
    "mean_reward": 0.7427208776867426,
    "entropy": 1.6135261163171033,
    "kl_from_reference": 0.1782333529049517
  },
  "relative_accuracy": 2.48,
  "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": 12.0,
      "target_interpolation": 0.65,
      "renyi_order": 0.7,
      "advantage_std": 1.3067151577740534,
      "loss": -0.07176633681520245,
      "policy_entropy": 1.6509356437662863
    },
    "fidelity": "mechanism reproduction on a candidate-policy model"
  },
  "history": [
    {
      "step": 1.0,
      "loss": -0.2828899475919372,
      "accuracy": 0.6328125,
      "mean_reward": 0.7028656058081177,
      "entropy": 1.7916693470770861,
      "kl_from_reference": 9.012214496923166e-05
    },
    {
      "step": 12.0,
      "loss": -0.30244941970870565,
      "accuracy": 0.6328125,
      "mean_reward": 0.7028083639263925,
      "entropy": 1.7833897783522616,
      "kl_from_reference": 0.00836969086979391
    },
    {
      "step": 24.0,
      "loss": -0.2624524436195601,
      "accuracy": 0.6640625,
      "mean_reward": 0.7270463058532356,
      "entropy": 1.766413591618015,
      "kl_from_reference": 0.025345877604038607
    },
    {
      "step": 36.0,
      "loss": -0.256849751202259,
      "accuracy": 0.6796875,
      "mean_reward": 0.7400182248545943,
      "entropy": 1.7415053774633096,
      "kl_from_reference": 0.050254091758745116
    },
    {
      "step": 48.0,
      "loss": -0.4034149473087151,
      "accuracy": 0.6484375,
      "mean_reward": 0.726761569803379,
      "entropy": 1.721161468772834,
      "kl_from_reference": 0.0705980004492212
    },
    {
      "step": 60.0,
      "loss": -0.4304795694774432,
      "accuracy": 0.578125,
      "mean_reward": 0.6777631524460729,
      "entropy": 1.7020845096585444,
      "kl_from_reference": 0.08967495956351106
    },
    {
      "step": 72.0,
      "loss": -0.4174265552634212,
      "accuracy": 0.640625,
      "mean_reward": 0.7206928683446362,
      "entropy": 1.6903445386092906,
      "kl_from_reference": 0.10141493061276532
    },
    {
      "step": 84.0,
      "loss": -0.5913653480706857,
      "accuracy": 0.6328125,
      "mean_reward": 0.7143837766781238,
      "entropy": 1.6704168534156836,
      "kl_from_reference": 0.12134261580637094
    },
    {
      "step": 96.0,
      "loss": -0.3320811726681811,
      "accuracy": 0.6796875,
      "mean_reward": 0.7427208776867426,
      "entropy": 1.6465280353754879,
      "kl_from_reference": 0.1452314338465673
    },
    {
      "step": 108.0,
      "loss": -0.25339657999706666,
      "accuracy": 0.671875,
      "mean_reward": 0.7374121020586228,
      "entropy": 1.6315928693303177,
      "kl_from_reference": 0.16016659989173737
    },
    {
      "step": 120.0,
      "loss": -0.07176633681520245,
      "accuracy": 0.6796875,
      "mean_reward": 0.7427208776867426,
      "entropy": 1.6135261163171033,
      "kl_from_reference": 0.1782333529049517
    }
  ],
  "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/2608.09745-sr-opsd/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"
  }
}