{
  "algorithm": "serpo",
  "dataset": "arithmetic-smoke",
  "baseline": {
    "accuracy": 0.1953125,
    "mean_reward": 0.359689058826365,
    "entropy": 1.791759469222057,
    "kl_from_reference": 0.0
  },
  "final": {
    "accuracy": 0.328125,
    "mean_reward": 0.49280061959272564,
    "entropy": 1.557815149442971,
    "kl_from_reference": 0.23394431977908514
  },
  "relative_accuracy": 0.68,
  "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": 10.0,
      "rubric_margin": 2.2184389626813044,
      "advantage_std": 0.9999958972636801,
      "loss": -0.21509762319503606,
      "policy_entropy": 1.6661424014176276
    },
    "fidelity": "mechanism reproduction on a candidate-policy model"
  },
  "history": [
    {
      "step": 1.0,
      "loss": 1.249000902703301e-16,
      "accuracy": 0.6484375,
      "mean_reward": 0.7263604799757403,
      "entropy": 1.7917205006414418,
      "kl_from_reference": 3.896858061362339e-05
    },
    {
      "step": 12.0,
      "loss": -0.03001047110993657,
      "accuracy": 0.328125,
      "mean_reward": 0.49280061959272564,
      "entropy": 1.7884745266450819,
      "kl_from_reference": 0.003284942576973876
    },
    {
      "step": 24.0,
      "loss": -0.13392860408213308,
      "accuracy": 0.359375,
      "mean_reward": 0.5185540240017683,
      "entropy": 1.779317027921861,
      "kl_from_reference": 0.012442441300194769
    },
    {
      "step": 36.0,
      "loss": -0.26849586649478635,
      "accuracy": 0.4140625,
      "mean_reward": 0.5586334742657426,
      "entropy": 1.7669366707816356,
      "kl_from_reference": 0.024822798440420124
    },
    {
      "step": 48.0,
      "loss": -0.22404082704285355,
      "accuracy": 0.34375,
      "mean_reward": 0.5053226115467783,
      "entropy": 1.742668706020548,
      "kl_from_reference": 0.04909076320150676
    },
    {
      "step": 60.0,
      "loss": -0.06389798795272639,
      "accuracy": 0.328125,
      "mean_reward": 0.49280061959272564,
      "entropy": 1.7133350926477184,
      "kl_from_reference": 0.07842437657433655
    },
    {
      "step": 72.0,
      "loss": -0.6308932399957752,
      "accuracy": 0.296875,
      "mean_reward": 0.468677279507009,
      "entropy": 1.6571915017618677,
      "kl_from_reference": 0.1345679674601863
    },
    {
      "step": 84.0,
      "loss": -0.1334780873236191,
      "accuracy": 0.296875,
      "mean_reward": 0.468677279507009,
      "entropy": 1.609475544862038,
      "kl_from_reference": 0.1822839243600174
    },
    {
      "step": 96.0,
      "loss": 0.039250896141855535,
      "accuracy": 0.3046875,
      "mean_reward": 0.47475757647766553,
      "entropy": 1.608345375894548,
      "kl_from_reference": 0.18341409332750713
    },
    {
      "step": 108.0,
      "loss": -0.06842581041378942,
      "accuracy": 0.3046875,
      "mean_reward": 0.47475757647766553,
      "entropy": 1.5635882809185628,
      "kl_from_reference": 0.2281711883034931
    },
    {
      "step": 120.0,
      "loss": -0.21509762319503606,
      "accuracy": 0.328125,
      "mean_reward": 0.49280061959272564,
      "entropy": 1.557815149442971,
      "kl_from_reference": 0.23394431977908514
    }
  ],
  "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/2607.26873-serpo/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"
  }
}
