{
  "algorithm": "erpo",
  "dataset": "arithmetic-smoke",
  "baseline": {
    "accuracy": 0.1953125,
    "mean_reward": 0.359689058826365,
    "entropy": 1.791759469222057,
    "kl_from_reference": 0.0
  },
  "final": {
    "accuracy": 0.609375,
    "mean_reward": 0.6995832246593663,
    "entropy": 1.7856474606584278,
    "kl_from_reference": 0.006112008563627598
  },
  "relative_accuracy": 2.12,
  "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": {
      "query_kl": 0.005663224607827989,
      "typicality_weight_mean": 0.999999999994,
      "response_policy_kl_coefficient": 0.0,
      "loss": -0.027563214836970204,
      "policy_entropy": 1.786096244614227
    },
    "fidelity": "mechanism reproduction on a candidate-policy model"
  },
  "history": [
    {
      "step": 1.0,
      "loss": 6.245004513516506e-17,
      "accuracy": 0.5390625,
      "mean_reward": 0.6292284645371297,
      "entropy": 1.7917548734106357,
      "kl_from_reference": 4.59581141915344e-06
    },
    {
      "step": 10.0,
      "loss": -0.00016750616874705582,
      "accuracy": 0.640625,
      "mean_reward": 0.7210101674899789,
      "entropy": 1.791710500120454,
      "kl_from_reference": 4.8969101600715185e-05
    },
    {
      "step": 20.0,
      "loss": -0.005186615591368897,
      "accuracy": 0.3984375,
      "mean_reward": 0.5472515549924221,
      "entropy": 1.7914106369753697,
      "kl_from_reference": 0.00034883224668538044
    },
    {
      "step": 30.0,
      "loss": -0.0024053319708897056,
      "accuracy": 0.6328125,
      "mean_reward": 0.7134201391761478,
      "entropy": 1.7911012978453305,
      "kl_from_reference": 0.0006581713767247208
    },
    {
      "step": 40.0,
      "loss": -0.0020625140491750873,
      "accuracy": 0.6328125,
      "mean_reward": 0.7141494489926746,
      "entropy": 1.790660359182211,
      "kl_from_reference": 0.001099110039844837
    },
    {
      "step": 50.0,
      "loss": -0.004766715556080173,
      "accuracy": 0.515625,
      "mean_reward": 0.6333557721904439,
      "entropy": 1.789911474692707,
      "kl_from_reference": 0.0018479945293478025
    },
    {
      "step": 60.0,
      "loss": -0.01626122374285259,
      "accuracy": 0.4765625,
      "mean_reward": 0.6043193950593897,
      "entropy": 1.7889874826908536,
      "kl_from_reference": 0.002771986531201008
    },
    {
      "step": 70.0,
      "loss": -0.01837273919423127,
      "accuracy": 0.4921875,
      "mean_reward": 0.6157657762020605,
      "entropy": 1.7880554979173826,
      "kl_from_reference": 0.0037039713046718506
    },
    {
      "step": 80.0,
      "loss": -0.00013732589288004138,
      "accuracy": 0.4609375,
      "mean_reward": 0.5928908405221923,
      "entropy": 1.7871416853857032,
      "kl_from_reference": 0.004617783836351637
    },
    {
      "step": 90.0,
      "loss": -0.006322111735438785,
      "accuracy": 0.515625,
      "mean_reward": 0.6332451022385487,
      "entropy": 1.786304419139956,
      "kl_from_reference": 0.0054550500820997145
    },
    {
      "step": 100.0,
      "loss": -0.027563214836970204,
      "accuracy": 0.609375,
      "mean_reward": 0.6995832246593663,
      "entropy": 1.7856474606584278,
      "kl_from_reference": 0.006112008563627598
    }
  ],
  "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/2608.23311-erpo/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"
  }
}