{
  "algorithm": "opdvr",
  "dataset": "arithmetic-smoke",
  "baseline": {
    "accuracy": 0.1953125,
    "mean_reward": 0.359689058826365,
    "entropy": 1.791759469222057,
    "kl_from_reference": 0.0
  },
  "final": {
    "accuracy": 0.6640625,
    "mean_reward": 0.733397346178058,
    "entropy": 1.7899356503961907,
    "kl_from_reference": 0.0018238188258643837
  },
  "relative_accuracy": 2.4,
  "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": {
      "implicit_opd_reward_mean": 0.006053729122212383,
      "relu_gated_reward_mean": 0.11547812669651236,
      "correct_trajectory_fraction": 1.0,
      "extra_loss_hyperparameters": 0.0,
      "loss": -0.007589046422977905,
      "policy_entropy": 1.7898752038832115
    },
    "fidelity": "mechanism reproduction on a candidate-policy model"
  },
  "history": [
    {
      "step": 1.0,
      "loss": 2.0816681711721685e-17,
      "accuracy": 0.5390625,
      "mean_reward": 0.6292284645371297,
      "entropy": 1.7917581148861492,
      "kl_from_reference": 1.354335906343541e-06
    },
    {
      "step": 10.0,
      "loss": -0.0002685231550778877,
      "accuracy": 0.6328125,
      "mean_reward": 0.7032413272184482,
      "entropy": 1.7917435996172046,
      "kl_from_reference": 1.5869604850670746e-05
    },
    {
      "step": 20.0,
      "loss": -0.0005377701272458769,
      "accuracy": 0.6484375,
      "mean_reward": 0.7199937913291299,
      "entropy": 1.7916816449966668,
      "kl_from_reference": 7.782422538788544e-05
    },
    {
      "step": 30.0,
      "loss": -0.0021128391399234184,
      "accuracy": 0.578125,
      "mean_reward": 0.6605260986340293,
      "entropy": 1.7915165597060985,
      "kl_from_reference": 0.0002429095159563772
    },
    {
      "step": 40.0,
      "loss": -0.0,
      "accuracy": 0.5703125,
      "mean_reward": 0.6541002914547227,
      "entropy": 1.7913903432825118,
      "kl_from_reference": 0.0003691259395429718
    },
    {
      "step": 50.0,
      "loss": -0.0002436910767334155,
      "accuracy": 0.6640625,
      "mean_reward": 0.733397346178058,
      "entropy": 1.791304693019279,
      "kl_from_reference": 0.00045477620277691825
    },
    {
      "step": 60.0,
      "loss": -0.00642302452869048,
      "accuracy": 0.6640625,
      "mean_reward": 0.7317825159641633,
      "entropy": 1.7910150815255195,
      "kl_from_reference": 0.0007443876965351217
    },
    {
      "step": 70.0,
      "loss": -0.0,
      "accuracy": 0.6796875,
      "mean_reward": 0.7407765142824253,
      "entropy": 1.7907664101369623,
      "kl_from_reference": 0.000993059085092456
    },
    {
      "step": 80.0,
      "loss": -0.00020483888004824645,
      "accuracy": 0.671875,
      "mean_reward": 0.7362776775028683,
      "entropy": 1.7906247812132121,
      "kl_from_reference": 0.0011346880088429193
    },
    {
      "step": 90.0,
      "loss": -0.006949358679262031,
      "accuracy": 0.6640625,
      "mean_reward": 0.7317825159641633,
      "entropy": 1.7902298575506275,
      "kl_from_reference": 0.0015296116714267147
    },
    {
      "step": 100.0,
      "loss": -0.007589046422977905,
      "accuracy": 0.6640625,
      "mean_reward": 0.733397346178058,
      "entropy": 1.7899356503961907,
      "kl_from_reference": 0.0018238188258643837
    }
  ],
  "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/2608.24696-opdvr/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"
  }
}