{
  "algorithm": "weak-guide-rlvr",
  "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": 0.7398431161754653,
    "entropy": 1.786694365533035,
    "kl_from_reference": 0.005065103689019571
  },
  "relative_accuracy": 2.48,
  "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": {
      "weak_prefix_fraction": 0.25,
      "prefix_surprise_mean": 1.799541871860641,
      "guided_entropy": 1.7870506939062338,
      "loss": -0.03471125580868825,
      "policy_entropy": 1.7871632734216112
    },
    "fidelity": "mechanism reproduction on a candidate-policy model"
  },
  "history": [
    {
      "step": 1.0,
      "loss": -1.3877787807814457e-17,
      "accuracy": 0.5,
      "mean_reward": 0.5962298893894342,
      "entropy": 1.7917477401385715,
      "kl_from_reference": 1.1729083484112992e-05
    },
    {
      "step": 10.0,
      "loss": -0.0006461321230525341,
      "accuracy": 0.578125,
      "mean_reward": 0.6595187519342955,
      "entropy": 1.7917255708696165,
      "kl_from_reference": 3.389835243826678e-05
    },
    {
      "step": 20.0,
      "loss": 7.12534206414601e-05,
      "accuracy": 0.6484375,
      "mean_reward": 0.7194847870111635,
      "entropy": 1.7915897332234891,
      "kl_from_reference": 0.00016973599856756006
    },
    {
      "step": 30.0,
      "loss": -0.007255242629957678,
      "accuracy": 0.5625,
      "mean_reward": 0.6483908604683593,
      "entropy": 1.7910476217347355,
      "kl_from_reference": 0.0007118474873196608
    },
    {
      "step": 40.0,
      "loss": 0.0004774604705596575,
      "accuracy": 0.5390625,
      "mean_reward": 0.6298203892915948,
      "entropy": 1.7907190070807577,
      "kl_from_reference": 0.0010404621412980525
    },
    {
      "step": 50.0,
      "loss": -0.005425862553499916,
      "accuracy": 0.6640625,
      "mean_reward": 0.7315481882787143,
      "entropy": 1.7904756326428042,
      "kl_from_reference": 0.001283836579250905
    },
    {
      "step": 60.0,
      "loss": -0.016762902846879132,
      "accuracy": 0.6484375,
      "mean_reward": 0.7261273225271052,
      "entropy": 1.789951165131678,
      "kl_from_reference": 0.0018083040903762386
    },
    {
      "step": 70.0,
      "loss": -0.017962597057663843,
      "accuracy": 0.5390625,
      "mean_reward": 0.65004577576101,
      "entropy": 1.7890586214470694,
      "kl_from_reference": 0.002700847774986081
    },
    {
      "step": 80.0,
      "loss": -0.0012378690129591363,
      "accuracy": 0.5078125,
      "mean_reward": 0.6276164415495833,
      "entropy": 1.78842350186273,
      "kl_from_reference": 0.003335967359325886
    },
    {
      "step": 90.0,
      "loss": 0.000979082421118607,
      "accuracy": 0.65625,
      "mean_reward": 0.7305749634100966,
      "entropy": 1.787637662389842,
      "kl_from_reference": 0.004121806832213417
    },
    {
      "step": 100.0,
      "loss": -0.03471125580868825,
      "accuracy": 0.6796875,
      "mean_reward": 0.7398431161754653,
      "entropy": 1.786694365533035,
      "kl_from_reference": 0.005065103689019571
    }
  ],
  "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/2608.27420-weak-guide-rlvr/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"
  }
}