{
  "algorithm": "grip",
  "dataset": "arithmetic-smoke",
  "baseline": {
    "accuracy": 0.125,
    "mean_reward": 0.3039061471113459,
    "entropy": 1.7917594692220566,
    "kl_from_reference": 0.0
  },
  "final": {
    "accuracy": 0.234375,
    "mean_reward": 0.41366344934761173,
    "entropy": 1.7739840548500698,
    "kl_from_reference": 0.017775414371985654
  },
  "relative_accuracy": 0.875,
  "training": {
    "steps": 100,
    "learning_rate": 0.08,
    "group_size": 4,
    "train_examples": 64,
    "validation_examples": 64,
    "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": {
      "interpolation_alpha_mean": 0.3261351282513234,
      "granular_reward_mean": 0.5637225003191624,
      "interpolated_policy_probability_mean": 0.18281051517029698,
      "extra_rollout_calls": 0.0,
      "loss": -0.022648735369353336,
      "policy_entropy": 1.7527963336144798
    },
    "fidelity": "mechanism reproduction on a candidate-policy model"
  },
  "history": [
    {
      "step": 1.0,
      "loss": 4.163336342344337e-17,
      "accuracy": 0.203125,
      "mean_reward": 0.38792858424974397,
      "entropy": 1.7917563379887897,
      "kl_from_reference": 3.131233264708648e-06
    },
    {
      "step": 10.0,
      "loss": -0.0023404965544135594,
      "accuracy": 0.203125,
      "mean_reward": 0.38792858424974397,
      "entropy": 1.7913069600700986,
      "kl_from_reference": 0.0004525091519563867
    },
    {
      "step": 20.0,
      "loss": -0.012140240826747556,
      "accuracy": 0.203125,
      "mean_reward": 0.38792858424974397,
      "entropy": 1.7904205760105274,
      "kl_from_reference": 0.0013388932115275091
    },
    {
      "step": 30.0,
      "loss": -0.0005694025542456391,
      "accuracy": 0.234375,
      "mean_reward": 0.41366344934761173,
      "entropy": 1.7901847427652884,
      "kl_from_reference": 0.001574726456766615
    },
    {
      "step": 40.0,
      "loss": -0.005671426003104801,
      "accuracy": 0.265625,
      "mean_reward": 0.43943013355655564,
      "entropy": 1.7889555244877868,
      "kl_from_reference": 0.002803944734268184
    },
    {
      "step": 50.0,
      "loss": -0.011292308245923546,
      "accuracy": 0.25,
      "mean_reward": 0.4265063625097134,
      "entropy": 1.7870091931066174,
      "kl_from_reference": 0.004750276115437812
    },
    {
      "step": 60.0,
      "loss": -0.0026577591438083845,
      "accuracy": 0.265625,
      "mean_reward": 0.43943013355655564,
      "entropy": 1.786376308492751,
      "kl_from_reference": 0.005383160729304236
    },
    {
      "step": 70.0,
      "loss": -0.03755439669095526,
      "accuracy": 0.265625,
      "mean_reward": 0.43943013355655564,
      "entropy": 1.7839775790411005,
      "kl_from_reference": 0.007781890180953859
    },
    {
      "step": 80.0,
      "loss": -0.027225501463915926,
      "accuracy": 0.265625,
      "mean_reward": 0.43943013355655564,
      "entropy": 1.7809596727552495,
      "kl_from_reference": 0.010799796466805737
    },
    {
      "step": 90.0,
      "loss": -0.028267656845137788,
      "accuracy": 0.25,
      "mean_reward": 0.4265063625097134,
      "entropy": 1.77781811140585,
      "kl_from_reference": 0.013941357816204382
    },
    {
      "step": 100.0,
      "loss": -0.022648735369353336,
      "accuracy": 0.234375,
      "mean_reward": 0.41366344934761173,
      "entropy": 1.7739840548500698,
      "kl_from_reference": 0.017775414371985654
    }
  ],
  "evaluation_protocol": {
    "tier": "l1_mechanism",
    "seeds": [
      42
    ],
    "formal_comparison": false,
    "claim_policy": "single/few-seed mechanism result; do not claim a stable improvement"
  },
  "schema_version": 2,
  "manifest_ref": "post-training:grip-arithmetic-smoke-seed42",
  "provenance": {
    "artifact_path": "docs/post-training/2608.25583-grip/metrics/grip-arithmetic-smoke-seed42.json",
    "dataset_fingerprint": "not recorded in historical artifact",
    "historical_migration": "historical-metrics-v2-2026-08-09",
    "original_code_commit": "not recorded"
  }
}
