{
  "algorithm": "iso-rlvr",
  "dataset": "arithmetic-smoke",
  "baseline": {
    "accuracy": 0.1953125,
    "mean_reward": 0.359689058826365,
    "entropy": 1.791759469222057,
    "kl_from_reference": 0.0
  },
  "final": {
    "accuracy": 0.5390625,
    "mean_reward": 0.65004577576101,
    "entropy": 1.689655629751765,
    "kl_from_reference": 0.10210383947028875
  },
  "relative_accuracy": 1.76,
  "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": 4.0,
      "fixed_spectrum_rank": 6.0,
      "advantage_std": 0.9999968104512983,
      "loss": -0.030899232702374824,
      "policy_entropy": 1.7056732824673746
    },
    "fidelity": "mechanism reproduction on a candidate-policy model"
  },
  "history": [
    {
      "step": 1.0,
      "loss": 6.938893903907228e-17,
      "accuracy": 0.5390625,
      "mean_reward": 0.6292284645371297,
      "entropy": 1.7917261262397064,
      "kl_from_reference": 3.334298234856459e-05
    },
    {
      "step": 12.0,
      "loss": -0.05295234934740736,
      "accuracy": 0.6640625,
      "mean_reward": 0.7315481882787143,
      "entropy": 1.789902449781725,
      "kl_from_reference": 0.0018570194403299946
    },
    {
      "step": 24.0,
      "loss": -0.058287311672336195,
      "accuracy": 0.65625,
      "mean_reward": 0.7282336768472688,
      "entropy": 1.784868892098176,
      "kl_from_reference": 0.006890577123878398
    },
    {
      "step": 36.0,
      "loss": -0.05732615665364468,
      "accuracy": 0.6484375,
      "mean_reward": 0.7237977073432597,
      "entropy": 1.7774410259638707,
      "kl_from_reference": 0.01431844325818352
    },
    {
      "step": 48.0,
      "loss": -0.038508882556109736,
      "accuracy": 0.484375,
      "mean_reward": 0.6099019169926555,
      "entropy": 1.7657608682227948,
      "kl_from_reference": 0.025998600999260216
    },
    {
      "step": 60.0,
      "loss": -0.1825592717955321,
      "accuracy": 0.546875,
      "mean_reward": 0.655827633812385,
      "entropy": 1.755962242238253,
      "kl_from_reference": 0.03579722698380193
    },
    {
      "step": 72.0,
      "loss": -0.2837042791947705,
      "accuracy": 0.5703125,
      "mean_reward": 0.6722921703880933,
      "entropy": 1.7443243931574544,
      "kl_from_reference": 0.04743507606459968
    },
    {
      "step": 84.0,
      "loss": -0.11307692667667202,
      "accuracy": 0.5234375,
      "mean_reward": 0.6389981814631177,
      "entropy": 1.7295073641450287,
      "kl_from_reference": 0.06225210507702678
    },
    {
      "step": 96.0,
      "loss": -0.2988463191998455,
      "accuracy": 0.4765625,
      "mean_reward": 0.6043193950593897,
      "entropy": 1.7060207317186438,
      "kl_from_reference": 0.08573873750341134
    },
    {
      "step": 108.0,
      "loss": -0.31618606553301465,
      "accuracy": 0.4921875,
      "mean_reward": 0.6159888140315496,
      "entropy": 1.6969426932150937,
      "kl_from_reference": 0.09481677600696087
    },
    {
      "step": 120.0,
      "loss": -0.030899232702374824,
      "accuracy": 0.5390625,
      "mean_reward": 0.65004577576101,
      "entropy": 1.689655629751765,
      "kl_from_reference": 0.10210383947028875
    }
  ],
  "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/2607.19331-iso-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"
  }
}