{
  "algorithm": "rstg",
  "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": 0.739174087613786,
    "entropy": 1.6401155901760665,
    "kl_from_reference": 0.15164387904598867
  },
  "relative_accuracy": 2.48,
  "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": 512,
    "teacher_prefill_calls": 512,
    "online_teacher_calls": 0,
    "drift_events": 0,
    "critic_updates": 0,
    "rollout_policy_refreshes": 7,
    "last_diagnostics": {
      "mechanism_id": 7.0,
      "negative_group": 1.0,
      "selected_tokens": 2.0,
      "advantage_std": 1.0187191356251568,
      "loss": -0.49920137749116966,
      "policy_entropy": 1.6223077343212537
    },
    "fidelity": "mechanism reproduction on a candidate-policy model"
  },
  "history": [
    {
      "step": 1.0,
      "loss": 6.938893903907228e-17,
      "accuracy": 0.5390625,
      "mean_reward": 0.6292284645371297,
      "entropy": 1.7917218396604269,
      "kl_from_reference": 3.762956162878595e-05
    },
    {
      "step": 12.0,
      "loss": -0.827707947448025,
      "accuracy": 0.671875,
      "mean_reward": 0.7332122087855273,
      "entropy": 1.787255948643643,
      "kl_from_reference": 0.004503520578410614
    },
    {
      "step": 24.0,
      "loss": -0.08339670165821211,
      "accuracy": 0.6640625,
      "mean_reward": 0.7331630184926088,
      "entropy": 1.7770447577228392,
      "kl_from_reference": 0.014714711499215402
    },
    {
      "step": 36.0,
      "loss": -1.1738151729855129,
      "accuracy": 0.6796875,
      "mean_reward": 0.7448392306205689,
      "entropy": 1.7584155728893678,
      "kl_from_reference": 0.03334389633268597
    },
    {
      "step": 48.0,
      "loss": -0.5171249336639698,
      "accuracy": 0.5390625,
      "mean_reward": 0.65004577576101,
      "entropy": 1.7373087252039534,
      "kl_from_reference": 0.05445074401810225
    },
    {
      "step": 60.0,
      "loss": -0.23474257920936875,
      "accuracy": 0.4921875,
      "mean_reward": 0.6159888140315496,
      "entropy": 1.7114822180962277,
      "kl_from_reference": 0.08027725112582723
    },
    {
      "step": 72.0,
      "loss": -0.3483056725270202,
      "accuracy": 0.515625,
      "mean_reward": 0.6332451022385487,
      "entropy": 1.6925416543411727,
      "kl_from_reference": 0.0992178148808821
    },
    {
      "step": 84.0,
      "loss": -0.4087086368670031,
      "accuracy": 0.5703125,
      "mean_reward": 0.6722060847364586,
      "entropy": 1.6828736770978747,
      "kl_from_reference": 0.10888579212417986
    },
    {
      "step": 96.0,
      "loss": -0.20721495829769995,
      "accuracy": 0.6484375,
      "mean_reward": 0.7251343062413547,
      "entropy": 1.6684793358886085,
      "kl_from_reference": 0.12328013333344585
    },
    {
      "step": 108.0,
      "loss": -0.3761732688899633,
      "accuracy": 0.6640625,
      "mean_reward": 0.7317825159641633,
      "entropy": 1.6545482559116707,
      "kl_from_reference": 0.13721121331038388
    },
    {
      "step": 120.0,
      "loss": -0.49920137749116966,
      "accuracy": 0.6796875,
      "mean_reward": 0.739174087613786,
      "entropy": 1.6401155901760665,
      "kl_from_reference": 0.15164387904598867
    }
  ],
  "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/2608.00782-rstg/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"
  }
}