{
  "algorithm": "sa-mrpo",
  "dataset": "arithmetic-smoke",
  "baseline": {
    "accuracy": 0.1953125,
    "mean_reward": 0.359689058826365,
    "entropy": 1.791759469222057,
    "kl_from_reference": 0.0
  },
  "final": {
    "accuracy": 0.28125,
    "mean_reward": 0.45564911291653953,
    "entropy": 1.6061716491012128,
    "kl_from_reference": 0.18558782012084224
  },
  "relative_accuracy": 0.44,
  "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": 9.0,
      "unsaturated_weight_max": 0.584838408639106,
      "advantage_std": 0.700885077554041,
      "loss": -0.073674600161796,
      "policy_entropy": 1.582241331905526
    },
    "fidelity": "mechanism reproduction on a candidate-policy model"
  },
  "history": [
    {
      "step": 1.0,
      "loss": 1.3877787807814457e-16,
      "accuracy": 0.3046875,
      "mean_reward": 0.47475757647766553,
      "entropy": 1.7917220358925046,
      "kl_from_reference": 3.743332954974294e-05
    },
    {
      "step": 12.0,
      "loss": -0.0011163807881804967,
      "accuracy": 0.28125,
      "mean_reward": 0.45564911291653953,
      "entropy": 1.790252497725624,
      "kl_from_reference": 0.001506971496430256
    },
    {
      "step": 24.0,
      "loss": -0.10868531904619938,
      "accuracy": 0.28125,
      "mean_reward": 0.45564911291653953,
      "entropy": 1.7821003787462704,
      "kl_from_reference": 0.009659090475785052
    },
    {
      "step": 36.0,
      "loss": -0.075910510915558,
      "accuracy": 0.28125,
      "mean_reward": 0.45564911291653953,
      "entropy": 1.7776298974793718,
      "kl_from_reference": 0.01412957174268352
    },
    {
      "step": 48.0,
      "loss": -0.07976334289962152,
      "accuracy": 0.28125,
      "mean_reward": 0.45564911291653953,
      "entropy": 1.747471068449301,
      "kl_from_reference": 0.04428840077275463
    },
    {
      "step": 60.0,
      "loss": -0.017346140062611565,
      "accuracy": 0.28125,
      "mean_reward": 0.45564911291653953,
      "entropy": 1.71087009689551,
      "kl_from_reference": 0.0808893723265443
    },
    {
      "step": 72.0,
      "loss": -0.045393650040268324,
      "accuracy": 0.28125,
      "mean_reward": 0.45564911291653953,
      "entropy": 1.678667612623207,
      "kl_from_reference": 0.11309185659884886
    },
    {
      "step": 84.0,
      "loss": -0.21770208096614688,
      "accuracy": 0.28125,
      "mean_reward": 0.45564911291653953,
      "entropy": 1.6540078325793628,
      "kl_from_reference": 0.13775163664269302
    },
    {
      "step": 96.0,
      "loss": -0.4683091709086584,
      "accuracy": 0.28125,
      "mean_reward": 0.45564911291653953,
      "entropy": 1.635122467662032,
      "kl_from_reference": 0.15663700156002275
    },
    {
      "step": 108.0,
      "loss": -0.013160237630985446,
      "accuracy": 0.28125,
      "mean_reward": 0.45564911291653953,
      "entropy": 1.6098397302291851,
      "kl_from_reference": 0.18191973899287
    },
    {
      "step": 120.0,
      "loss": -0.073674600161796,
      "accuracy": 0.28125,
      "mean_reward": 0.45564911291653953,
      "entropy": 1.6061716491012128,
      "kl_from_reference": 0.18558782012084224
    }
  ],
  "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/2608.16072-sa-mrpo/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"
  }
}