{
  "algorithm": "v-rubrics",
  "dataset": "arithmetic-smoke",
  "baseline": {
    "accuracy": 0.125,
    "mean_reward": 0.3039061471113459,
    "entropy": 1.7917594692220566,
    "kl_from_reference": 0.0
  },
  "final": {
    "accuracy": 0.21875,
    "mean_reward": 0.40066753689370793,
    "entropy": 1.5197399863907717,
    "kl_from_reference": 0.2720194828312833
  },
  "relative_accuracy": 0.75,
  "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": {
      "visual_faithfulness_credit": 0.25,
      "reasoning_consistency_credit": 0.8752011827573534,
      "instruction_following_credit": 0.7083333333333334,
      "prefix_localized_fraction": 0.5,
      "loss": 0.31793914631363795,
      "policy_entropy": 1.4066122987983527
    },
    "fidelity": "mechanism reproduction on a candidate-policy model"
  },
  "history": [
    {
      "step": 1.0,
      "loss": -3.3306690738754696e-16,
      "accuracy": 0.203125,
      "mean_reward": 0.38792858424974397,
      "entropy": 1.791569788357861,
      "kl_from_reference": 0.00018968086419379258
    },
    {
      "step": 10.0,
      "loss": -0.11768317690034491,
      "accuracy": 0.203125,
      "mean_reward": 0.38792858424974397,
      "entropy": 1.7825461083343201,
      "kl_from_reference": 0.00921336088773493
    },
    {
      "step": 20.0,
      "loss": -0.1961297923448979,
      "accuracy": 0.203125,
      "mean_reward": 0.38792858424974397,
      "entropy": 1.7551300348633236,
      "kl_from_reference": 0.03662943435873147
    },
    {
      "step": 30.0,
      "loss": -0.32483954062062526,
      "accuracy": 0.203125,
      "mean_reward": 0.38792858424974397,
      "entropy": 1.7418067505617063,
      "kl_from_reference": 0.049952718660348894
    },
    {
      "step": 40.0,
      "loss": -0.2788941382734464,
      "accuracy": 0.203125,
      "mean_reward": 0.38792858424974397,
      "entropy": 1.7237846793161862,
      "kl_from_reference": 0.06797478990586889
    },
    {
      "step": 50.0,
      "loss": -0.05896152558426701,
      "accuracy": 0.203125,
      "mean_reward": 0.38792858424974397,
      "entropy": 1.6843746225594334,
      "kl_from_reference": 0.10738484666262132
    },
    {
      "step": 60.0,
      "loss": -0.5550942466684874,
      "accuracy": 0.203125,
      "mean_reward": 0.38792858424974397,
      "entropy": 1.6583300797955673,
      "kl_from_reference": 0.13342938942648772
    },
    {
      "step": 70.0,
      "loss": -0.16083136968914324,
      "accuracy": 0.21875,
      "mean_reward": 0.40066753689370793,
      "entropy": 1.6279132926896438,
      "kl_from_reference": 0.16384617653241132
    },
    {
      "step": 80.0,
      "loss": 0.12591654182946188,
      "accuracy": 0.21875,
      "mean_reward": 0.40066753689370793,
      "entropy": 1.5904438130122498,
      "kl_from_reference": 0.20131565620980532
    },
    {
      "step": 90.0,
      "loss": -0.6288704148981783,
      "accuracy": 0.21875,
      "mean_reward": 0.40066753689370793,
      "entropy": 1.5360396789254165,
      "kl_from_reference": 0.25571979029663905
    },
    {
      "step": 100.0,
      "loss": 0.31793914631363795,
      "accuracy": 0.21875,
      "mean_reward": 0.40066753689370793,
      "entropy": 1.5197399863907717,
      "kl_from_reference": 0.2720194828312833
    }
  ],
  "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:v-rubrics-arithmetic-smoke-seed42",
  "provenance": {
    "artifact_path": "docs/post-training/2608.25580-v-rubrics/metrics/v-rubrics-arithmetic-smoke-seed42.json",
    "dataset_fingerprint": "not recorded in historical artifact",
    "historical_migration": "historical-metrics-v2-2026-08-09",
    "original_code_commit": "not recorded"
  }
}