{
  "protocol": {
    "dataset": "OpenAI GSM8K official train/test JSONL",
    "adapter": "gsm8k-candidate",
    "train_examples": 512,
    "validation_examples": 128,
    "steps": 300,
    "learning_rate": 0.08,
    "group_size": 4,
    "seed": 42,
    "baseline": "untrained candidate policy",
    "fidelity": "mechanism reproduction on a candidate-policy model"
  },
  "baseline": {
    "accuracy": 0.1640625,
    "mean_reward": 0.31261065795890286,
    "kl_from_reference": 0.0
  },
  "methods": {
    "constitutional-ai": {
      "accuracy": 0.84375,
      "mean_reward": 0.8616564155478119,
      "kl_from_reference": 1.0213545205106576,
      "cumulative_critiques": 300,
      "cumulative_revisions": 69,
      "human_preference_labels": 0
    },
    "rrhf": {
      "accuracy": 0.8125,
      "mean_reward": 0.8400585043733009,
      "kl_from_reference": 0.8344118615935469,
      "ranked_responses_per_group": 6,
      "last_ranking_pairs": 15,
      "last_ranking_violations": 6
    },
    "raft": {
      "accuracy": 0.84375,
      "mean_reward": 0.8616564155478119,
      "kl_from_reference": 0.8788547637092913,
      "sampled_responses_per_update": 4,
      "kept_responses_per_update": 1,
      "kept_fraction": 0.25
    }
  },
  "interpretation": "All gains are relative to the same untrained candidate policy. They validate objective mechanics and are not free-generation or original-paper model results.",
  "schema_version": 2,
  "manifest_ref": "experiments:p0-missing-post-training-gsm8k-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/experiments/p0-missing-post-training-gsm8k-seed42.json",
    "historical_migration": "historical-metrics-v2-2026-08-09",
    "original_code_commit": "not recorded",
    "dataset_fingerprint": "not recorded in historical artifact"
  }
}