{
  "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": {
    "slic-hf": {
      "accuracy": 0.78125,
      "mean_reward": 0.8073573303541953,
      "kl_from_reference": 0.2511690037013696,
      "calibration_margin": 0.5,
      "off_policy_preferences": true
    },
    "steerlm": {
      "accuracy": 0.8515625,
      "mean_reward": 0.8654098929838473,
      "kl_from_reference": 0.9111972686932943,
      "attribute_dimensions": 4,
      "attribute_conditioned_sft": true
    },
    "spin": {
      "accuracy": 0.859375,
      "mean_reward": 0.8691367545869129,
      "kl_from_reference": 0.12941252077316928,
      "opponent_refreshes": 18,
      "external_preference_labels": 0
    }
  },
  "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:p1-alignment-candidates-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/p1-alignment-candidates-gsm8k-seed42.json",
    "historical_migration": "historical-metrics-v2-2026-08-09",
    "original_code_commit": "not recorded",
    "dataset_fingerprint": "not recorded in historical artifact"
  }
}