{
  "protocol": {
    "dataset": "arithmetic-generate",
    "maximum_examples": 48,
    "seeds": [
      42,
      43,
      44
    ],
    "warmup_steps": 20,
    "rl_steps": 6,
    "group_size": 4,
    "model": "30378-parameter GRU causal LM",
    "tokenizer": "auditable character tokenizer",
    "verifier": "exact final numeric answer + format/length diagnostics"
  },
  "results": {
    "ipo": {
      "baseline_accuracy_mean": 0.0,
      "final_accuracy_mean": 0.0,
      "baseline_mean_reward": 0.1,
      "final_mean_reward": 0.1,
      "final_format_rate_mean": 0.0
    },
    "simpo": {
      "baseline_accuracy_mean": 0.0,
      "final_accuracy_mean": 0.0,
      "baseline_mean_reward": 0.1,
      "final_mean_reward": 0.1333333333,
      "final_mean_reward_std": 0.0471404521,
      "final_format_rate_mean": 0.3333333333,
      "final_format_rate_std": 0.4714045208
    },
    "luspo": {
      "baseline_accuracy_mean": 0.0,
      "final_accuracy_mean": 0.0,
      "baseline_mean_reward": 0.1,
      "final_mean_reward": 0.1,
      "final_format_rate_mean": 0.0
    },
    "coba-rl": {
      "baseline_accuracy_mean": 0.0,
      "final_accuracy_mean": 0.0,
      "baseline_mean_reward": 0.1,
      "final_mean_reward": 0.1,
      "final_format_rate_mean": 0.0
    }
  },
  "conclusion": "All four methods executed token-level free rollouts and multi-seed updates. At this deliberately small budget no method improved exact generation; SimPO improved format/reward only. This negative result is retained rather than being described as a paper-level reproduction gain.",
  "schema_version": 2,
  "manifest_ref": "experiments:free-generation-post-training-seeds42-44",
  "evaluation_protocol": {
    "tier": "l1_mechanism",
    "seeds": [
      42,
      43,
      44
    ],
    "formal_comparison": true,
    "claim_policy": "formal multi-seed comparison"
  },
  "provenance": {
    "artifact_path": "docs/experiments/free-generation-post-training-seeds42-44.json",
    "historical_migration": "historical-metrics-v2-2026-08-09",
    "original_code_commit": "not recorded",
    "dataset_fingerprint": "not recorded in historical artifact"
  }
}