{
  "schema_version": 2,
  "manifest_ref": "post-training:pact",
  "algorithm": "pact",
  "dataset": "arithmetic-generate",
  "protocol": {
    "maximum_examples": 48,
    "seeds": [
      42,
      43,
      44
    ],
    "warmup_steps": 20,
    "rl_steps": 3,
    "group_size": 2,
    "device": "cpu",
    "model": "30378-parameter GRU causal LM and 27713-parameter prefix critic",
    "verifier": "exact final numeric answer + format/length diagnostics"
  },
  "baseline": {
    "accuracy": 0.0,
    "mean_reward": 0.1,
    "format_rate": 0.0
  },
  "final": {
    "accuracy": 0.006944444444444444,
    "accuracy_std": 0.009820927516479826,
    "mean_reward": 0.10555555555555551,
    "format_rate": 0.0
  },
  "relative_accuracy": null,
  "mechanism_diagnostics": {
    "update_order": "actor-then-critic",
    "critic_target": "post-actor current-token IS * verifier reward",
    "critic_final_loss_by_seed": [
      0.5486676692962646,
      0.5055667757987976,
      0.5244128704071045
    ],
    "post_actor_ratio_mean_by_seed": [
      0.9804801940917969,
      0.9997652173042297,
      0.9916771650314331
    ]
  },
  "evaluation_protocol": {
    "tier": "l1_mechanism",
    "seeds": [
      42,
      43,
      44
    ],
    "formal_comparison": false,
    "claim_policy": "small-model mechanism validation only; no LLM or paper-level capability claim"
  },
  "provenance": {
    "artifact_path": "docs/experiments/pact-arithmetic-seeds42-44.json",
    "dataset_fingerprint": "e0a4daa74dd67573b73023217b8b4872c50ae499974ab887bbc2dd249bf85173",
    "command": "auto-research post-train --algorithm pact --dataset arithmetic-generate --maximum-examples 48 --steps 3 --group-size 2 --seeds 42,43,44 --offline --device cpu --cpu-threads 2"
  },
  "conclusion": "Three CPU seeds execute actor-then-critic updates. Baseline exact accuracy is zero, so relative improvement is undefined. This tiny-budget result does not demonstrate stable capability gain."
}