{
  "purpose": "Describe settings separately; not an executable evaluation config",
  "explicit_historical_evaluation_settings": {
    "verifier_repeats_per_proof": 8,
    "max_new_tokens": 65536,
    "temperature": 1.0,
    "top_p": null,
    "top_k": null
  },
  "unchanged_checkpoint_generation_defaults": {
    "do_sample": true,
    "temperature": 1.0,
    "top_p": 0.95,
    "top_k": 20,
    "eos_token_id": [248046, 248044],
    "pad_token_id": 248044
  },
  "step_index_base": 0,
  "no_error_sentinel": -1,
  "verifier_prompt": "prompts/proof_verifier.md",
  "notes": [
    "Historical API configuration did not explicitly fix top_p/top_k; runtime defaults require confirmation for exact reproduction.",
    "Checkpoint default sampling values are not asserted to be the historical serving parameters.",
    "The bundled full model includes MTP and vision tensors; this staging step does not convert weights or strip model components.",
    "The intended pessimistic decision requires eight valid no-error judgments; legacy parse-failure handling belongs to the separately versioned evaluator."
  ]
}