{
  "snapshotUTC": "2026-10-15T06:55:03.592078+00:00",
  "source": "https://github.com/DennisKoshta/upbringing/blob/main/analysis/grader_comparison.py",
  "evidence": "https://github.com/DennisKoshta/upbringing/blob/main/analysis/evidence/first-run-step-00024.jsonl.gz",
  "scope": "Only the retained step-24 batch from a stopped first run; grader vulnerability, not demonstrated learned reward hacking.",
  "recomputed": {
    "step": 24,
    "rollouts": 256,
    "multiple_answer_blocks": 172,
    "different_answer_read": 166,
    "old_rewarded": 1,
    "new_rewarded": 1,
    "false_positives": [
      {
        "nums": [
          20,
          93,
          85
        ],
        "target": 12,
        "first_answer": "(20 + 85) / 93 = 1.08651",
        "answer_old_grader_read": "(20 + 85) - 93"
      }
    ],
    "false_negatives": [
      {
        "nums": [
          20,
          62,
          1
        ],
        "target": 82,
        "first_answer": "20 + 1 * 62",
        "answer_old_grader_read": "4000 + (550 x 10) = 10,500 pounds"
      }
    ],
    "share_of_text_after_first_answer": 0.8588700664975601
  }
}