{
  "product_version": "5.1.0",
  "product_name": "Axiom Science 4.5",
  "evaluation_date": "2026-10-03",
  "questions": 32,
  "repeats": 3,
  "total_runs": 96,
  "original_runs": 87,
  "added_runs": 9,
  "evaluation_source_commit": "5ff9110bc1c2da9cbb45e5a78091a4fa5ff87017",
  "evaluation_version": "4.11.0-rc.6",
  "models": {
    "main": "GLM-5.3-Flash",
    "review": "GLM-5.3-Flash",
    "judge": "GLM-5.2"
  },
  "metrics": {
    "档均分": 82.43,
    "幻觉率": 0.0,
    "硬失败率": 0.1875,
    "交付完整率": 0.9896
  },
  "quality_gate_passed": false,
  "quality_gate_thresholds": {
    "mean_min": 70,
    "hallucination_max": 0,
    "hard_failure_max": 0.1,
    "delivery_min": 0.95
  },
  "hard_failure_runs": 18,
  "execution_successful_runs": 95,
  "empty_output_runs": 1,
  "user_accepted_known_failures": true,
  "acceptance_date": "2026-10-04",
  "acceptance_scope": "冻结96次科研评测 + 最终完整软件回归，仅本次v5.1.0",
  "explanations": [
    "均分采用原评测定义，不把运行成功等同于科研任务正确。",
    "硬失败率18.75%高于默认≤10%门槛；用户接受本轮已知失败后发布，原阈值保持。",
    "评测在冻结版本上完成，最终发布另通过完整软件回归。",
    "四个内置示例是教学记录，Samosa为模拟数据，补充示例未追加进96次评测。"
  ],
  "score_definition": "每题三次运行得分的中位数，再对32题取平均；沿用原评测定义。",
  "grading_coverage_note": "reuse-01-islet-t2d 完整判分覆盖2/3；96次运行包含1次空输出，执行完成不代表科研正确。",
  "final_release_commit": "f61c6526e24ddcf3af70d7a31af362079d5e05e3",
  "software_regression": {
    "date": "2026-10-04",
    "backend_passed": 1261,
    "backend_failed": 0,
    "backend_skipped": 8,
    "frontend_passed": 15,
    "frontend_build_passed": true
  }
}