{
 "generated": "2026-11-20",
 "sha": "",
 "node": {
  "date": "2026-10-13",
  "updated": "2026-10-13",
  "kind": "report",
  "category": "Agent 工程",
  "group": "agents",
  "page": "/deepdive/grader-read-only/",
  "summary": "作弊不是边角料：METR 的 o3 在 RE-Bench 上 30.4% 的运行作弊，GPT-5.6 Sol 的时间视野因作弊记账方式不同差约 24 倍。对比 Karpathy autoresearch 的文字约定与 METR 的强制隔离，给出九条评分器隔离清单，并回头审自己的实验",
  "version": 1,
  "tags": [
   "reward-hacking",
   "autoresearch",
   "evals",
   "verification",
   "agent-engineering",
   "metr",
   "benchmark-integrity",
   "harness"
  ],
  "langs": [
   "zh",
   "en"
  ],
  "words": 15299,
  "n_sections": 10,
  "dd_id": "DD·0153",
  "related": [
   "ai-native-rd",
   "agent-engineering-evolution",
   "loop-engineering",
   "ai-rd-automation-index"
  ],
  "id": "dd:grader-read-only",
  "type": "Article",
  "article": "grader-read-only",
  "name": "评分器不可写：长时程自动实验里的 reward hacking 与度量失效",
  "title": "评分器不可写：长时程自动实验里的 reward hacking 与度量失效",
  "slug": "grader-read-only",
  "kind_zh": "深度报告",
  "entry": "agent",
  "n_ents": 6,
  "n_entities": 7
 },
 "page": "https://xiaopingfeng.com/graph/e/dd/grader-read-only/",
 "out": [
  {
   "type": "covers",
   "sentence": "评分器不可写：长时程自动实验里的 reward hacking 与度量失效讲到METR",
   "direction": "out",
   "id": "org:metr",
   "name": "METR",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "评分器不可写：长时程自动实验里的 reward hacking 与度量失效讲到Anthropic",
   "direction": "out",
   "id": "org:anthropic",
   "name": "Anthropic",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "评分器不可写：长时程自动实验里的 reward hacking 与度量失效讲到Shopify",
   "direction": "out",
   "id": "org:shopify",
   "name": "Shopify",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "评分器不可写：长时程自动实验里的 reward hacking 与度量失效讲到OpenAI",
   "direction": "out",
   "id": "org:openai",
   "name": "OpenAI",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "评分器不可写：长时程自动实验里的 reward hacking 与度量失效讲到Claude",
   "direction": "out",
   "id": "app:claude",
   "name": "Claude",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "评分器不可写：长时程自动实验里的 reward hacking 与度量失效讲到GitHub",
   "direction": "out",
   "id": "org:github",
   "name": "GitHub",
   "role": "mention",
   "status": "agent",
   "evidence": []
  }
 ],
 "in": [
  {
   "type": "part_of",
   "sentence": "评分器能被写到，数字就不算数属于评分器不可写：长时程自动实验里的 reward hacking 与度量失效",
   "direction": "in",
   "id": "ddsec:grader-read-only-sec-tldr",
   "name": "评分器能被写到，数字就不算数",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "同一批运行，三个答案属于评分器不可写：长时程自动实验里的 reward hacking 与度量失效",
   "direction": "in",
   "id": "ddsec:grader-read-only-open",
   "name": "同一批运行，三个答案",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "作弊有多常见属于评分器不可写：长时程自动实验里的 reward hacking 与度量失效",
   "direction": "in",
   "id": "ddsec:grader-read-only-sec-01",
   "name": "作弊有多常见",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "评分器有 五个面属于评分器不可写：长时程自动实验里的 reward hacking 与度量失效",
   "direction": "in",
   "id": "ddsec:grader-read-only-sec-02",
   "name": "评分器有 五个面",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "约定与强制属于评分器不可写：长时程自动实验里的 reward hacking 与度量失效",
   "direction": "in",
   "id": "ddsec:grader-read-only-sec-03",
   "name": "约定与强制",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "没人作弊，数字也不对属于评分器不可写：长时程自动实验里的 reward hacking 与度量失效",
   "direction": "in",
   "id": "ddsec:grader-read-only-sec-04",
   "name": "没人作弊，数字也不对",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "当作弊 决定曲线属于评分器不可写：长时程自动实验里的 reward hacking 与度量失效",
   "direction": "in",
   "id": "ddsec:grader-read-only-sec-05",
   "name": "当作弊 决定曲线",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "评分器 隔离清单属于评分器不可写：长时程自动实验里的 reward hacking 与度量失效",
   "direction": "in",
   "id": "ddsec:grader-read-only-sec-06",
   "name": "评分器 隔离清单",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "回头审 自己的实验属于评分器不可写：长时程自动实验里的 reward hacking 与度量失效",
   "direction": "in",
   "id": "ddsec:grader-read-only-sec-07",
   "name": "回头审 自己的实验",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "这篇 没证明什么属于评分器不可写：长时程自动实验里的 reward hacking 与度量失效",
   "direction": "in",
   "id": "ddsec:grader-read-only-sec-08",
   "name": "这篇 没证明什么",
   "status": "derived",
   "evidence": []
  }
 ]
}