{
  "schemaVersion": "0.11.0",
  "canonical": "https://www.pystone.net/notes/proxy-metric-distortion-campbell-goodhart-literature/",
  "atlas": "https://www.pystone.net/?node=proxy-metric-distortion-campbell-goodhart-literature#knowledge-atlas",
  "markdown": "https://www.pystone.net/notes/proxy-metric-distortion-campbell-goodhart-literature.md",
  "context": "https://www.pystone.net/notes/proxy-metric-distortion-campbell-goodhart-literature.context.json",
  "knowledgeVersion": "224c990773de.5fa8af6e39fa",
  "build": {
    "siteCommit": "224c990773de166d23a886306577dd90379529ce",
    "notesCommit": "5fa8af6e39fa3891d1b9b4832bfa6c4e0ecaaf0a",
    "builtAt": "1970-01-01T00:00:00.000Z",
    "version": "224c990773de.5fa8af6e39fa"
  },
  "id": "note:proxy-metric-distortion-campbell-goodhart-literature",
  "slug": "proxy-metric-distortion-campbell-goodhart-literature",
  "title": "代理指标失真：Campbell 定律与 Goodhart 类效应文献摘录",
  "type": "note",
  "visibility": "public",
  "idStability": "rename-stable",
  "author": {
    "name": "Perrin Yong",
    "profile": "https://www.pystone.net/profile/"
  },
  "publisher": {
    "name": "Perrin Yong",
    "profile": "https://www.pystone.net/profile/"
  },
  "aliases": [],
  "summary": "代理指标失真：Campbell 定律与 Goodhart 类效应文献摘录",
  "contentRole": "unspecified",
  "isMoc": false,
  "mocRecognition": "none",
  "generated": false,
  "attribution": "unspecified",
  "domain": "99-资料来源、摘录与归档",
  "tags": [],
  "mocs": [],
  "contentHash": "037f0356229aaf3f0e9ca3a14a610c3b97462f93fd2d1d046d1a76dd7c957ab6",
  "assets": [],
  "headings": [
    {
      "depth": 1,
      "text": "代理指标失真：Campbell 定律与 Goodhart 类效应文献摘录",
      "anchor": "代理指标失真campbell-定律与-goodhart-类效应文献摘录",
      "citation": "https://www.pystone.net/notes/proxy-metric-distortion-campbell-goodhart-literature/#%E4%BB%A3%E7%90%86%E6%8C%87%E6%A0%87%E5%A4%B1%E7%9C%9Fcampbell-%E5%AE%9A%E5%BE%8B%E4%B8%8E-goodhart-%E7%B1%BB%E6%95%88%E5%BA%94%E6%96%87%E7%8C%AE%E6%91%98%E5%BD%95"
    },
    {
      "depth": 2,
      "text": "Donald T. Campbell：Assessing the Impact of Planned Social Change",
      "anchor": "donald-t-campbellassessing-the-impact-of-planned-social-change",
      "citation": "https://www.pystone.net/notes/proxy-metric-distortion-campbell-goodhart-literature/#donald-t-campbellassessing-the-impact-of-planned-social-change"
    },
    {
      "depth": 3,
      "text": "研究对象与论证",
      "anchor": "研究对象与论证",
      "citation": "https://www.pystone.net/notes/proxy-metric-distortion-campbell-goodhart-literature/#%E7%A0%94%E7%A9%B6%E5%AF%B9%E8%B1%A1%E4%B8%8E%E8%AE%BA%E8%AF%81"
    },
    {
      "depth": 3,
      "text": "能支持与不能支持什么",
      "anchor": "能支持与不能支持什么",
      "citation": "https://www.pystone.net/notes/proxy-metric-distortion-campbell-goodhart-literature/#%E8%83%BD%E6%94%AF%E6%8C%81%E4%B8%8E%E4%B8%8D%E8%83%BD%E6%94%AF%E6%8C%81%E4%BB%80%E4%B9%88"
    },
    {
      "depth": 2,
      "text": "David Manheim 与 Scott Garrabrant：Categorizing Variants of Goodhart's Law",
      "anchor": "david-manheim-与-scott-garrabrantcategorizing-variants-of-goodharts-law",
      "citation": "https://www.pystone.net/notes/proxy-metric-distortion-campbell-goodhart-literature/#david-manheim-%E4%B8%8E-scott-garrabrantcategorizing-variants-of-goodharts-law"
    },
    {
      "depth": 3,
      "text": "四类失效机制",
      "anchor": "四类失效机制",
      "citation": "https://www.pystone.net/notes/proxy-metric-distortion-campbell-goodhart-literature/#%E5%9B%9B%E7%B1%BB%E5%A4%B1%E6%95%88%E6%9C%BA%E5%88%B6"
    },
    {
      "depth": 3,
      "text": "边界",
      "anchor": "边界",
      "citation": "https://www.pystone.net/notes/proxy-metric-distortion-campbell-goodhart-literature/#%E8%BE%B9%E7%95%8C"
    },
    {
      "depth": 2,
      "text": "对知识网络的作用",
      "anchor": "对知识网络的作用",
      "citation": "https://www.pystone.net/notes/proxy-metric-distortion-campbell-goodhart-literature/#%E5%AF%B9%E7%9F%A5%E8%AF%86%E7%BD%91%E7%BB%9C%E7%9A%84%E4%BD%9C%E7%94%A8"
    }
  ],
  "claims": [],
  "outgoing": [
    {
      "id": "note:software-quality-as-feedback-system",
      "title": "软件质量为什么是反馈系统而不是测试阶段",
      "url": "https://www.pystone.net/notes/software-quality-as-feedback-system/",
      "atlas": "https://www.pystone.net/?node=software-quality-as-feedback-system#knowledge-atlas",
      "label": "软件质量为什么是反馈系统而不是测试阶段",
      "origin": "explicit",
      "humanReviewed": true,
      "context": "本文保存本次知识编译实际使用的两份来源提要。内容以转述为主，不替代原文；综合判断见软件质量为什么是反馈系统而不是测试阶段。",
      "citation": "https://www.pystone.net/notes/proxy-metric-distortion-campbell-goodhart-literature/#%E4%BB%A3%E7%90%86%E6%8C%87%E6%A0%87%E5%A4%B1%E7%9C%9Fcampbell-%E5%AE%9A%E5%BE%8B%E4%B8%8E-goodhart-%E7%B1%BB%E6%95%88%E5%BA%94%E6%96%87%E7%8C%AE%E6%91%98%E5%BD%95"
    },
    {
      "id": "note:campbells-law-metrics-dysfunction",
      "title": "“坎贝尔定律” —— 为什么不能把考核当手段",
      "url": "https://www.pystone.net/notes/campbells-law-metrics-dysfunction/",
      "atlas": "https://www.pystone.net/?node=campbells-law-metrics-dysfunction#knowledge-atlas",
      "label": "坎贝尔定律底座材料",
      "origin": "explicit",
      "humanReviewed": true,
      "context": "不支持：把二手材料中的具体故事自动归给 Campbell。现有坎贝尔定律底座材料所含“术后第 31 天”例子未在这篇原文中核实。",
      "citation": "https://www.pystone.net/notes/proxy-metric-distortion-campbell-goodhart-literature/#%E8%83%BD%E6%94%AF%E6%8C%81%E4%B8%8E%E4%B8%8D%E8%83%BD%E6%94%AF%E6%8C%81%E4%BB%80%E4%B9%88"
    }
  ],
  "incoming": [
    {
      "id": "note:software-quality-as-feedback-system",
      "title": "软件质量为什么是反馈系统而不是测试阶段",
      "url": "https://www.pystone.net/notes/software-quality-as-feedback-system/",
      "atlas": "https://www.pystone.net/?node=software-quality-as-feedback-system#knowledge-atlas",
      "label": "软件质量为什么是反馈系统而不是测试阶段",
      "origin": "explicit",
      "humanReviewed": true,
      "context": "坎贝尔定律底座材料保存了一张二手截图，指出考核压力会诱发对指标的操纵。其方向与Donald T. Campbell 的原始论文一致，但截图中的“术后第 31 天”例子没有在该论文中得到核实，不能作为本页证据。Campbell 讨论的是社会指标进入决策、奖惩和权力分配后，记录方式与被观察过程都会承受腐化压力；他也明确说明当时列举的许多证据主要是轶事，并未证明所",
      "citation": "https://www.pystone.net/notes/software-quality-as-feedback-system/#%E4%BB%A3%E7%90%86%E6%8C%87%E6%A0%87%E4%B8%BA%E4%BD%95%E4%BC%9A%E5%9C%A8%E5%8F%97%E5%8E%8B%E5%90%8E%E5%A4%B1%E7%9C%9Fgoodhart-%E4%B8%8E-campbell-%E4%B8%8D%E6%98%AF%E4%B8%80%E5%8F%A5%E5%8F%A3%E5%8F%B7"
    },
    {
      "id": "note:software-quality-as-feedback-system",
      "title": "软件质量为什么是反馈系统而不是测试阶段",
      "url": "https://www.pystone.net/notes/software-quality-as-feedback-system/",
      "atlas": "https://www.pystone.net/?node=software-quality-as-feedback-system#knowledge-atlas",
      "label": "软件质量为什么是反馈系统而不是测试阶段",
      "origin": "explicit",
      "humanReviewed": true,
      "context": "Manheim 与 Garrabrant 对 Goodhart 类效应的分类进一步说明，“指标被优化后失效”至少包含四种不同机制：",
      "citation": "https://www.pystone.net/notes/software-quality-as-feedback-system/#%E4%BB%A3%E7%90%86%E6%8C%87%E6%A0%87%E4%B8%BA%E4%BD%95%E4%BC%9A%E5%9C%A8%E5%8F%97%E5%8E%8B%E5%90%8E%E5%A4%B1%E7%9C%9Fgoodhart-%E4%B8%8E-campbell-%E4%B8%8D%E6%98%AF%E4%B8%80%E5%8F%A5%E5%8F%A3%E5%8F%B7"
    },
    {
      "id": "note:sources-excerpts-archive",
      "title": "资料来源、摘录与归档",
      "url": "https://www.pystone.net/notes/sources-excerpts-archive/",
      "atlas": "https://www.pystone.net/?node=sources-excerpts-archive#knowledge-atlas",
      "label": "资料来源、摘录与归档",
      "origin": "explicit",
      "humanReviewed": true,
      "context": "01-外部文章与转载摘录中的“代理指标失真：Campbell 定律与 Goodhart 类效应文献摘录”导航项",
      "citation": "https://www.pystone.net/notes/sources-excerpts-archive/#01-%E5%A4%96%E9%83%A8%E6%96%87%E7%AB%A0%E4%B8%8E%E8%BD%AC%E8%BD%BD%E6%91%98%E5%BD%95"
    }
  ],
  "contentMarkdown": "# 代理指标失真：Campbell 定律与 Goodhart 类效应文献摘录\n\n> [!source] 来源摘录\n> 本文保存本次知识编译实际使用的两份来源提要。内容以转述为主，不替代原文；综合判断见[[10-计算机、信息技术与工程/03-软件工程与质量保障/软件质量为什么是反馈系统而不是测试阶段|软件质量为什么是反馈系统而不是测试阶段]]。\n\n## Donald T. Campbell：Assessing the Impact of Planned Social Change\n\n- 作者：Donald T. Campbell\n- 原始版本：1976 年 Dartmouth College Occasional Paper；1979 年发表于 *Evaluation and Program Planning*, 2(1), 67–90\n- 本次核对版本：2011 年授权重印，*Journal of MultiDisciplinary Evaluation*, 7(15), 3–43\n- 原文：[期刊页面与 PDF](https://doi.org/10.56645/jmde.v7i15.297)\n\n### 研究对象与论证\n\nCampbell 讨论社会改革与公共项目怎样被评估，重点不是反对量化，而是分析评价进入行政决策后会发生什么。他把方法论问题、统计推断问题和评价与社会权力结构的相互作用放在同一框架中，并用教育、犯罪统计等案例说明：指标一旦影响资源、奖惩和合法性，记录者与被评价者就会承受改变数据或行为的压力。\n\n最常被概括为“坎贝尔定律”的判断是：定量社会指标越被用于社会决策，越容易遭受腐化压力，也越可能扭曲它原本要监测的过程。原文同时承认，许多例子是轶事性证据；它没有证明任何指标一经使用就必然失效，也没有主张放弃测量。\n\n### 能支持与不能支持什么\n\n- 支持：当指标连接高强度奖惩时，测量系统本身成为因果回路的一部分，不能继续假定数据生成过程不变。\n- 支持：应同时检查记录方式是否被操纵，以及现实过程是否为了指标而改变。\n- 不支持：把所有量化、目标管理或自动化门禁都判定为有害。\n- 不支持：把二手材料中的具体故事自动归给 Campbell。现有[[50-商业、产品、创业与组织/组织管理与工程协作/“坎贝尔定律” —— 为什么不能把考核当手段|坎贝尔定律底座材料]]所含“术后第 31 天”例子未在这篇原文中核实。\n\n## David Manheim 与 Scott Garrabrant：Categorizing Variants of Goodhart's Law\n\n- 作者：David Manheim、Scott Garrabrant\n- 年份：2018\n- 类型：机制分类论文，arXiv:1803.04585\n- 原文：[arXiv 页面与 PDF](https://arxiv.org/abs/1803.04585)\n\n### 四类失效机制\n\n论文把“代理指标被优化后不再可靠”拆成四类，而不是把所有失败都归为同一种作弊：\n\n1. **回归型 Goodhart**：代理与目标存在噪声；选择代理的极端值时，也选择了更多偶然噪声，目标的真实改善小于代理表现。\n2. **极端型 Goodhart**：优化把系统推到原有观测范围之外，原先成立的代理—目标关系在分布外失效。\n3. **因果型 Goodhart**：代理和目标的相关来自共同原因、错误方向或中介关系；直接干预代理并不会产生目标，甚至会破坏原机制。\n4. **对抗型 Goodhart**：会响应规则的主体主动操纵输入、报告或行为，使观察者看到的代理与真实目标分离。\n\n### 边界\n\n这是一套诊断分类，不是估计各种失效频率的实证研究。同一事件可能同时包含多类机制；分类也不会自动告诉决策者应该取消指标、换指标还是降低激励强度。迁移到软件测试、性能优化和知识网络时，属于机制类比，仍需用具体失败、反例和结果校准。\n\n## 对知识网络的作用\n\n两份来源共同把“指标不好”改写为可诊断问题：代理与目标如何关联、优化是否离开适用范围、干预方向是否正确、参与者是否会适应规则。它们支持在设计指标时预先记录目标、代理、激励和可能的替代行为；不支持只要换一个更复杂指标就能消除代理问题。\n"
}