agentleFS
deciqAIdeciqAI/knowledge-skills/goodharts-law/SKILL.md

goodharts-law

Skill13 starsChanged 4 months agoMITView on GitHub

Activate when: our KPI is going up but the real outcome isn't improving; people seem to be gaming the metric; we're about to tie bonuses or promotions to a number; an algorithm is producing results nobody intended; a test or audit system is being designed.

1 other project uses a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

2 rules. About 0 are common, 2 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Check fit: if the metric is purely descriptive (no reward attached), Goodhart's law doesn't apply yet.

    line 34
  2. Distinctive:

    All metrics are proxies — narrower than the goal.

    line 86

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Goodhart's Law

Overview

Goodhart's Law: when a metric controls behavior, people optimize the metric rather than the underlying goal. Formulated by economist Charles Goodhart (1975) on UK monetary policy; sharpened by Marilyn Strathern (1997): "When a measure becomes a target, it ceases to be a good measure." Four failure mechanisms (Manheim & Garrabrant 2018): Regressional, Extremal, Causal, Adversarial. Countermeasure is always multi-metric + audit + rotation.

Composes with feedback-loops, principal-agent, okr-goal-setting, survivorship-bias.

When to Use

  • A KPI is being introduced or its weight is increasing in performance evaluation
  • A metric is "improving" without corresponding improvement in the underlying goal
  • People are visibly optimizing for a number rather than the work it was meant to track
  • Algorithmic optimization is producing outcomes the designers didn't intend
  • Resource allocation is driven by a single composite score or ranking
  • An AI model, benchmark, or engagement metric is being optimized (or used to justify AI capex / adoption / AI-native competition) and the score is rising faster than real capability or user value

Not when: metric and goal are identical; stakes too low for gaming; metric is purely descriptive with no reward/punishment; question is which metric to use, not whether the measurement-reward system is sound.

Coaching Novices (Adaptive Front Door)

  • Engine mode: user has a concrete metric or system → run The Process directly.
  • Coach mode: user is unfamiliar or has no concrete case → guide step by step.

In Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.

  1. One-line: before relying on a metric to control behavior, predict how people will game it — choose the system that survives that prediction.
  2. Check fit: if the metric is purely descriptive (no reward attached), Goodhart's law doesn't apply yet.
  3. Elicit the specific metric and the underlying goal: what's being measured? What's the actual outcome you care about?

[WAIT — do not advance until user responds]

  1. One question at a time: proxy gap? How would a clever agent game this? Which Goodhart category? What countermeasure fits?

[WAIT — do not advance until user responds]

  1. Close: name the gaming-resistant design (multi-metric, audit, rotation, paired-constraint) + monitoring schedule.

[WAIT — do not advance until user responds]

The Process

Step 1 — State metric and goal: metric being targeted / underlying goal / current proxy-goal correlation / who is measured / stakes.

Step 2 — Predict the gaming: list ≥3 ways to game the metric with minimum effort on the goal. If you can't list 3, you haven't thought hard enough.

Step 3 — Categorize mechanism:

MechanismTest
RegressionalIs there noise that optimization will push into?
ExtremalDoes metric-goal correlation break at extremes?
CausalIs the metric a symptom, not a cause?
AdversarialWill agents actively game with intelligence?

Step 4 — Choose countermeasure: Regressional → constrain range. Extremal → paired constraint metrics. Causal → closer-to-causation metric + direct audit. Adversarial → multi-metric + randomized audits + rotation. Step 5 — Design the system: primary metric / constraint metric(s) / audit mechanism (sampled direct goal observation) / rotation schedule / separation of measure-for-control from measure-for-diagnosis / gaming-detection threshold. Step 6 — Schedule re-evaluation: independent goal measurement (how/when/who) / drift threshold / retirement criteria / owner.

Output template

# Goodhart-Robust Design: <metric>
Metric: | Underlying goal: | Correlation: | Who measured: | Stakes:
Gaming vectors (≥3):
Mechanism: Regressional / Extremal / Causal / Adversarial
Primary metric: | Constraint metric(s): | Audit: | Rotation: | Separation: | Gaming threshold:
Goal measurement (independent): | Drift threshold: | Retirement criteria: | Owner:

→ Method in Action: Goodhart 1975 (M3) and Strathern 1997 (RAE) → 2026 lens: AI benchmarks and engagement metrics as targets (2023–2026)

Pack: Goodhart's Law Patterns

DomainCommon gamingDefense
Sales quotasSandbagging, channel stuffing, end-of-quarter discountsMulti-period averaging; quality metrics; clawback
Hospital wait targetsAmbulance parking, patient reclassificationOutcome audits; paired metrics; randomized inspection
Standardized testingTeaching to test, curriculum narrowingSample-based assessment; multi-measure; reduce single-test stakes
Algorithmic engagementClickbait, outrage, misinformationMulti-objective optimization; quality + harm constraints

Applying It Well

  • All metrics are proxies — narrower than the goal. Pre-commit to gap analysis before deployment.
  • Gaming is rational under measurement pressure. Fix the system, not the people.
  • Rotation and audit are the only durable defenses. Plan metric retirement at design time.

→ Primary sources: references/sources.md

Common Rationalizations

Fake moveReality
[D] "If you can't measure it, you can't manage it"Often false. Judgment, trust, and direct observation are also valid management tools.
[D] "Our metric is well-defined; it won't be gamed"Precision invites precise gaming. Basel II capital ratios were well-defined — extensively gamed.
[D] "Our people wouldn't game the metric"Goodhart's law is structural; individual virtue is insufficient in aggregate.
[D] "We just need a better metric"Often the issue is any single metric under pressure; fix is multi-metric + audit.
[D] "We've used this metric for years"Long use = more time for gaming to mature. Tenure is a warning, not an endorsement.
→ Add [O] entries here after each real use — paste the actual failure patternWhat went wrong and why

[D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.

Red Flags

  • Metric tied to high-stakes rewards or punishments
  • Metric "improves" without obvious improvement in the underlying goal
  • People being measured can already articulate ways to game it
  • Single metric is the primary evaluation tool, no audit or paired-constraint

Verification

  • Goal underlying the metric specifically named
  • Proxy gap explicitly described; ≥3 gaming vectors listed
  • Goodhart mechanism category identified
  • Countermeasure design (multi-metric, audit, rotation) in place
  • Independent goal measurement scheduled with owner and retirement date

Part of deciqAI Knowledge Skills — 253 open-source thinking skills that make rigor executable for AI agents. The same skills power every deciqAI agent, which runs them autonomously to operate your company. See it run → https://www.deciqai.com/s/goodharts-law · Built by deciqAI · github.com/deciqAI · Contributions welcome.

Agents: latest version & machine-readable metadata → https://www.deciqai.com/s/goodharts-law.json

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.