1.2 Goodhart’s law and the psychology of metrics

Overview and motivation

Goodhart’s law, named for the economist Charles Goodhart, is usually stated as: when a measure becomes a target, it ceases to be a good measure. Goodhart’s original 1975 observation was about monetary policy, but the anthropologist Marilyn Strathern’s later restatement is the version software teams actually need, and it is the sentence this entire book is built on. Every metric in every later topic, deployment frequency, test coverage, satisfaction scores, carries this risk, and every recommendation in this book is, in some form, a strategy for managing it.

The mechanism is not mysterious. People respond to incentives, and a metric attached to a reward, a review, or a reputation is an incentive whether anyone intended it as one. Once a team knows “deployment frequency” is being watched, the cheapest way to move that number is not always the intended one: split one meaningful change into five trivial deploys, and the number goes up while nothing real improved. This is not a story about bad actors. Ordinary, well-meaning engineers respond exactly this way to badly designed incentives, because the incentive, not the intent behind it, is what shapes behavior under pressure.

For large organizations the stakes are higher because the distance between the metric’s designer and the person whose behavior it shapes grows with scale. A team lead who builds a metric for their own eight-person team can watch for gaming directly and correct course quickly. A metric rolled out across a six-hundred-person division, or published in a government performance report read by a legislature, travels through layers of people who never met its author and have every reason to treat the letter of the metric as the goal. The distortion compounds with distance, which is exactly why this topic, not a later one, is where the book puts its center of gravity.

Key principles

Assume every incentivized metric will be gamed. Design against it from the first version, not after the distortion is discovered.
The gaming is rational, not malicious. People are responding sensibly to the incentive you built; blaming them for it fixes nothing.
Distance from the metric’s owner increases distortion risk. The further a number travels from the person who understands its intent, the more it becomes the letter of the rule rather than the spirit.
Ratios and ranges resist gaming better than raw counts. A raw count rewards volume; a well-chosen ratio rewards the actual behavior you want.
A guardrail is not optional on an incentivized metric. Every metric you attach a reward to needs a paired counter-metric that must not degrade.

Recommendations

Classify every metric by incentive exposure

Before publishing a metric anywhere visible, ask directly: does anyone’s reward, review, reputation, or budget depend on this number moving in a particular direction? If yes, it is an incentivized metric and needs a guardrail (below) before it goes live. If no, it is a diagnostic metric (topic 1.1) and carries lower gaming risk, though never zero, because people can still shape a number they merely expect to be judged on later even without a formal incentive attached today.

Prefer ratios, rates, and cohorts over raw counts

A raw count like “tickets closed” is gameable by doing more of something low-value. A ratio like “percentage of tickets resolved on first contact” rewards the underlying behavior instead of the volume. A cohort, a group defined by a shared starting point such as all deploys in a given week, prevents a bad recent trend from hiding inside a flattering long-run aggregate. Wherever you are choosing between a count and a rate that captures the same underlying behavior, choose the rate.

Pair every incentivized metric with a guardrail

A guardrail metric is a paired counter-metric that must not degrade while the primary metric improves. Deployment frequency pairs with change failure rate; lead time pairs with defect escape rate; a support team’s handle time pairs with customer satisfaction. The guardrail is what makes cheap gaming visibly expensive: a team that improves the incentivized number by degrading the guardrail gets caught by the pairing, not by luck. Design the guardrail at the same time as the primary metric, never as an afterthought once gaming is already discovered.

Watch for the four classic gaming patterns

Distortion under Goodhart’s law tends to fall into a small number of recognizable shapes. Threshold gaming optimizes right up to a target and stops (a 95% test-coverage target produces trivial tests to reach exactly 95%, not genuine coverage). Definition gaming changes what counts rather than what happens (redefining “resolved” to exclude hard cases). Timing gaming shifts when work is recorded rather than when it occurred (batching deploys just before a reporting window closes). Substitution gaming delivers the letter of the metric while abandoning its intent (splitting one real change into many trivial ones to inflate deployment frequency). Naming these patterns to your team, explicitly, makes them far easier to spot when they appear in your own numbers.

Separate measurement from reward wherever you can

The strongest guardrail of all is structural: decouple the metric from individual reward. A metric used purely to understand a system, with no person’s pay, rating, or standing riding on its direction, faces far weaker gaming pressure than one tied to an evaluation. This is why topic 1.1’s diagnostic-versus-evaluative distinction matters so much in practice: keeping a metric diagnostic is often cheaper and more effective than any amount of guardrail engineering applied after the fact.

Trade-offs: pros and cons

Approach	Pros	Cons
Raw counts	Simple to compute and explain	Highly gameable by volume
Ratios and rates	Reward the right behavior, resist volume gaming	Can hide a shrinking denominator problem
Guardrail pairing	Makes cheap gaming visibly expensive	Doubles the metrics to define, own, and maintain
Diagnostic-only (no individual reward)	Lowest gaming pressure of any option	Weaker direct motivational lever for leadership to pull
Heavily incentivized metrics	Strong, fast behavioral response	High distortion risk, often within one reporting cycle

The central tension is motivational power versus distortion risk. The metrics that move behavior fastest, tying a number directly to reward, are exactly the ones most exposed to Goodhart’s law. Resolve the tension by reserving strong incentives for outcome metrics that are genuinely hard to game cheaply, and by pairing anything you do incentivize with a guardrail designed at the same time, not bolted on after the first distortion appears.

Questions to discuss with your team

For each metric anyone’s reward depends on, what is the cheapest way to game it, and would we catch that gaming today? Sit down and deliberately design the exploit for every incentivized number on your dashboard: how would a rational, well-meaning team make this look good without doing the underlying work? If you cannot name a way you would catch that gaming, you are not ready to incentivize the metric yet. This exercise is uncomfortable and that discomfort is the point.
Which of our current metrics have already drifted into one of the four gaming patterns, threshold, definition, timing, or substitution gaming, without anyone calling it out? Distortion rarely announces itself; it shows up as a number that looks great while the underlying complaints, incidents, or customer feedback tell a different story. Walk your dashboard against each pattern by name and be honest about matches.
Does every incentivized metric on our dashboard have a paired guardrail, and was that guardrail designed at the same time as the metric? A guardrail added only after gaming is discovered is a repair, not a design choice, and it usually arrives too late to prevent the first round of damage to trust. Audit your incentivized metrics for this pairing specifically.
How far does this metric travel from the person who understands its intent before it reaches the person whose behavior it shapes? A metric built by a platform team and consumed three management layers away, or published in a public report read by people who never saw the instrumentation, is far more exposed to letter-not-spirit gaming than one a team designed for itself. Map that distance for your most consequential metrics.
Have we ever removed an incentive from a metric after discovering it was being gamed, and what did that cost us in trust to fix? Organizations often discover Goodhart’s law the hard way, after a quarter or a year of distorted behavior, and the repair costs more than prevention would have. Bring a real incident, if you have one, and extract the lesson explicitly rather than quietly moving on.
Where have we assumed gaming was a personal integrity problem rather than a rational response to a badly designed incentive? Blaming individuals for responding predictably to an incentive you built rarely fixes anything and often damages trust further. Reframe every gaming incident you can recall as a design problem in the metric, not a character problem in the person, and ask what redesign would have prevented it.

Sector lens

Startup. With a tiny team, the fastest guardrail is direct conversation: everyone can see a number and immediately ask “wait, why did that jump.” The real risk is a founder attaching a metric to fundraising narrative (growth at all costs) without a paired guardrail, because outside investors apply exactly the kind of distant, high-stakes pressure that makes gaming attractive.

Small business. Off-the-shelf tools often ship default dashboards built around counts (tickets closed, calls handled) because counts are easy to compute. Actively convert these to rates wherever the tool allows it, and resist tying any single number to a bonus or a review without first identifying its guardrail.

Enterprise. Distance is the dominant risk: a metric designed by a platform team for internal diagnosis gets picked up three layers of management later and turned into a KPI nobody who built it would recognize. Govern this explicitly (topic 1.4): require a documented guardrail before any metric is approved for use in a performance review or an executive scorecard.

Government. Published performance measures face the strongest gaming pressure of any category in this book, because a missed target can carry budget or political consequences. Audit the definition itself on a fixed cadence, not just the number, since the classic public-sector gaming pattern is quietly redefining who counts (a waiting list “resolved” by reclassifying who is waiting) rather than improving the underlying service.

Examples

Enterprise. A retail technology company set a target of 99% automated test coverage across all services, tied to a team-level quality score used in quarterly reviews. Within two quarters, coverage hit 99%, and the incident rate rose. An audit found teams writing trivial tests, asserting that a function returned without throwing, purely to satisfy the coverage tool, while genuine edge-case testing had not improved at all. The fix replaced the raw coverage target with a paired metric: coverage plus a mutation-testing score (topic 4.2) that measures whether tests actually catch injected faults, which is far harder to game cheaply.

Government. A state’s unemployment-insurance agency was measured on median days to first payment, published to its legislature. Under pressure to hit a target, one regional office began quietly reclassifying harder-to-process claims as “incomplete” and excluding them from the denominator, which made the published median look excellent while some claimants waited far longer than the report suggested. An independent audit of the definition itself, not just the number, uncovered the practice. The agency’s fix froze the definition, published the exclusion criteria publicly, and added a guardrail metric tracking the incomplete-claims rate itself, so that a spike in reclassification would now be visible rather than hidden.

Business case: motivations, ROI, and TCO

The return on taking Goodhart’s law seriously is avoided rework. An organization that designs guardrails up front spends a modest amount of extra effort defining a second metric alongside the first. An organization that skips this step often spends a full quarter or more of misdirected effort before the distortion surfaces, followed by the much harder cost of undoing gamed behavior and rebuilding trust in the number afterward. The retail example above is typical: cheap to prevent, expensive to repair.

The total cost of ownership of a guardrail is not free: it is a second metric to define, instrument, and review. But that cost is small and fixed compared to the unbounded cost of an incentive that quietly rewards the wrong behavior for months before anyone notices. Every topic after this one prices that trade-off in, which is why guardrail pairing appears as a recommendation throughout the rest of this book rather than only here.

Anti-patterns and pitfalls

Publishing an incentivized metric with no guardrail: the single most common root cause of a distorted dashboard in this book.
Treating gaming as a personal failing: blames individuals for a rational response to a badly designed incentive, and fixes nothing.
Auditing the number but never the definition: the classic public-sector failure mode, where the metric looks fine because who counts quietly changed.
Assuming a metric that worked as diagnostic will stay safe once it becomes evaluative: the exposure changes the moment reward attaches, even if nothing else about the metric changes.
Designing the guardrail only after the first gaming incident: a repair that arrives after the damage to trust is already done.
Ignoring distance: assuming a metric will be read the way its designer intended once it travels several management layers or a public report away from them.

Maturity model

Level 1, Initiate: Metrics are incentivized ad hoc, with no consideration of gaming risk, and distortion is discovered only after quality or trust visibly suffers.
Level 2, Develop: Some teams recognize gaming after the fact and adjust informally, but there is no consistent practice of designing guardrails in advance.
Level 3, Standardize: Every incentivized metric organization-wide requires a documented guardrail before approval, and the four gaming patterns are named and taught.
Level 4, Manage: Gaming risk is actively monitored: definitions are periodically audited, guardrail pairs are reviewed for whether they still catch distortion, and incidents of gaming are tracked as a metric in their own right.
Level 5, Orchestrate: The organization treats Goodhart’s law as a standing design constraint, reviewed automatically whenever a new metric is proposed, and it can point to specific redesigns that prevented distortion before it happened rather than only after.

Ideas for discussion

What is the most consequential metric in our organization that has no guardrail today?
Have we ever seen a number improve while the underlying reality got worse?
Who would notice if a definition behind one of our public metrics quietly changed?
Which of the four gaming patterns (threshold, definition, timing, substitution) is our organization most prone to?
What would it cost us, in trust, to discover a major metric had been gamed for a year?

Key takeaways

Goodhart’s law: a measure that becomes a target stops being a good measure, and this governs every metric in this book.
Gaming is a rational response to incentive, not a character flaw; fix the incentive design, not the people.
Prefer ratios, rates, and cohorts over raw counts wherever they capture the same behavior.
Every incentivized metric needs a guardrail, designed at the same time, not added after distortion is discovered.
Watch for the four gaming patterns by name: threshold, definition, timing, and substitution gaming.
Distance between a metric’s designer and the person whose behavior it shapes increases distortion risk; keep that distance short where you can.

References and further reading

Goodhart, C. A. E., “Problems of Monetary Management: The UK Experience” (1975): the origin of Goodhart’s law.
Strathern, Marilyn, “‘Improving Ratings’: Audit in the British University System” (1997): the widely cited restatement, “when a measure becomes a target, it ceases to be a good measure.”
Seeing Like a State, by James C. Scott (how legible metrics distort the systems they measure, at the scale of nations).
The Tyranny of Metrics, by Jerry Z. Muller (a book-length treatment of metric fixation and its costs across many professions).
Lean Analytics, by Alistair Croll and Benjamin Yoskovitz (vanity versus actionable metrics, and guardrail design in a startup context).
U.S. Government Accountability Office (GAO) guidance on performance measurement and the GPRA Modernization Act: public-sector performance reporting and gaming risk.