delidecTry the demo →

delidec / Research / COO

COO · The Operator on WatchOrchestration Layer29 Jun 2026

The Metrics That Mislead: Goodhart’s Law in the Operating Review

When a measure becomes a target, it stops measuring. We catalog how operating metrics corrupt under pressure and offer a discipline for choosing indicators that resist gaming.

The Substitution at the Heart of the Operating Review

Goodhart's Law is usually stated and then misunderstood. The failure is not that people cheat when you measure them; the failure is structural and precedes any bad intent. Every metric is a proxy standing in for something you actually care about but cannot observe directly. You care about durable customer value; you watch monthly active users. You care about engineering throughput; you watch tickets closed. The gap between the proxy and the true objective is small at rest, which is why the metric looked good when you chose it. Pressure widens that gap. When you attach consequence to the proxy — a bonus, a board slide, a headcount decision — you create a gradient, and the organization optimizes along the gradient of least resistance. The shortest path from here to a higher number almost never runs through the true objective. It runs through the gap.

Four Mechanisms of Corruption

Metrics do not all degrade the same way, and the remedy depends on the mode. In the orchestration layer, where you are routing attention and capital across many teams against a small set of indicators, naming the mechanism is the first act of control.

The Trade-off You Cannot Escape

The instinct after a gaming incident is to add more metrics — surround the gamed number with guardrails until no single one can be moved in isolation. This works, but it is not free, and pretending it is free is its own error. Every indicator you add to the operating review is attention you spend and a new surface to optimize against. A dashboard with forty numbers does not resist Goodhart; it diffuses accountability until no one owns the true objective at all, and it teaches the organization that the review is a compliance ritual rather than a steering instrument. The real trade-off is between completeness and legibility. A metric set rich enough to be ungameable is too large to reason about; a set small enough to drive decisions is always gameable by someone who studies it longer than you do. There is no measurement architecture that escapes this. There is only the choice of where to sit on the curve, made deliberately rather than by accretion.

A Discipline for Choosing Indicators

The defense is not a better metric but a posture toward the ones you have. Choose proxies whose cheapest way to improve is the thing you actually want — a metric is well-built when the gap between proxy and objective is narrow and the path through the gap is expensive, so gaming costs more than performing. Pair every efficiency or speed measure with a quality or durability measure that moves in the opposite direction when the first is gamed, and watch the pair, never the single number. Hold a few indicators deliberately uninstrumented and unannounced, sampled by hand, so that what you inspect is not fully known to those being inspected; the moment a check becomes a published target it joins the gradient and stops checking anything.

Above all, treat tenure of a metric as a liability, not an asset. A number's resistance to gaming decays with exposure: the longer a target has hung on the wall, the more cycles the organization has had to learn its shortcuts, and the more confidently you are being shown what you asked to see. The operating review's job is not to confirm that the targets were hit. It is to keep asking whether the targets still measure what made them worth choosing — and to retire them, on a schedule, before the answer quietly turns to no.

← All researchPut a question to the board →