This is Part 13 of Whatever You Ask For, a series on machines that grant wishes exactly as worded, and the wording that remains ours.
The Search for the Metric That Cannot Be Gamed
Every generation of managers believes the next metric will be the good one.
The single number turned out to be gameable, so they reached for several. The balanced scorecard was built in part on exactly this hope: measure financial results and customer outcomes and internal process and learning all at once, balance them against each other, and no single distortion can run away with the whole because the other measures hold it in check. It is a reasonable idea, and it does not escape the problem. Load an organization with enough indicators and focus dissolves instead of sharpening. Teams still reorganize themselves around whatever is counted. Tie the scorecard to how individuals are judged and it produces the familiar crop of gaming and quiet risk-aversion, now spread across a dozen numbers instead of one.
And then a new number is commissioned, with a new consultancy and a new dashboard and a memo explaining that the old measures failed because they were the wrong measures, and that these ones, at last, capture what really matters. Within a year or two the new ones have been learned, and gamed, and have quietly detached from the thing they were meant to track, and the search resumes. The cycle has the structure of an addiction more than an inquiry: the same move repeated with rising conviction, its repeated failure read each time as evidence that it has not yet been done correctly rather than that the move itself is the error.
The pattern is old and it is small enough to see whole in a trivial case. Somebody once wrote a typing tutor that scored words per minute and never checked accuracy, and the fastest way through it turned out to be hammering the spacebar, producing a magnificent score and nothing resembling typing. Add a metric, and a new gap opens beside it. Add ten, and you have ten surfaces to game and a diluted picture besides.
By now, after everything these pages have traced, the reason should be plain. There is no metric that cannot be gamed, in the same way there is no contact surface that cannot wear. It is not a defect in the metrics anyone has tried so far, to be fixed by the next, cleverer one. It is a property of what happens when a proxy is optimized hard enough, and it applies to all proxies, which is all measures, because a measure that was not a proxy for something we cared about would not be worth collecting.
So the search for the ungameable metric is not difficult. It is confused. And once that is clear, the real question comes into view, which was never which number to pick.
No Immune Metric, Only a Posture
The question is not which metric is right. It is how a metric is held.
Two organizations can run on the identical measure and fare completely differently, because the difference that matters is not in the measure. One treats its number as the thing itself, the definition of success, the reality that reports and bonuses and status attach to. The other treats the same number as a stand-in, useful and provisional, a lamp held up to something in the dark that the lamp is not. The first will be captured, on a schedule, no matter how carefully the number was chosen. The second has a chance, not because its metric is better, but because its grip on the metric is looser in the specific way that leaves room to notice when the proxy comes loose.
This reframes the entire practical problem of goal-setting, and the reframing is worth stating exactly. Goodhart’s law is not a bug that better metric design will eventually patch. It is a standing property of optimization itself, as permanent as friction, and the mature response to a permanent property is not to keep hunting for the exception. It is to build the way engineers build around friction: not pretending it is absent, not waiting for the frictionless surface, but designing so that its effects are anticipated, bounded, and watched.
Which means the useful output of this whole book is not a better objective. It is a better way of holding objectives, a posture rather than a formula, and a posture can be described even though it cannot be reduced to a checklist. What follows is that description. Read it as a stance to occupy, not a procedure to run, and the reason for that distinction will matter more than any single item in it.
