The Proxy and the Prize

The intro left us with a promise: every case of Goodhart’s Law begins with an innocent, sensible decision. Not a villain, not a con — just someone facing something they genuinely care about and can’t see directly, reaching for the next best thing they can see. That reach is where the whole story starts. So before we watch any number go bad, let’s slow down on the moment before it does: why we measure a stand-in at all, and what quietly holds it together.

Why we ever measure a proxy

Start with what you actually care about. Call it the goal (or target): the real outcome you’re after — a child who has learned, a patient who is healthier, a codebase that works, a scientist doing good science. Now notice the cruel thing these have in common: you cannot see any of them directly, cheaply, or in time to act. “Did this child learn?” is answered honestly only years later. “Is this patient healthier?” needs a doctor, a history, a follow-up. Goals are slow, fuzzy, and expensive to observe.

So you reach for a proxy: a cheap, fast, visible measurement that stands in for the goal. Test scores stand in for learning. A&E four-hour wait time stands in for system health. Reported incidents stand in for safety. Lines of code stand in for programmer output. Citation counts stand in for scientific merit. None of these is the goal — they’re windows onto it.

Worked example — health → the four-hour wait. England’s National Health Service wanted “A&E departments are serving patients well.” Unmeasurable in the moment. So they picked a proxy: the share of patients seen, treated, and admitted or discharged within four hours. Before anyone’s career depended on it, this was a genuinely good proxy. A hospital with enough staff, free beds, and a slick process moves people through in under four hours; a swamped, under-resourced one doesn’t. Measure the wait, and you’d learn something real about the health of the system. That’s the honest starting point — and, as we’ll see, honesty is exactly the thing that doesn’t last.

The pitfall to name here is the opposite of cynicism: don’t conclude “measuring is bad.” Refusing to measure means flying blind, which is worse. A proxy is a reasonable, often necessary tool. The problem is never that we chose to measure — it’s what happens next.

When a proxy is doing its job

A proxy earns its keep when it’s cheap to read, hard to fake, and moves in step with the goal for reasons you didn’t have to arrange. Right up until someone starts pushing on it, a good proxy is one of the most useful things you can have. The skill this course builds isn’t “stop using proxies” — it’s knowing the precise conditions under which one is still telling you the truth.

The correlation you borrowed but didn’t build

Here’s the load-bearing idea for the entire course, so read it twice. A proxy is trustworthy only because of a correlation you did not create and do not control. In the wild — before any stakes — better teaching causes higher scores; a healthier, better-run hospital causes shorter waits; safer operations cause fewer incidents. The proxy tracks the goal because, out in the world, the goal quietly drives the proxy.

That correlation is borrowed, not guaranteed. You didn’t wire scores to learning; the world did, as a side effect of how learning and tests happen to relate when nobody’s manipulating either. And a borrowed correlation holds under exactly one condition: nobody is optimising the proxy. The moment effort is aimed at the number itself rather than the goal beneath it, the link that made the number meaningful is the first thing to snap — because it was never a law of nature, just an incidental by-product you were leaning on.

Attaching a stake: the fateful act

Everything so far has been calm. Now we do the one thing that turns a useful proxy into a trap: we attach a stake to it. A bonus for hitting the number. A league table that ranks you by it. Funding, promotion, or a headline that follows it. The size of that reward — how hard reality now presses on people to move the number — is what we’ll call optimisation pressure.

Watch what optimisation pressure does. The instant the number becomes the reward, the rational move is no longer to improve the goal — it’s to improve the proxy, by whatever route is cheapest. And gaming the proxy is almost always cheaper than achieving the goal. Teaching better is hard and slow; drilling last year’s exam paper is fast. Genuinely healthier operations take years; reclassifying a patient as “not yet officially waiting” takes a keystroke. So effort floods toward the proxy — and remember what the proxy’s value rested on: a borrowed correlation that only held while nobody was pushing. Pushing is precisely what we just started.

Here’s the punchline, and it’s the shape of every Goodhart disaster to come: you get the number without the substance. Scores rise while learning stalls. Waits “improve” while sick people queue in ambulances outside the door, uncounted. Incident reports fall not because work got safer but because reporting got punished. The proxy soars, the goal sags, and the two — once faithful companions — walk off in opposite directions.

Every row is the same skeleton in a different costume: a proxy that was honest in the wild, a stake bolted on, and effort rerouted from the goal to the gameable number.

Watch the correlation crumble

Reading about it is one thing; watching it happen is another. Below, every dot is a person being measured. Their real quality runs along the horizontal axis; the proxy score we judge them by runs up the vertical. At zero pressure, every dot sits neatly on the honest diagonal — the proxy equals reality, and the number tells the truth.

Now drag the optimisation-pressure slider up from 0 and watch three things at once: the measured score rockets (the number looks like a triumph), the real value sags (the thing you cared about is quietly getting worse), and the proxy ↔ reality correlation crumbles toward zero — and can even flip negative. That last part is the sting: when the correlation goes negative, your top scorers are your best gamers, not your best performers. The number hasn’t just stopped tracking the goal; it’s started pointing the wrong way.

Notice why it happens in the model: each person has a hidden real quality and a completely separate gaming ability. At zero stakes, gaming ability is dormant and irrelevant — the proxy reads real quality. Add pressure, and the people who happen to be good gamers pull their scores up regardless of their real quality, while the effort spent gaming actively drags that real quality down. The borrowed correlation had no defence against this, because it was never built to survive someone optimising the number. It was a loan, and the slider is the world calling it in.

So when is a proxy safe?

If proxies break under pressure, is measuring hopeless? No — and this is the foreshadowing for the fixes lesson. A proxy stays honest under three conditions, and loses one of them at your peril:

• Low stakes. The smaller the reward riding on the number, the less optimisation pressure, and the longer the borrowed correlation survives. A metric you merely watch is far safer than one you pay for.
• Hard to game. If moving the proxy without moving the goal is expensive or impossible, gaming isn’t worth it and the link holds. Proxies die fastest when cheating them is cheaper than achieving the goal.
• One of several cross-checks. A single proxy is a single point of failure. Watch several imperfect measures that are gamed in different ways, and gaming one tends to betray you on another — no single number becomes the whole prize.

So the map of this lesson is simple, and it’s the seed of everything ahead: we measure a proxy because the goal is invisible; we can trust the proxy only because of a correlation we borrowed from a world with no stakes; and the moment we attach a stake, optimisation pressure floods toward the number and snaps the very link that made it meaningful. Number up, substance down. That gap — soaring measure, sinking reality — is Goodhart’s Law, and now you can name every part of it.

Recap

Next we’ll open up the fateful act itself: how exactly the seams split under pressure — Campbell’s Law, and the four distinct flavours of Goodhart that decide which way your number betrays you.