What To Do About It

Four lessons in, you have a superpower and a matching curse. The superpower: you can now spot a metric going bad from across the room. The curse: you might be tempted to sneer at every number, fold your arms, and declare all measurement a lie. That reaction feels sophisticated. It is actually its own failure mode — and this lesson is about not falling into it.

So we do two honest things. First we draw the boundaries of the law — the places where “Goodhart!” is wrong, overstated, or being used as a dodge. Then we open the toolkit: the fixes that genuinely blunt Goodhart in the wild, each tied back to one of the three parents from lesson 04 (incentives, the principal–agent gap, the feedback loop), and each carrying its own cost. Because — as you learned with the alignment toolkit — a fix that names its own limitation is the only kind worth trusting.

Part A — Where the model lies to you

Every sharp idea attracts a lazy caricature, and Goodhart’s is especially seductive because it sounds wise to reject numbers. Before the fixes, three corrections — each in the shape the trap → why it’s wrong → what’s true instead.

Trap 1: “Goodhart means stop measuring”

The trap. You watch four lessons of metrics rotting and conclude the safe move is to measure nothing — to run on gut and vibes, because any number will eventually betray you.

Why it’s wrong. Flying blind is worse than an imperfect metric, not better. A pilot in cloud doesn’t rip out the altimeter because instruments can occasionally mislead — they’d hit a mountain. An unmeasured organisation doesn’t escape Goodhart; it just swaps a gameable number for something even easier to game: politics, charisma, whoever argues loudest in the meeting. Remove the metric and you don’t remove the incentive to look good — you remove your only evidence of whether anyone actually is good.

What’s true instead. Goodhart is not an argument against measurement. It’s an argument against naive single-metric targets under high stakes — against handing the controls to one gameable dial and rewarding it hard. You still need instruments. You just don’t wire a bonus, a firing, or a headline to a lone number and walk away. Keep measuring; stop worshipping the measure.

Trap 2: “Every measure inevitably collapses”

The trap. If a measure becomes a target it ceases to be a good measure — so, fatalistically, every metric is doomed the instant you use it. Why bother improving any of them?

Why it’s wrong. Goodhart is a risk that scales, not a law of instant doom. How fast (and whether) a measure rots depends on three knobs you already met in lesson 04:

• Optimisation pressure — how hard and how concentrated the reward is. A metric nobody’s pay depends on can stay honest for years.
• Gameability — how easy it is to move the number without moving the goal. Some proxies are genuinely hard to fake (an independently audited outcome, a physical result you can inspect).
• Stakes — how much rides on this single number. Low-stakes measures barely tempt anyone.

Crank all three and the metric detaches almost immediately (the school whose scores decide pay). Leave them low and the same measurement can read true for a very long time (the same scores when nobody’s judged on them).

What’s true instead. Goodhart describes a gradient, not a cliff. Measures that are low-stakes, hard to game, or that nobody is judged on can stay trustworthy indefinitely. The law tells you where the danger concentrates — high pressure × high gameability × high stakes — so you can spend your defensive effort exactly there instead of despairing everywhere.

Trap 3: “Goodhart!” as a lazy dodge

The trap. The caveat becomes a shield. “You can’t measure my team — any target is just Goodhart.” Quote the law loudly enough and you’ve excused yourself from all accountability, all comparison, all scrutiny.

Why it’s wrong. This weaponises a real insight into an all-purpose escape hatch. Yes, targets can be gamed — but “this metric is imperfect” does not imply “therefore measure nothing and never hold me to anything.” That’s the fatalism of Trap 1 dressed up as expertise. A surgeon who refuses any outcome tracking, a fund that rejects any benchmark, a team that won’t accept any goal — each is using a genuine caveat to buy total immunity from feedback, which is exactly the environment where quiet failure hides best.

What’s true instead. The honest use of Goodhart is surgical, not blanket. It says: this specific number, under this specific pressure, will detach in this specific way — so here’s how I’ll defend it. It should make your measurement smarter and your defences sharper, never make you unaccountable. If someone invokes Goodhart to end a conversation about whether they’re delivering, they’ve inverted the law: the point is to keep measuring well, not to stop.

Part B — The fixes that actually work

Now the constructive half. None of these fixes eliminates Goodhart — remember lesson 04’s punchline, that Goodhart is emergent from three parents, and lesson-before’s, that no single tool zeroes the problem. Each fix instead attacks one parent and pays for it somewhere. Deploy them in combination, name the cost, and you can keep a metric honest far longer than the naive version ever could.

Fix 1: Baskets of metrics that are hard to game together

Mechanism. Instead of one dial, track several that pull in tension — so gaming one shows up as damage in another. Pair a quantity metric with a quality metric and a counter-metric that gets worse exactly when someone games the first. This attacks the incentives parent: it dilutes the reward across numbers so no single lever is worth yanking, because yanking it lights up a neighbour.

Worked example. A support centre that rewards only “calls closed per hour” gets agents hanging up on customers to pad the count. So build a basket:

• Calls closed per hour — the throughput you want (but gameable alone).
• Customer satisfaction (CSAT) — closing calls by hanging up tanks this.
• Repeat-contact rate — “resolving” a call that isn’t fixed sends the customer straight back, spiking this counter-metric.

Now the only way to look good on all three at once is to actually resolve problems quickly — which is the real goal. Gaming throughput alone now costs you on satisfaction and repeat contacts. The basket is hard to game together.

Its limitation. Baskets are only as good as their tension. If all your metrics are correlated (they all rise when you game the same thing), you’ve just built a bigger dial, not a counterweighted one. And a sufficiently clever agent can sometimes game the whole basket at once. Baskets also add complexity and can slow decisions — more numbers, more argument about which one matters this quarter.

Fix 2: Hold targets loosely — signals, not fates

Mechanism. Use a metric to start a conversation, not to auto-trigger reward and punishment. Treat the number as a signal worth investigating (“why did this dip?”) rather than a fate that mechanically decides pay, promotion, or funding. This attacks the incentives parent directly by turning down the stakes on any single number — and lower stakes mean lower optimisation pressure, which (Trap 2) is one of the three knobs that drive the rot.

Worked example. A sales manager notices one rep’s numbers slipped this month. The autopilot response: dock the bonus automatically. The loose-target response: ask why — and discover the rep spent the month rescuing the firm’s biggest at-risk account, which no monthly figure captures. Held loosely, the metric prompted a useful question. Held as a fate, it would have punished exactly the behaviour the company most wanted.

Its limitation. Loosely-held targets are soft, and softness can shade into no accountability at all (straight back to Trap 3). If every miss can be explained away, the metric stops disciplining anyone. Holding targets loosely demands managers with judgment and the willingness to use it — it doesn’t scale to a spreadsheet that grades itself, which is precisely why organisations reach for hard targets in the first place.

Fix 3: Keep humans in the loop — qualitative audit

Mechanism. Put judgment and spot-checks where a formula can’t reach. A human auditor can see the spirit being violated even when the letter is satisfied — the gamed call, the cherry-picked patient, the padded report — because they can look at things no metric encodes. This attacks the principal–agent parent: it shrinks the information asymmetry that lets the agent game the letter undetected. Even random, unannounced checks work, because the agent can’t predict when the fog will lift.

Worked example. A hospital ranked on mortality could quietly refuse its sickest patients to flatter the stat (lesson 04). Add a clinical-audit committee that periodically reviews declined cases, not just treated ones, and the manoeuvre becomes visible — a surgeon turning away operable patients now shows up to a human reviewer even though it never appears in the mortality number. The audit sees what the metric structurally cannot.

Its limitation. Human audit is costly, slow, and subjective. It doesn’t scale — you can’t spot-check everything, so you sample. It introduces the auditor’s own biases and can be captured or gamed in turn (who audits the auditor?). And judgment is contestable in a way a number isn’t, which can make decisions harder to defend. It’s a powerful flank, not a cheap one.

Fix 4: Reward outcomes, not proxies, where you can

Mechanism. Move the incentive closer to the real goal. A proxy exists because the true outcome was hard to measure — but where you can measure the outcome (even slowly), reward that instead. Deferred, outcome-based pay and long time horizons pull the reward off the gameable short-term proxy and onto the thing you actually wanted. This attacks the feedback loop parent by lengthening the loop: if the reward only lands after the true outcome is visible, short-term gaming stops paying.

Worked example. Pay a mortgage broker a commission the instant a loan closes and you incentivise volume regardless of whether the borrower can repay (a compact retelling of 2008). Defer part of that pay and claw it back if the loan defaults within three years, and suddenly the broker cares about the outcome — a repaid loan — not the proxy of a signed contract. The reward now tracks the real goal because it waits for the real goal to reveal itself.

Its limitation. Outcomes are slow and noisy — which is the whole reason you reached for a proxy in the first place. Waiting years for the true result means slow feedback, and luck muddies the signal (a good loan can default in a recession; a bad one can survive a boom). You often can’t fully defer pay — people need to eat this year — so you’re forced back onto some proxy. This fix narrows the gap between proxy and goal; it rarely closes it.

Fix 5: Keep the true goal explicit — and expect to rotate metrics

Mechanism. Two moves that travel together. First, keep the true goal written down and in view, so nobody mistakes the proxy for the point — the metric is a stand-in, and saying so out loud is a cheap, constant defence. Second, treat every metric as perishable: a proxy that’s honest today will be gamed tomorrow, so plan to refresh or rotate it before gaming catches up. This attacks the feedback loop parent by resetting the loop — you retire the metric before its validity fully decays, denying the runaway spiral time to complete.

Worked example. A content platform knows any single engagement metric will eventually be gamed into clickbait (lesson 03). So it states the true goal explicitly — “long-term user value” — and rotates its operational proxies: this quarter, watch-time; next, a satisfaction survey; next, a downstream retention measure. By the time creators have reverse-engineered one proxy, the platform has moved to another, and the explicit true goal keeps every rotation pointed the same way. The metric is treated as ammunition to be spent, not a monument to be defended.

Its limitation. Rotation is disruptive and expensive — people need stable targets to plan against, and constantly changing the number breeds confusion, cynicism (“why bother, they’ll change it again”), and lost comparability across time. Keeping the true goal explicit sounds free but requires real discipline; goals drift, get reinterpreted, and quietly collapse back into “hit the proxy.” Perishability is a mindset you have to keep paying for, not a one-time fix.

The fixes at a glance

The decision checklist

You don’t need to memorise five fixes. You need five questions to ask before you set any target — each one routes you to the right tool and forces you to name the cost. Run this list and you’ve internalised the whole course:

1. How gameable is this proxy? Can someone move the number without moving the goal? The easier it is to fake, the less weight it can safely bear — and the more you need a basket or an audit.
2. How high are the stakes? How much rides on this single number (pay, firing, funding, headlines)? High stakes × high gameability is the danger zone — hold the target more loosely, or spread the stake across a basket.
3. What’s my counter-metric? What number gets worse exactly when someone games this one? If you can’t name one, you have a lone dial, not a basket — go build the counterweight.
4. Who audits with judgment? Where’s the human who can see the spirit being violated while the letter is satisfied? A formula alone can’t catch cherry-picking — name the auditor and the spot-check.
5. When will I retire it? Every metric is perishable. Decide now how you’ll refresh or rotate it before gaming catches up — and keep the true goal written down so every rotation points the same way.

Recap

That’s the course. You started with a factory stamping seventy-tonne nails and you end with a five-question checklist that would have saved it. Goodhart’s Law isn’t a reason to give up on numbers — it’s the discipline that lets you keep using them with your eyes open. The next step is the final exam: a graded, one-way run through the whole model. Treat every practice question you’ve met so far as exactly that — practice — and go prove it stuck.