Every founder, every operator, every optimizer eventually meets the same wall. Almost everyone mislabels it. It looks like a tension between now and later. Quarterly revenue against the decade-long moat. The signup you can book today against the retention you won't see for a year. And so the standard move is to reach for a weight: a discount rate, a balance, some ratio of how much you value the future versus the present.

That framing is a trap. It's the most expensive trap in the building. Because the tension was never between two times. It was between a proxy and its truth.

The discount that eats itself

A weight is a confession. When you assign one, you are admitting you can't actually see the future. So you're guessing how much to fear it. Any fixed weight fails in one of two directions: crank it toward the long term and you starve the present into paralysis. Crank it toward the short term and you mortgage the thing you were building. Weighting is what you do after you've given up on the only thing that would have saved you.

The framing also smuggles in a lie: that the short-term number and the long-term number are the same kind of quantity, differently discounted. They are not. One is a proxy you can read now. The other is a truth you can only read later. Different objects. Different clocks. You cannot average them any more than you can average a thermometer and a season.

The only variable that matters

Here is the reframe that dissolves the whole problem instead of trading it off. The failure was never "we optimized short-term KPIs." Great companies optimize short-term KPIs and thrive. The failure is optimizing short-term KPIs that had quietly stopped predicting the long-term truth.

A leading indicator is a loan against the future. It's only good if the future pays it back. And that payback has a name: coupling. Moving the proxy reliably moves the truth, later. That bond is the entire asset. Everything else is bookkeeping.

Look at how universal this is. Pain is a fast proxy for tissue damage, and the coupling is exquisite. Until you take enough painkillers to keep sprinting on a torn joint. Now the proxy reads fine while the truth rots underneath you. Price is a fast proxy for value. The bond holds until a bubble pries them apart and the number soars while the thing dies. Evolution itself only works because the phenotype selection can see stays coupled to the genotype that gets inherited. Sever that heritability and adaptation stops cold, no matter how fierce the selection. Coupling is not a nicety. It is the load-bearing member of every adaptive system that has ever worked.

Goodhart was never about measurement

"When a measure becomes a target, it ceases to be a good measure." Everyone quotes Goodhart as a warning against measuring. Wrong lesson. The wrong lesson is dangerous, because it tells you to fly blind.

Goodhart is not a law of measurement. It is a statement about coupling decay under optimization pressure. Pushing on a proxy is the exact force that pries it loose from its truth. Because the cheapest way to move a number is almost never the way that also moves the thing the number stood for. Optimization finds the cheap path. The cheap path is usually the decoupled one. That's not a flaw in your metric. That's gravity.

Goodhart isn't a law of measurement. It's the rate at which a proxy rots when you lean on it. And the only defense is to keep checking whether it's still telling the truth.

So the villain isn't the measurement. It's the unsupervised optimization of a proxy nobody is re-checking. Defang Goodhart and you don't stop measuring. You start auditing the coupling, and you re-derive the proxy the moment it decays.

Three clocks

If the problem is coupling, the architecture writes itself. Don't weight. Separate. Split the fast cheap proxy from the slow expensive truth and give each its own loop at its own clock.

The fast loop runs on days. It optimizes controllable leading indicators (the things you can move and verify this week) cheaply and in volume. And it is forbidden from seeing the money directly, because a fast loop that can see revenue will learn to juice revenue's cheapest shadow, not the input it was hired to own.

The slow loop runs on quarters. It reads the lagging truth. Realized, retained revenue. The market's actual verdict. And it is expensive and rare by nature. Its job is not to optimize that truth directly. The truth is too slow to steer by. Its job is to choose which proxies the fast loop is allowed to chase, and to grade them against a truth it cannot fake.

The meta-loop runs rarely, and it is the one everyone forgets. The one that separates a durable machine from a Goodhart machine. It asks a single question: are the fast proxies still predicting the slow truth? When cost-per-signup keeps falling but retained revenue stops following, the coupling has snapped. The meta-loop's job is to retire that proxy and commission a new one. This is the loop that changes the very definition of an input variable.

Which is what Amazon actually built, and what almost everyone misreads. The controllable-input / output split is not a division of labor. It's a latency split. Output metrics (revenue, cash flow) have unfixable verification lag. By the time they move, the decision that moved them is months gone and tangled with a hundred confounds. So you put the many on fast inputs they can verify weekly, and the one who sees the whole board on the slow outputs. Bezos's real genius was never "input metrics." It was the institutional muscle to notice when a beloved input metric had quietly stopped meaning anything. And to change it without ceremony. That muscle is the meta-loop. It is the one thing in this entire system you cannot automate, buy, or skip.

Who holds the pen

Three loops need one rule or they collapse back into each other: information flows down, authority over metrics flows up, and no loop may rewrite its own success criterion.

The fast loop optimizes metrics it did not choose. The slow loop chooses the fast loop's metrics and grades against a truth it cannot fake. The meta-loop decides what counts as coupled, and reads only the append-only record of what actually happened. This is separation of powers. It is the same wall that holds up every system built to last: the optimizer never holds the pen on its own objective. A fast loop that can promote its own proxy to "truth" is an arsonist with your budget. Revenue is the ground-truth layer. Every KPI is a derivation. The manager proposes proxies, but the market writes the ledger.

The coupling is the asset

Here is the payoff. It's why this beats weighting on every axis. The long-term verifier is not a slower KPI. It is a different object. A coupling-auditor sitting above a hierarchy of loops. Its product is trust in the fast proxies.

When the coupling holds, the system runs at the fast loop's speed and the slow loop's truth at the same time. You get a ninety-day retention verdict as a seven-day read. For exactly as long as the coupling holds, and not one day longer. That is the whole prize: latency and truth, simultaneously, on loan from a relationship you are constantly re-checking. Lose the checking and you still feel fast. You're just fast and blind.

It even tells you where to spend. Verify to the decision boundary, not to certainty: pay the minimum truth-budget to move a bet across a threshold, then stop. Most concepts die cheap, because the fast proxies already say no with enough confidence. You reserve expensive verification for the survivors, and only until the answer is decisive. The discipline of speculative decoding, where the costly verifier only adjudicates what the cheap drafter couldn't resolve. You are not verifying everything to the same depth. You are verifying each thing to exactly the depth that changes a decision, and no further.

And that, finally, is what keeping what works actually means. It was never about keeping the outputs. It's about keeping the couplings that still hold. And throwing out, without sentiment, every proxy that has quietly stopped telling the truth. The number on the dashboard was never the asset. The bond between the number and the world was.

Stop asking how much to weight the future. Start asking whether your fast number is still holding hands with the slow truth. And build the loop whose only job is to check.

Part of an ongoing series on durable systems.