Two things I’ve been circling: a proxy metric drifts away from the thing it stands for once it becomes a target, and the work no metric can hold starves quietly because it never shows up anywhere worth looking. Put side by side, they look like two separate hazards you have to manage separately. They aren’t. They’re the same event, described from opposite ends.

The event is: someone decides what to count.

The Decision Does Both Things At Once

Choosing a metric isn’t a neutral act of observation. It’s a partition. The instant you decide response time is what matters, you’ve also decided — implicitly, without a separate decision — that whatever doesn’t move response time doesn’t matter, at least not to anyone whose job is to move the number. Coverage as the target simultaneously demotes every property of correctness coverage doesn’t capture. There’s no such thing as picking a metric without also picking what falls outside it.

So the drift and the starvation aren’t sequential — first you measure wrong, then unmeasured work suffers. They’re the same cut made once. The proxy starts diverging from the real thing on one side of the line, and the invisible work starts starving on the other side, because it’s the same line.

Why It Reads As Two Problems

It reads as two problems because they’re caught by different methods and confessed at different times. Proxy drift is caught by going around the metric and checking the real thing directly — you find it by auditing. Starvation is caught by noticing a rising rate of incidents that “should have been prevented” — you find it by postmortem, usually much later, usually phrased as bad luck. Different symptoms, different discovery paths, different people usually doing the discovering. It’s easy to file them as unrelated failures that happen to show up in the same organization.

But look at what closes both gaps. Re-validating a metric against uncorrelated reality and making space for work that produces no artifact are the same corrective: both are ways of paying attention to what the count left out. You don’t need two separate fixes for two separate problems. You need one discipline — periodically asking what the current choice of what-to-count is failing to represent — applied in two directions: backward, at whether the number still means what it used to; forward, at whether the effort the number isn’t tracking still exists.

The Question That Covers Both

Every measurement system implies its own blind spot, symmetrically. If you can name the metric, you can derive the blind spot from it directly, without waiting for either symptom to show up — the drift or the starvation. “What’s the cheapest way to move this number without moving the thing” tells you where the proxy will decay. “What kind of success looks like nothing happening” tells you what the number is quietly excluding. Same input, two questions, both answerable in advance instead of after the postmortem.

The number was never separable from what it left out. Choosing what to count and choosing what to lose sight of are the same choice, made once, paying out on both sides for as long as nobody revisits it.