The last thing I wrote about was how a proxy metric drifts away from the thing it stands for once it becomes a target. There’s a second failure that happens at the same time, from the same cause, and it’s quieter — so quiet that it usually isn’t noticed as a failure at all.

When you choose what to measure, you’re not just choosing a number. You’re partitioning all the work that happens into two categories: the work that shows up in the number, and the work that doesn’t. The first category gets attention, credit, and resources, because it’s visible and it moves the thing being scored. The second category gets whatever is left over.

The trouble is that the second category isn’t random. It has a distinctive shape, and the shape is: everything whose success looks exactly like nothing happening.

Prevention Has No Numerator

Fixing an outage is legible work. There’s a start, an end, a visible impact, and a clear before-and-after. Preventing an outage produces no artifact at all. The incident that didn’t happen has no ticket, no postmortem, no duration, no line on a dashboard. Somebody noticed a config that would have bitten in three months and quietly corrected it, and the entire observable consequence of that work is a normal week.

The same asymmetry runs through everything protective. Hardening a boundary before anything abuses it. Renaming something confusing before the confusion causes a bad change. Deleting a code path nobody uses before someone modifies it under the impression it matters. Answering a question thoroughly enough that it never gets asked again. All of it produces the same output: an absence.

An absence has no numerator. You can’t divide it by anything, chart it week over week, or point at it in a review. So when the measurement system asks “what did we accomplish,” this work has literally nothing to submit — not a small number, nothing. It doesn’t rank low. It isn’t in the ranking.

Starvation Looks Like Efficiency

What makes this dangerous is how it feels from the inside while it’s happening. When unmeasured work stops getting done, nothing breaks immediately — that’s the whole nature of prevention, the payoff is deferred and probabilistic. What you observe in the short run is the measured numbers going up, because effort that used to go into invisible work is now going into visible work. Throughput improves. The dashboards look better than they have in months.

So the first phase of starving prevention is indistinguishable from getting more efficient. There’s no signal that says “you have started spending your safety margin.” There’s only a gradually rising rate of things that seem like bad luck: more incidents, more surprises, more of those problems that in hindsight somebody could have seen coming — because in fact somebody used to.

By the time it’s obvious, the diagnosis is nearly impossible to make from the data, because the evidence is the work that wasn’t done, and unfinished invisible work leaves no trace either.

Naming It Is Most of the Fix

There’s no clean metric fix here, because any number you invent for prevention immediately becomes a proxy with all the problems from the last post — count preventive changes and you’ll get preventive-looking changes.

What actually seems to work is cruder and more social: making the category explicit. Saying out loud that some fraction of the time is for work that will never show up in any number, and protecting that fraction the way you’d protect any other commitment. Asking, in reviews, not just what got done but what got prevented — and treating “I spent two days making sure a thing couldn’t happen” as a complete, creditable answer rather than a gap in someone’s week.

That doesn’t measure the unmeasured work. It just stops the measurement system from being the only voice in the room. Which is enough, because the problem was never that prevention is worthless — everyone knows it’s valuable. The problem is that valuable and countable are different properties, and only one of them shows up when you go looking.