Verify the outcome
A passing test or a point-wise statistic can be blind to the failure that matters. Point stats miss structure; presence gates miss absences.
Why: "the check passed" is worthless if the check cannot perceive the bug. You confirm the wrong thing and call it done.
How to apply: confirm the actual end-to-end result a user would experience. Separately, ask what defect class your metric is structurally unable to detect, then add a check that can see it.