Short answer

It compares a baseline and revision against the same inputs and rubric. Fixed cases make regressions visible.

Boundary

Version cases, include missing-evidence and refusal tests, and inspect failures before adoption.