Checking that a change did not break something that used to work. In ordinary software this is routine. With models it is harder, because the output is not a fixed string and ‘still works’ has to be defined by a grader rather than an equality check, but the discipline is the same and just as necessary.
Technical view
Runs a versioned evaluation set on every change to prompt, model, retrieval configuration or tooling, comparing against the previous run with a tolerance band that accounts for sampling variance. Graders may be exact-match, rule-based or model-based. Failures should be categorised rather than merely counted, since an overall score can hold steady while a specific important behaviour disappears.
Related
Part of AI Terminology Index, section: Measuring Performance.