loading
loading
An eval you re-run on every change to catch quality that silently got worse.
You lock in a golden dataset and a passing score, then run it in CI so that tweaking a prompt, swapping a model, or bumping temperature can't quietly break cases that used to work. The point is the diff: 'this prompt edit fixed 3 cases but broke 7' is invisible without it. Example: a gate that fails the PR if accuracy on your 100-case set drops below 92%.
Plainly
Think of Regression Test as a simple recipe for doing the work better. An eval you re-run on every change to catch quality that silently got worse.
In practice
Use it when you need a repeatable method instead of guessing from vibes. In practice, define the owner, input, output, and failure mode before you rely on it.