Tiny evals that catch big regressions

Every time I swapped a model or a prompt, some quiet behaviour broke. Fifty hand-written cases with exact-match or simple regex checks now run in under a minute before any change.

{"input": "Extract the date: 'see you on the 3rd of May'", "expect": "05-03"}

It is not a benchmark, and that is the point: it encodes the failures I have actually seen.