Retrieval-augmented answers over a real document set from your vertical. You write the golden questions, with answers and supporting passages, before any build, including questions the documents cannot answer, because refusing correctly is part of the job.
And midweek, the ground moves: a model you depend on changes price or disappears, while you are mid-build. Your evals tell you what to do next. That is the taste of real life, scheduled.
It ends with
A golden-question set, a retrieval design, and a migration record showing you survived the change.