Build an assistant that answers only from governed domain data and shows its evidence. Teams will define a domain slice, curate a small corpus, implement grounded retrieval with cite‑or‑fail, and add refusal logic and PII guards. Session will capture traces linking prompt to sources to filters to answer and practice a review ritual to refine boundaries and tests. Most time is hands‑on in pairs with quick checkpoints. You leave with a slim checklist and a repeatable evaluation harness.