Release notes by agents
Before each release, someone read every linked ticket and wrote the customer docs and email by hand. Now a person pastes one link. Code pulls the ticket IDs, because an ID has to be exact. Five model agents read the page, classify each ticket in parallel, draft the docs and email, and review all of it against the source. Nothing ships until a person approves the reviewer's flags.
One release, start to finish
Seven tickets go in. Step through what happens to them, and who does each part. REL-4821 is invented; the stages and prompts are the real ones.
How I know it works
Code checks what can be counted. Golden sets test what needs judgement: cases I answered before the model ran. The reviewer and these evals shipped with the pipeline, and they re-run on every prompt or model change.
Each eval sits on the stage it tests. Pick one above to light it up. Swipe the chain sideways.