# Definition of done for an AI feature

Source: https://nativelayer.ai/definition-of-done. Use it on any vendor, including NativeLayer AI.
Done means your team can run it without us. All six exist in your repositories and accounts.

## 1. `evals/` — An evaluation suite that runs in CI

- [ ] A golden set of real cases with the expected answer written down.
- [ ] Every prompt, model or retrieval change is judged against it in CI.
- [ ] A regression shows up in the build, not in a demo.

## 2. `infra/` — Infrastructure as code, versioned and reversible

- [ ] Your Google Cloud project, your identity provider, your logging.
- [ ] Model and prompt deploys are versioned and reversible.
- [ ] The rollback path is documented in the runbook.

## 3. `RUNBOOK.md` — A runbook written for the person on call

- [ ] Each alert: what it means, the first action, the rollback command, who to call.
- [ ] How to operate, monitor, and roll back.
- [ ] What “normal” looks like on the dashboards.

## 4. `decisions/` — One record per model, prompt and data decision

- [ ] What was chosen, what was rejected, and the evidence from the evaluation suite.
- [ ] Data-handling implications reviewed before production.
- [ ] Written so your team can defend the decision later without us.

## 5. `dashboards` — Cost, latency and quality your team reviews weekly

- [ ] Cost and latency, reviewed weekly by your team.
- [ ] Quality from the evaluation suite over time, not a one-off demo.
- [ ] Alerting and runbooks in place before launch.

## 6. `your repo` — All of it in your repositories and accounts from the first commit

- [ ] Nothing transfers at the end because nothing ever left.
- [ ] A repository with a green build.
- [ ] Your engineers pair with ours until they run it alone.