Module 10 Summary
What this module established
A gate names a metric, a bar, a comparison and a severity - anything less is a suggestion. Include newly-broken-cases at zero: it catches the change that improves the average while breaking a case that used to work.
Carry forward
- Run the eval set on a schedule as well as on deploys. Model updates change your system without touching your code.
- Prompts and retrieval settings are code and must roll back with it. The old binary running the new prompt was never tested.
- Make gates overridable with a recorded reason and review the overrides monthly. Unskippable gates get deleted; silent ones get ignored.
Before moving on
Move on when your gates run in CI with recorded overrides, the eval runs on a schedule, and a rollback has been proven to restore configuration too.
