Phase 08Evals are the deliverable418 of 434
A dataset made of things that actually happened
Fifty real inputs beat five hundred invented ones. Build the pipeline that turns production traffic into test cases.
Build40 minAI pair
Build this in your own editor
This one runs on your machine rather than in the browser workbench. Work to the outcomes below, then come back and mark it complete.
You should now be able to
- Capture and label real inputs as eval cases
- Cover the failure distribution, not just the happy path
- Keep the dataset out of the prompts it grades
Loading…