eval-driven development
Deciding what a good output looks like and writing evals for it before changing an AI feature, then running them on every change.
In more detail
It is test-driven development for probabilistic systems: define success criteria first, design outputs so code can grade them, and let the eval results, not impressions, decide whether a change ships. Every production failure becomes a new case before the fix is written.
Related terms
- AI eval — A repeatable test of an AI feature: a set of inputs, a way to grade each output, and a score you track across changes.
- golden set — A curated set of inputs with known good outcomes, used to check that an AI feature still behaves after every change.
Taught in
- Eval-Driven Development — AI Engineering: LLM Apps, RAG & Agents