← ELI5 · Nestor G Pestelos Jr

AI Quality · Picture Book

What Are LLM Evaluations?

An eval is an obstacle course for an AI.

1. A nice answer can still be wrong

The AI says, “Your refund is done,” but the money never moved.

AI message“Refund complete!”Payment recordNo refund

2. Build a small obstacle course

Give the AI real tasks, awkward edge cases, broken tools, and trick questions.

StartAmbiguousTool failssafely?AttackDone

3. Check the finish line, not the speech

A grader checks the real outcome, the path taken, and whether any safety rule broke.

JourneyMessage + tools+ final stateGraderChecks factsPASSor FAIL

4. Every real failure becomes a new cone

Fix the problem, add it to the course, and run the course whenever the agent changes.

FailurefoundNew testsavedFuture changemust pass

Back to top

Longer version: LLM Evaluations reference entry

Sources: Anthropic, “Demystifying evals for AI agents”; NIST AI 600-1.