DESIGNING EVALUATIONS (EVALS)

CHAPTER SUMMARY Evals are structured tests with defined inputs, scoring rules, and pass or fail criteria. A dataset holds many eval cases, each ideally paired with a trusted golden reference answer that is reviewed and kept up to date. Task success metrics tell you whether the job was done, and step-level metrics tell you how it was done and where it went wrong; both are needed. Failure modes should be probed on purpose with cases built to trigger them, and a healthy dataset mixes normal, edge, and adversarial cases. KEY TERMS REVIEW Eval: a structured test of an AI system's quality. Golden reference answer: a pre-verified correct answer stored beside a test input. Dataset: a collection of eval cases. Normal case: an everyday typical request. Edge case: an ambiguous or unusual request. Adversarial case: a deliberately tricky or malicious input. Task success metric: a measure of whether the whole job was done correctly. Step-level metric: a measure of something that happened during the run. Failure mode: a recognizable way that a system can go wrong. Leakage: test answers accidentally visible to the system being tested.