También: evals, llm evals, ai evals

Measuring whether a model actually works. Covers offline metrics, held-out and adversarial test sets, LLM evaluation harnesses, human review, and regression tracking between versions.