Evals and quality

Normal software has a right answer. You click the button, the total should be $42.10, and a test checks that it is. AI features rarely have one right answer, and they do not give the same answer twice. So the question "How would you know it works?" is harder, and interviewers ask it constantly.

The answer is evals: a repeatable way to measure the quality of a model's outputs against criteria you defined. This lesson covers the Proof gate before launch: what to test, how to grade it and how to keep quality from slipping later.