Eval a prompt on a real model

What to take away

• An eval is a fixed set of inputs with known answers. Run it before and after every prompt or model change.
• Structured outputs make grading a plain comparison in code. Use a model as judge only when there's no single right answer.
• Most misses come from labels nobody defined. Write the definitions the customer's team already uses.
• Watch for regressions, not just the score: a change that fixes two cases and breaks one is worth knowing about.
• Ten cases teach the idea. A real golden set comes from real tickets, including the ugly ones, and grows with every bug report.