Evals are the unit tests of prompts

Every prompt change ships against a golden set. Every incident becomes a case. And the LLM judging the outputs is treated as the biased witness it is.