LLM Evaluation Harness: How to Know an AI Feature Is Reliable Enough to Ship

An AI feature needs a way to prove that changes are helping, not quietly breaking behavior. That is the job of an evaluation harness.