Domain Experts as Eval Builders

LLMs are general. Verification is specific. The people who know what "correct" looks like should be defining the tests — not ML engineers, not prompt hackers.