LLM Evals Start as Thirty Test Cases, Not a Platform

Most LLM features ship on vibes and break on the next prompt tweak. A small regression set, cheap deterministic checks, and a judge you actually calibrate will catch most of it.