Rubric-based eval harness for LLM/agent outputs, backed by the typesafe-ai/jev evaluation model via Vercel AI Gateway. Cheap enough (~$0.04/1M input tokens, one round trip per case) to run on every PR.