A RAG evaluator that admits what it can't judge

Fail-closed groundedness, deterministic corroborators, and a self-test — because an evaluator should be more trustworthy than the thing it grades. The quiet flaw in "LLM-as-judge" evals Most tools tha...