Build a small Python RAG evaluation harness, separate retrieval failures from answer errors, and test missing evidence and document permissions.