A controlled comparison of a top-5 RAG pipeline and a full 127,000 token prompt on the same 12 questions, same system prompt and same model. Graded blind on correctness, completeness and grounding.