Wrap the RAG assistant from part 5 in a containerized DeepEval suite judged by gemma4 on the WEC Inference API — no OpenAI key anywhere. Then use it to answer a real question: should we swap our generation model? Same quality, 5.5× the tokens: the eval says no. Every command, number, and error is real.