How do you make RAG citations actually accurate and verifiable?

Our users are lawyers who will challenge every citation. The model retrieves the right passage but attributes it to the wrong document, quotes slightly wrong, or fabricates a page number. We've tightened prompts and post-process to verify, but verification adds latency and still misses edge cases. What's the architecture teams use for high-stakes citation accuracy — not just grounding, but verifiable attribution at the sentence level?