A production RAG system for enterprise legal research is experiencing accuracy issues. Users report that Claude provides confident answers citing specific case law, but lawyers verify that the cited cases don't support the stated proposition — the case law exists in the corpus but the holding is being mischaracterized. Retrieval metrics show high recall (relevant documents are being retrieved) but the grounding step is failing. The vector similarity scores for retrieved documents average 0.72. Which architectural change most directly addresses this specific failure mode?

Attribution-aware generation with quote validation is the correct fix for case law mischaracterization. The problem is not retrieval (high recall confirmed) — the problem is that Claude is reading retrieved case documents and generating inaccurate summaries of holdings. Requiring Claude to output exact verbatim quotes for each proposition, then validating those quotes exist in the retrieved documents, creates two catching mechanisms: (1) Claude cannot fabricate a holding — it must find the actual text, (2) the downstream validation layer can deterministically verify quotes against source documents, rejecting mischaracterized propositions before they reach the user. Option B (cross-encoder reranking) improves retrieval precision but the problem is generation, not retrieval — the relevant documents are already being retrieved (high recall). Better reranking would surface better chunks but Claude can still mischaracterize a holding from a highly ranked chunk. Option C (BM25 hybrid) also addresses retrieval, not the generation grounding problem. Option A (k=15) adds more context but more documents means more opportunities for mischaracterization without fixing the core grounding mechanism.