Retrieval-Verified Generation: A Release Contract for Retrieval-Augmented Answering

Question answering over an organization's documents can promise accuracy, or a structure of responsibility. We describe Retrieval-Verified Generation (RVG), a controller that releases a sentence only when it is attributed to evidence valid at the question's reference time and visible to the requester, and hands everything else to a person. A judge-free audit finds no attribution violation in any audited setting, including 23,401 releases from three live runs over 13 knowledge bases in six languages (1,869 to 5,233,235 passages), and recorded decisions replay exactly. Given identical model outputs the controller never changed its judgement (0/3,307), but asking a question three times changes the end-to-end decision for 10.6% of questions, 99.9% of it through writer sampling, concentrated on high-risk questions. Attributed answers are still wrong 4.61% of the time on static knowledge bases (1.64% after reference defects are removed) and 3.5% (2.8%) at 5.23 million passages. Between 15% and 38% of answerable questions are handed off, and about two thirds of withheld drafts are correct. At its deployed operating point RVG's error rate is lower than the same model's plain RAG (6.2% vs 8.0% at 5.23 million passages; 8.4% vs 13.6% over 13 knowledge bases) at the cost of fewer correct answers, which pays when a wrong answer costs more than three to six hand-offs. At equal coverage it is not more accurate than plain RAG with post-hoc selection. We also report an opt-in hand-off packet and a decision cache adopted by product decision after two pre-registered acceptance tests narrowly failed.