When Generated Answers Become Evidence: Do Errors Compound in Self-Updating Retrieval-Augmented Generation?

A self-updating support system stores the answers its users accept, so the next similar question can be answered from them. In deployed retrieval-augmented generation (RAG) systems the stored entry is a generated answer, and a retrieved entry becomes the evidence for the next generated answer. We measure that loop on real repeat questions from four CQADupStack forums, with community duplicate links as ground truth, three retrievers (bge-small, MiniLM, BM25), answers generated by gpt-5.4-mini and replicated on android with an open-weight model, and only user acceptance simulated. Four hypotheses stated before the runs: (H1) source bias, not supported: storing the question with the generated answer is the safest entry design we have measured; (H2) compounding, weak: entries generated from entries form and drift slightly, but cause a small share of wrong answers; (H3) the first-order crossover estimate from our previous paper needs no depth correction; (H4) reopen demotion is the best countermeasure, keeping the gain at low wrong-acceptance rates and improving the outcome at high ones. Includes a 300-answer faithfulness audit with two judges. Follow-up to "When Accepted Answers Become Evidence" (doi:10.5281/zenodo.23089817). Code, generated answers and the scripts that produce every table: https://github.com/pavansky/self-updating-retrieval