TRACE-RLM: An Adaptive Trust Controller for Evidence Verification, Provenance Tracking, and Abstention in Long-Context Question Answering

Long-context language models and retrieval-augmentedgeneration systems can answer complex questions from largeevidence collections, yet their outputs may remain unsup-ported, incomplete, or based on broken multi-hop reasoning.This paper presents TRACE-RLM, a non-destructive AdaptiveTrust Controller that keeps direct long-context inference as the primary answer generator while adding hybrid retrieval, claim–evidence certificates, provenance tracking, conflict checks, tar-geted verification, proof completion, and selective abstentionas a trust layer. A key methodological contribution is aninference-time gold-leak barrier introduced after an integrity audit found that earlier experimental paths could expose gold answers or evidence; all contaminated historical results are excluded from the final evidence base. On a clean balanced MuSiQue evaluation of 500 examples, Direct Long Context achieved 30.6% overall exact match and 61.2% answerableexact match, compared with 20.2% and 40.4% for stan-dalone TRACE, motivating the controller-based design. In a preliminary balanced Gemini pilot of 30 examples, Direct plus TRACE improved overall exact match from 33.3% to 53.3%, repairing six Direct errors without damaging any correct Direct answer. However, offline abstention studies under an answerable-pass-through constraint of at least 85% showed that unanswerable abstention plateaued near 42.9% de-spite semantic, provenance, retrieval-rescue, structured-proof, and contrastive-sufficiency refinements. These results position TRACE-RLM as a trust and provenance architecture that may complement direct long-context inference, while also revealing a practical trade-off between stronger abstention and preservation of correct answers under the current architecture.,