Provenance-Aware PCR6 for Conflicting Evidence in Retrieval-Augmented Generation

Resumen

Retrieval-augmented generation (RAG) must often combine passages that are ambiguous, redundant, or mutually contradictory. This study asks whether proportional conflict redistribution from the Dezert-Smarandache theory (DSmT) provides a measurable advantage over Dempster-Shafer (DS) combination when both rules receive the same evidence and the same control for source dependence. Eight fusion methods were evaluated on 500 RAMDocs (Retrieval with Ambiguity and Misinformation in Documents) questions in three conditions. Under duplicated misinformation, provenance-aware PCR6 (Proportional Conflict Redistribution rule #6) reached 80.6% exact-set accuracy and 92.31% set F1, against 78.2% and 90.98% for provenance-matched DS; the exact-set difference was 2.4 percentage points (95% bootstrap interval [1.2, 3.8]) and the F1 difference was 1.33 percentage points (95% bootstrap interval [0.60, 2.33]). With NLI (natural language inference)-derived masses, PCR6 and DS selected identical answer sets, but PCR6 reduced Brier error. A preregistered 1,000-item multi-source experiment using singletons, proper unions, and total ignorance was followed by external validation on 428 RAMDocs items. In controlled partial-ignorance conflict, PCR6+ reduced BetP (betting probability transformation) Brier error from 0.3185 for PCR6 to 0.2733 and remained exactly neutral to an added vacuous BBA (basic belief assignment), although the DS rule achieved the lowest BetP Brier error of the three rules in that same controlled evaluation (0.1878). On the prespecified 260-item RAMDocs applicability subset, ordinary PCR6 had lower BetP Brier error than PCR6+ (0.2354 versus 0.2629), while both achieved set F1 = 0.7718. Exploratory decision sensitivity showed that maximum belief gave PCR6+ its best hard-decision result (exact set 0.5077; set F1 0.7859), whereas DSmP (Dezert-Smarandache probability transformation) sharpened rankings but worsened the RAMDocs Brier score. The evidence therefore supports a bounded claim: PCR6+ corrects a structural multi-source weakness of PCR6 under broad or vacuous focal evidence, but the best fusion and decision rules depend on the downstream loss and the mapping from text to BBAs.