Certified Domain Consistency for Multi-Domain Retrieval: Label-Free Per-Domain Contamination Control with Conformal Risk Guarantees

arXiv cs.AI · 6 hours ago

Researchers introduce C3R, a control layer for multi-domain retrieval systems that certifies per-domain contamination budgets without requiring query-time labels, using conformal risk control to prevent returning relevant but wrong-domain evidence. The method uses a two-split scheme that guarantees violation-free performance across resampled calibrations while maintaining recall better than existing cascade approaches at equal contamination levels. The approach enables retrieval systems to either certify domain consistency or abstain rather than silently mixing evidence across domains.