arXiv cs.LG
7/17/2026

Certified Domain Consistency for Multi-Domain Retrieval: Label-Free Per-Domain Contamination Control with Conformal Risk Guarantees
Short summary
C3R is a drop-in control layer for multi-domain retrieval that certifies per-domain contamination budgets without query-time labels, abstaining when guarantees cannot be met. It uses a two-split scheme with risk-controlling prediction sets and finite-sample transfer bounds from inferred to true domains. Across a thousand resampled calibrations the certificate never violates, while an LLM-judged probe shows wrong-authority grounding rises with contamination and falls under control.
- •C3R certifies per-domain contamination budgets in multi-domain retrieval without query-time labels
- •Two-split scheme with finite-sample transfer bounds never violated across 1000 resampled calibrations
- •LLM-judged probe confirms wrong-authority grounding increases with contamination and decreases under control
Generated with AI, which can make mistakes.
Is this a good recommendation for you?