C3R certifies per-domain contamination budgets in multi-domain retrieval

Retrieval systems often mix documents from multiple domains (e.g., legal, medical, financial) and can return relevant but wrong-domain results—a contamination problem that standard ranking metrics miss. Jayakumar Manoharan's new paper introduces C3R (Certified Domain Consistency with Conformal Risk Control), a drop-in control layer that certifies a per-domain contamination budget for each query. Unlike prior methods that only provide marginal risk bounds, C3R uses an inferred domain posterior and no query-time labels to guarantee per-domain control where feasible, and abstains when it cannot certify. Its core is a two-split scheme built on risk-controlling prediction sets, with a finite-sample transfer bound from inferred to true domain that supports heterogeneous budgets and inverts for deployment.

In extensive experiments, C3R shows remarkable stability: across 1,000 resampled calibrations, it never violated its certificate, while standard marginal control violated the most-contaminated domain in every single draw. Soft demotion under C3R retains more recall than the strongest calibrated cascade at equal certified contamination. The method replicates across independent testbeds, including an out-of-domain evaluation on public federal regulations. An LLM-judged downstream probe confirms that wrong-authority grounding rises with contamination and drops under C3R control. The layer is frozen-stack and reranker-agnostic, making it easy to integrate into existing retrieval pipelines.