RiskWeave: Allocating Evidence by Omission Cost in Medication-Safety RAG

Medication-safety RAG is governed by its most consequential omission: under a ten-item evidence budget, nine relevant passages cannot compensate for a missing contraindication, severe drug–drug interaction, or organ-specific dose rule. RiskWeave turns context construction into clinical-priority support allocation. It infers required medication–patient facets, scales their aggregate omission cost by pre-retrieval query risk, and greedily maximizes a monotone-submodular objective over provenance-bearing assertions; conflict reservation, facet-targeted re-exposure, and residual-support thresholds carry the same support state into citation and abstention. A 200-question frozen-pool control holds candidate access, passage retrieval and packing, prompt, generator, and budgets constant; a matched ablation removes priority weights while retaining the other policy stages. On 400 annotated questions, Facet Coverage@10 is 0.731, versus 0.689 for a flat cross-encoder (paired difference 0.042, 95% CI 0.018–0.067) and 0.612 for uniform submodular coverage. Across 847 questions, blinded grading gives RiskWeave 87.8% acceptability, versus 84.1% and 81.7%, respectively. In the frozen pool, unsupported high-priority claims are 8/200 versus 18/200 for uniform coverage; removing priority weights lowers high-priority coverage from 0.782 to 0.641. With graph-free exposure, RiskWeave reaches 0.715 coverage versus 0.674 for the cross-encoder; with Llama-3.1-70B-Instruct, acceptability is 83.8% versus 80.0%. On 423 eICU questions from 208 hospitals, it reaches 85.3% acceptability versus 78.7% for the strongest tested comparator. At the clinical interface, complete RiskWeave assistance lowers clinician major errors from 88/992 to 23/992 and median task time from 10.4 to 5.8 minutes.