Post-Hoc Knowledge Grounding for Verifiable Multi-Agent Search and Rescue Decision Support

Abstract

Multi-agent systems powered by large language models (LLMs) show promise for complex decision-support tasks, yet their outputs often contain hallucinations that undermine trust in safety-critical domains such as search and rescue (SAR). We present a post-hoc verification framework that grounds LLM agent outputs in a probabilistic knowledge graph derived from approximately 12,000 curated cases from the International Search and Rescue Incident Database (ISRID). Our approach extracts structured claims from agent outputs, performs probabilistic reasoning to detect statistical anomalies, and produces tiered decisions (accept, flag, reject) with explicit evidence chains. The framework operates downstream of a seven-agent SAR system, verifying terrain predictions (where the subject is likely found) and status predictions (the subject's likely condition). These claims directly inform search prioritization. Experiments demonstrate that coordinators following our system's recommendations achieve 71.0% accuracy, compared to 51.6% when accepting all LLM outputs, a 19.4 percentage point improvement. The system detects anomalies with an 81.3% F1 score while maintaining practical coverage (46.3% of claims verified). Ablation studies confirm anomaly detection as the critical component, contributing +5.7 percentage points over grounding alone. Our framework provides a practical path toward trustworthy AI in emergency response.