Architectural Safety Mechanisms for Multi-Agent Clinical LLM Systems Under Knowledge Base Distribution Shift

The study evaluated whether multi-agent architectures with large language models (LLMs) featuring explicit safety verification maintain clinical guideline compliance when their knowledge bases shift over time or across institutions. The research tested four architectural modes on 50,000 synthetic patients with type 2 diabetes, chronic kidney disease, and hypertension under four shift conditions: baseline, temporal drift, institutional terminology change (11 term pairs with similarity degradation to 0.36), and metadata erasure. The task involved medication reconciliation with contraindication detection. The linear multi-agent architecture without safety mechanisms failed under terminology shift, with guideline compliance score dropping from 1.00 to 0.36. The stateful graph architecture with a safety floor maintained a score of 1.00 across all shift conditions. However, the safety mechanism added 32.2 seconds of latency per patient compared to 12.5 seconds in single-agent mode. Architectural choice determines guideline compliance more than retrieval quality or model size.