MTS-Bench: A Manchester Triage System Benchmark for Language Model Triage Safety

The study compared the safety of three systems for patient triage using the Manchester Triage System (MTS). Tested were: GPT-5.1, DR. INFO without MTS, and DR. INFO with MTS retrieval on 156 clinical cases covering all five MTS priority levels. GPT-5.1 undertriaged 44.2% of cases including 75% of Red and 73.4% of Orange cases, producing 6 dangerous misses. Both DR. INFO configurations undertriaged only 11.5% of cases with no dangerous misses. When a misleading GP referral statement was added, GPT-5.1 undertriaged all 8 Red cases, while DR. INFO with MTS retrieval continued without errors. Adding objective clinical data reduced undertriage in DR. INFO with MTS retrieval from 19.2% to 3.8%. Conclusion: Replacing a general-purpose language model with DR. INFO system with MTS retrieval substantially reduced undertriage and dangerous miss rates.