LogiMed-RoB Benchmark Reveals LLM Reasoning Collapse

The LogiMed-RoB benchmark, based on Cochrane Risk of Bias 2.0 logic, shows that while LLMs achieve 98.88% atomic consistency, end-to-end logical consistency drops to 45.13% in medical risk assessment. This error compounding effect is particularly severe in open-weight architectures.