Medical AI that refuses to guess — diagnosis by deterministic logic, not language model hallucination

A medical AI system stops using language models for diagnosis and instead routes patient symptoms through a rigid decision tree built from expert medical knowledge. In practice, this means the AI can explain exactly why it ranked each diagnosis, catches its own mistakes before they reach a doctor, and doesn't invent symptoms that patients never reported.

Healthcare AI has a fundamental problem: when a language model guesses wrong about a patient's symptoms, there's no way to trace back why or catch it before harm. This system separates the parts of the problem — language models do what they're actually good at (reading messy patient descriptions), but diagnosis lives in a deterministic box that can't hallucinate. The prototype tested on 42 hard pediatric neurology cases got the right diagnosis in the top five 88% of the time, which is meaningful enough to suggest the architecture works. This matters because it shows one concrete way to make medical AI safer without requiring either (a) language models to stop hallucinating, or (b) doctors to trust black boxes.