AI
Preprint applies temporal constraints to reduce surgical-workflow ‘hallucinations’

John S. H. Baxter and Pierre Jannin frame impossible workflow transitions as measurable hallucinations and combine temporal-logic constraints with probabilistic modelling. Simulations on robot-assisted hysterectomy phase recognition report roughly 10% better accuracy and removal of most topological errors.
Why it matters: Safety-critical AI needs constraints that reject impossible sequences, not only aggregate accuracy. Impact: The result is promising for assurance layers, but it is a preprint, author-reported, simulation-based and not independently reproduced.
arXiv preprint




