Seeing more without deciding more: eight principles for bounding the authority of AI in clinical decision support, from one physician-governed system

A self-contained condensation, for international readers, of the ten-paper H-DKPS series on physician-governed clinical AI (the eightfold principles of knowing-equilibrium). It reports no new experiment; every number is taken from the paper that reports it. Clinical decision support increasingly lets a language model propose while something else must decide. We report what one physician-governed system had to satisfy for that division to hold, condensed from a ten-paper series. A deterministic kernel is the only path to an output, and the physician decides. Four principles bound the decision itself: safety, bounded liveness, safety over trajectories and validation at the human interface. Four more extend what the system can see — reasons, measurements, rules and populations — without extending what it may do: none can refuse a physician's order. Each principle was pushed until its method stopped working, and the stopping points are reported as numbers. Two lessons recur: conforming to an invariant set did not entail the property it was meant to secure, and a check that always passes must first be questioned. No part reports a clinical outcome.