How the device line separates defensible hospital AI from regulatory liability

Hospital AI lives or dies on where it draws the device line. The HORUS Decision Twin keeps its simulation strictly on the decision-support side: it surfaces distributions and scenarios, but it never autonomously acts.

By treating the guardrail as a first-class asset — recommender-mode by default, evidence-based promotion, and a kill switch — the architecture stays defensible under regulatory scrutiny.

The lesson for the industry: the constraint that looks like a limitation is the thing that makes the AI deployable.