Every AI-generated output that affects a patient record passes through a clinician review step before it lands. The clinician sees the suggestion, the supporting evidence, and an easy path to accept, edit, or reject.
We measure how often clinicians accept versus edit. A model that's accepted unchanged ninety percent of the time is doing real work; one that's edited heavily isn't ready.
The review step is not a gate to be removed at scale. It's the design.
