It would be easy to build a version of this product that quietly optimizes for one number — findings per scan, or time saved per review — and let that number creep into places it shouldn't. Oralytix is built to resist that specifically.
Every finding the model produces is labeled as a candidate for review, never a diagnosis. Every candidate carries a visible confidence score instead of a flat "detected" label, so a 52% finding reads differently than a 96% one. And structurally, nothing written by the model reaches the patient record without passing through a dentist's confirm, dismiss, or modify decision first.