A free pose model logs 15 good squats. Ours greyed one out and refused to score it. The hip wasn't visible, so we wouldn't guess. That grey rep in the replay is the whole point. It's rep 12 in Maria's session, a synthetic post-ACL patient, where the hip keypoint drops below our visibility gate. The engine stops and prints "Rep not scored, hip not visible." No fabricated angle, no confident bluff. Across our labeled synthetic set, 3 of 3 occluded reps abstained and 0 angles were invented. Why refuse instead of estimate? Monocular knee-angle error runs 9.3 to 21.9 degrees (Nature Scientific Reports, 2025). An AI that quietly guesses through an occlusion isn't handing a clinician verified data, it's handing them a liability. So we built the layer that knows when it doesn't know. Pose estimation is free now. Every PT and wellness app can count reps. The hard, billable part is the brain above the keypoints 🧠: confidence-gated scoring, population-adaptive thresholds, and an auditable session report a clinician can actually act on. The one rep it declines to score is what makes the rest of the scores worth trusting. We'd genuinely like to hear how other teams building remote therapeutic monitoring draw the line between "verified" and "we can't tell." It's a hard call, and getting it wrong is how trust breaks. #DigitalMSK #RTM #PhysicalTherapy #HealthAI #poseestimationtechnology