In 2020, an automated soccer camera filmed most of a match locked onto a bald linesman's head instead of the ball. The system wasn't broken. It was doing exactly what it was built to do. Under floodlights, a shiny bald head reads to an object detector like a ball, scoring around 98% confidence, higher than the actual ball. The camera follows the highest-confidence guess and swings off the pitch. It really happened, at Inverness. (Pixellot has since patched that exact case, but the failure underneath it didn't go away.) So we built Plumbline, a layer that sits after your detector and checks the physics instead of the pixels. Three plain-code tests run on every "ball" it's handed: → Could a real ball move like that from where it just was? → Does the on-screen motion match? → Could an object that size even be a ball at that distance? A head, at head height, barely moving, fails all three. The 98% detection gets rejected, the camera stays on the real ball, and every accept or reject writes down the physics reason behind it. We rebuilt that 2020 moment to test the gates, then ran the same three checks on real broadcast footage with a real detector. Same code both times, and the impossible "ball" never gets through. None of this is about a better detector. A physically impossible detection is wrong no matter how sharp the model gets. If you build anything that acts on what a camera sees, that is the one you want caught before it ever reaches air. #ComputerVision #SportsTech #BroadcastTech #MachineVision