The jersey that counted as a person
An empty chair kept Ash convinced someone was there. A one-change experiment explained why.
Record date Sep 28, 2026 · Published
Leaving the room should have produced a person-lost event. It did not. The service was healthy, but its view of the room was wrong.
The suspect was an Eagles jersey draped over a gaming chair. In the empty-room diagnostic, the scene produced 124 person detections, including confidence values around 0.89–0.93. The torso-like appearance was a hypothesis; the next test checked the environmental trigger.
Only the jersey was removed. The chair stayed put, and camera and software settings stayed the same. The repeated diagnostic returned zero person detections and 224 chair detections.
Raising a threshold would have been a poor answer to this particular problem: the wrong predictions were already highly confident. Removing the jersey resolved the tested scene, while the broader recognition problem stayed open.
It was a useful reminder to look at what the camera actually sees. A service can be running correctly and still be persistently wrong about the room.