A 1.0 confidence score can tell you less than a 0.92. That sounds strange until you read @vangrid_io's early access API docs. They explain that a 1.0 score backed by two devices may carry less weight than 0.92 backed by fourteen. The score measures how strongly independent observations agree, not the probability that a scene is correct. The docs also describe flagging conflicting reports instead of quietly merging them into one answer. Picture a loading bay. One phone records it while a truck blocks the entrance; another films it after the truck leaves. You'd want to know when each recording happened and how many observations support the result. This is a design described in early access documentation, not evidence that the full system works at scale today. But it's a useful question for physical AI: what sits behind the confidence score?