SceneTrace

SceneTrace follows one object through a short, fixed-camera video and builds an evidence-backed history of it: when it was first and last seen, what happened to it, and who, if anyone, the footage shows handling it. Every statement in an answer links to the original frames behind it. When the footage does not show what happened, SceneTrace says so instead of telling a fluent story. The problem: ask a general video model "who took the bottle?" and it will usually name someone, even when the act is hidden or the person only walked past. In offices, shared spaces and small investigations, a confident wrong answer is worse than no answer. SceneTrace separates what the camera saw from what it did not. How it works: you choose the object on one frame by describing it or drawing a box. SAM 3.1 tracks it and the people near it. The engine turns changes in the object's motion into candidate events and pulls the original frames before, during and after each one. Muse Spark assesses those frames against a strict schema, then deterministic rules decide whether each event is supported, possible or rejected. Closeness is never treated as contact. Questions go to a bounded Muse Spark investigator that can read the stored history but may only submit structured claims; the server checks every claim against the stored events and writes the sentence itself. An optional Muse Wonder review checks the answer against the frames in a fresh context and can block a claim, never promote one. Evidence can be sealed into a content-addressed bundle whose expected head is kept separately, so tampering is detected. The demo uses a real office clip filmed with my colleagues' permission. I walk past the stain-remover spray bottle and am cleared, because the bottle does not move. The bottle then disappears behind two people. The lift itself is hidden, so SceneTrace reports a possible pick-up by the one person the bottle was next seen moving with, and states that the footage does not establish who took it. Measured so far: the claim validator blocked all 30 planted false claims and none of 10 clean ones; a cold analysis takes about half a minute and costs around USD 0.09. Real-video accuracy has not yet been measured on a labelled set, and the README says so. It runs locally or in Docker, and the recorded runs replay without a key or any model calls.