Yula

Marketplace listings are written from memory or supplier sheets, and the mismatch surfaces when the buyer opens the parcel. Writing them is also a chore. With ClaimCheck, a seller points a camera at the product and talks. Muse Voice Transcribe streams live captions. Muse Spark turns each sentence into facts a camera could check. SAM 3.1 hunts for the evidence on every frame and coaches the seller with a single line: closer, hold still, show me the rest. The shutter fires itself. The mask flattens the label, Muse Spark reads it, and plain Python rules decide counts, fibre percentages and colors. Wrong facts are struck through while filming, and evidence that can't be seen leaves a claim open rather than calling it wrong. It also flags damage the seller didn't mention, like a hole or a stain, for them to confirm or dismiss. The seller leaves with a title, a description, a fact table with evidence behind each row, a photo in which only the background is generated, and a proof reel tracked by SAM 3.1 video. A person edits and decides; nothing is published automatically. Then the footage answers for itself. A buyer types or says a question nobody planned for, like "does it have a pocket?". SAM 3.1 video searches the whole take, Muse Spark confirms what it found, and the answer is that moment, spotlighted. Settled facts are answered from their verdicts in code, so a question can't overturn them. If two models can't find something, it says "not in this video" and that this isn't a no. If no video could show it, it says so instead of guessing. On 28 real Wikimedia Commons photos labelled by hand, 70 of 74 claims were correct, 4 were left open and none were wrong. On 25 questions about generated clips of products it wasn't tuned on, 24 got the right kind of answer, with nothing wrongly claimed as found. I picked and labelled these sets myself, and they are small. A real seller filming through the live flow has not been measured yet. A replay of one real run, with no install or key: https://coderunner24.github.io/claimcheck/ The demo video runs the live app and live Meta models on the app's sample clip, which was made with Muse Image and uses a scripted seller. All code, prompts, the evaluation set and the UI were built during the event.