Alibi

Alibi shows which scratches on a car are new since you got it, and which were already there. Who it is for: anyone handing back a car that isn't theirs, like a rental, a friend's car or your parents'. Rental companies now scan returned cars with AI and can bill within minutes; one renter was billed $350 for two small dents and had no proof (InvestigateTV and Arizona's Family, June 2026). Your only record is a phone video nobody can search. How you use it: film one slow lap of the car when you get it and one when you hand it back. Alibi makes a sealed sheet that puts the same spot from both videos side by side and says which marks were already there, which are new and which need your eye. It also lists the panels you never filmed. How it works: - Both videos are hashed and timestamped before any model sees them. - Meta SAM 3.1 follows 16 text prompts (car parts, plates, faces) through each shaky walk. Code turns the tracks into 22 panels, so a door is only compared with the same door, and plates and faces get blurred. - Meta Muse Spark 1.3 boxes the marks on one blurred panel at a time. - Plain code, not a model, matches marks across the two walks in each panel's own coordinates and writes every label. - Muse Spark Open, Meta's open-weights model on Baseten, takes a second look before any "already there" or "new" stands. It can only lower the label the distance gives to "needs your eye", never raise it, because a false "new" is worse than no answer. - Muse Voice turns what you say while filming ("trunk lid, two small chips") into timed turns that code pins to the panel on screen as you said it. Results: on a friend's Ford Focus with blue tape stand-ins (one strip added and one removed between the walks) and one real scrape, Alibi got 5 of the 6 test marks right (one only in part), made no false "new", and sent the one it missed to "needs your eye" instead of guessing. In my own checks of 44 labels from four runs, every "already there" (9) and every "new" (2, the same strip in two runs) was right. The example's live run made 504 model calls in 13 minutes for $6.52. Every model answer is stored, so the run folder rebuilds the same labels and sheet data byte for byte, with no new model calls. Try the example, no install: https://walkaround-web-production.up.railway.app/run/hero_focus5e?replay=instant Limits: I tuned the rules on that footage, so it is not a held-out test. Muse Spark boxed 17 of the 53 real marks I labelled by hand on this car's crops, glass and wheels are left to your eye, and model answers vary between runs (the same videos once scored 2 of 6). No real renter has used it yet. And the stand-ins are tape because I wasn't going to scratch my friend's car for a hackathon. Also submitted for Best Build on Meta Open Weight Models: the only model step that compares the two walks runs on Muse Spark Open (meta-models/muse-spark-open-1026 on Baseten), and every call is logged with the model id that answered.