Ctrl+Alt+Deliver

Tally Pack: Muse places the order. Tally proves it was packed right. When an AI agent like Muse buys something for you, a person still packs the box by hand, and 2-3 orders in every 100 ship with a wrong item (about $22 each). The customer blames the agent. Tally Pack checks every box at the packing desk before it ships, using only a camera: no barcodes, no new hardware. A fixed overhead camera watches the desk. SAM 3.1 tracks every item and the box. A plain-code ledger and verifier (zero model calls) checks five invariants: the box matches the order, nothing appears or vanishes, no track teleports, no ID switches, every product is clearly identified. If the evidence is consistent it issues PASS or FAIL with the exact reason ("missing: 0 lays in box, order says 1"). If not, Muse Glimmer 30B investigates: it looks at image crops of the items and picks one candidate repair (e.g. merge two IDs after a tracker switch), and the repair is kept only if the verifier confirms it. If nothing fits, the box goes to a human. Glimmer can never clear a box on its own. Customers' free-text orders ("2 maggi, one parle g and a coke") are compiled into validated contracts by Muse Glimmer. Results on PackBench (30 synthetic fixed-camera packing clips, each also run with an injected tracker ID switch): 15/15 packing errors caught, 0/30 false clearances, 58/60 correct verdicts; the 2 others were honestly sent to a human. Everything runs on one A100 with Meta open weights. Live demo: https://yashwanth-maram.github.io/tally-pack/