TinyTMS

tinyTMS.IO checks a freight load in at the dock with three witnesses and one judge. Every witness is a Meta model: - Vision: SAM 3.1, through the Meta Model API, counts pallets in the load photo with the concept prompt "pallet". - Paper: Muse Spark 1.3 reads a photo of the bill of lading (a Mexican Carta de Porte) into fields: pieces, weight, seal, shipper, consignee, hazmat. - Voice: Muse Voice Transcribe turns the driver's voice note into words; plain code reads the count and the seal from them, and leaves a field empty, with the reason, when the driver hedges or corrects a number. - Offline, served on our own hardware: Muse Glimmer 30B, Meta's open weights (4-bit, through mlx_vlm on our own Mac), reads the same paper and photo with no network at all. - The driver's explanation: on a HOLD, Muse Spark 1.3 explains the judge's reasons to the driver in Spanish or English. It never sees the image or the audio, and it cannot change the verdict. - Every SAM 3.1 answer can be drawn back onto its own photo from its signed receipt, and only if the receipt verifies, so a reviewer sees exactly what was counted. That includes a wrong answer we kept as it was: asked for a projector, SAM boxed the VR headset. The judge is plain Python with no model inside: ACCEPT or HOLD, and fail-closed, so a missing answer is a HOLD. Every check-in becomes an Ed25519-signed receipt that names the model and provider behind each witness. Anyone can verify who said what.