Ultimate Truth

Robots are about to take language models at their word, and a model's word is often wrong: in our own log, every wrong name a model gave the robot vacuum ("pet cat", "dog", "plant") passed the usual format check. A robot that believes that will eat the wrong thing or hurt someone. This project is the missing layer between a model and a body: Meta's models may propose, and the mind decides. Built on ultimateTruth, an embodied mind with a cognitive model on a mesh chip and a deep brain that rehearses in a world model, the Meta layer makes SAM 3.1 the robot's eye, runs Muse Glimmer 30B on the robot's own NVIDIA Jetson GPU with no cloud, and runs Muse Spark 1.3 on the Meta Model API for the same prompts. A model names a thing, reads a page, or composes a procedure; a person can point and say through a WebXR page. Every proposal passes a gate, is held until the body has watched the thing do what was claimed, is rehearsed by the mind before it can act, and is kept with its verdict in a public proof register. The result is a robot that learns what things are at once, from a picture plus what it saw the thing do: no retraining, no new round of trial and error, and a receipt for every decision. Measured on nine things: 8 of 9 named from a picture, 9 of 9 with the body's observations, and a live rehearsal that caught a well-formed but useless plan. The whole chain runs offline on open weights or online on the API, is Apache-2.0, and can sit between any language model and any actuator.