TWINFORGE

Twinforge turns visual motion into a prediction you can test. Tap an object, draw a box, or choose Hands. Meta SAM 3.1 finds its shape in the actual submitted image. Twinforge tracks the observed pixels and forecasts where the object will appear 0.5 to 1 second later. Lock a target before that moment arrives; the target stays fixed while the interface reveals the later observation and pixel error. We built this for developers and educators who need to distinguish a convincing overlay from a real prediction. It accepts uploaded video, camera input and randomized 3D scenes. The scene renderer never supplies positions, velocities, forces or other hidden physics to the observer. Every run can export image hashes, SAM receipts, observation timestamps, frozen predictions and later assessments. Actual SAM trials include bottle selection, two-hand segmentation and recorded real hand movement. Failed trials remain in the evidence. Camera support is implemented, but physical-phone accuracy is not claimed. This prototype predicts 2D image position; unexpected motion remains difficult, and its uncertainty display is not a guarantee.