LineProof: Contact-Verified Robot Simulation

LineProof makes robot progress auditable: it checks whether an object stays grasped, travels above the table, lowers with support and releases stably, instead of accepting a lucky final position. During development, this exposed a real false positive: a cube slipped during transfer but landed near its target. The corrected scorer rejects that trace, with a regression test. Built for the Intel online robotics track, the prototype uses MuJoCo, pinned public SO-101 assets, two arms, two cameras and 12 bounded joint targets. A scripted IK teacher controls the left arm; the right arm stays parked. The nominal cube trial sustained a 79.54 mm lift and ended 9.14 mm from its target. Ten small cube perturbations produced six passes and four failures, all retained. Separately, a random-weight ACT model converts to OpenVINO and runs on an Intel i5-1245U CPU. Five measured calls gave a 42.649 ms OpenVINO median versus 60.634 ms in PyTorch, with numerical parity checked on two fixtures. These are runtime checks, not learned task results or a production speedup claim. The repository includes reproducible commands, licenses, raw traces, scoring logic, failure evidence and 13 passing tests. This is a partial research prototype: there is no task-trained checkpoint, VLM planner, model-driven action, coordinated dinner-table sequence, Core Ultra test or physical robot test. The next gate is verified demonstration collection and bounded closed-loop policy training.