This project trains the machine that follows KernelMind AI OS. A compact neural world model learns the transition function of an eight-bit virtual computer from current state, capability flags, and one of thirteen OS actions. Training covers the exhaustive 53,248-state/action corpus.

The evaluation separates held-out one-step transition accuracy from six-step autoregressive rollout fidelity. The Space places the neural prediction beside the deterministic reference transition for direct inspection.

It is a learned virtual-machine dynamics model, not a hypervisor and not a model with access to the host operating system.

The 6,937-parameter model reached 99.962% exact transition accuracy and 100% blocked-action accuracy across 5,325 held-out transitions. In a separate autoregressive audit, it reproduced every step and final state across 2,000 randomized six-action rollouts.