Regulatory Sandbox: AI Kill-Switch Containment Simulator

The EU AI Act lets a high-risk AI system be tested inside a supervised "regulatory sandbox" — a bounded environment with an enforceable stop mechanism if the system's behaviour drifts too far. This simulator renders that oversight loop as a real kinematics problem: a test agent wanders a bounded 3D space under a noisy, damped random walk, an oversight loop samples its position and outward radial velocity at a tunable interval, estimates time-to-impact against the sandbox boundary, and fires a kill-switch whenever that estimate falls below the reaction latency plus the sampling gap. Whether the stop lands before the boundary is actually crossed depends on the race between reaction latency, monitoring cadence and how erratically the agent is exploring — tune all three live and watch the contained-vs-escaped tally update in real time.