Regulatory Sandbox: AI Kill-Switch Containment (2D)

Sandbox Oversight Controls

Sandbox boundary radius 6.0

Kill-switch reaction latency 300 ms

Monitoring sample interval 120 ms

Agent exploration temperature 1.8

Oversight Loop

Monitoring — agent nominal.

Distance to boundary

0.00

Predicted time-to-impact

—

Contained (safe stop)

Escaped (violation)

Containment success rate

—

How it works

Article 57–58 of the EU AI Act lets developers test a high-risk AI system inside a bounded "regulatory sandbox" under live supervision, with an enforceable stop mechanism if the system drifts outside agreed limits. This simulates that oversight loop as a 2D kinematics-and-estimation problem, not just a policy label.

agent position p(t), velocity v(t)  (Ornstein–Uhlenbeck-style walk, 2D)
dv = -γ·v·dt + σ·√dt·ξ            ξ ~ N(0,1) per axis (x,y)
dp = v·dt

distance to boundary   d  = R − |p|
outward radial speed   vr = (p·v) / |p|     (vr ≤ 0 → agent safe, TTI = ∞)
time-to-impact          TTI = d / vr          (vr > 0)

trigger condition (checked every monitoring interval Δt_s):
   TTI  <  reaction latency L + Δt_s   →  kill-switch fires now

outcome, decided after L has elapsed:
   |p(t_fire + L)| < R   →  CONTAINED  (stop landed in time)
   |p(t_fire + L)| ≥ R   →  ESCAPED    (boundary breached before shutdown)

Reaction latency — how slow the real stop mechanism is (network, actuator, human-in-the-loop confirmation). Higher latency needs the trigger to fire earlier.
Monitoring interval — how often the oversight system samples the agent's state. A coarse interval can miss a fast excursion between two samples.
Exploration temperature — how aggressively the test agent probes its environment; higher σ means faster, harder-to-predict radial velocity spikes.
An escape here is exactly the regulatory failure mode a sandbox's stop condition exists to prevent: the containment measure fired, but too late relative to how fast the system was moving toward the limit.
Drag the spatial view to pan, scroll/pinch to zoom — the strip chart below tracks distance-to-boundary (blue) and predicted TTI (amber) over the last few seconds so you can see the trigger threshold being approached before it fires.

Agent (nominal) Kill-switch firing Contained history Escaped history

drag to pan · scroll to zoom

Regulatory Sandbox: AI Kill-Switch Containment Simulator (2D)