Experiment Note
The runtime shield monitored measured physical gap g_i and estimated time to collision:
TTC_i = g_i / max(v_i - v_(i-1), epsilon)
when the follower was closing on the predecessor. If gap dropped below g_min or TTC dropped below T_min, the controller replaced the trust-weighted cooperative command with a conservative fallback command.
The shield was logged through activation counts. These counts were treated as intervention-frequency diagnostics, not as a metric to minimize independently. A controller with zero shield activations can still be unsafe if it never detects danger; a controller with many activations may be correctly preventing an attack-driven failure.
Technical Essay
The shield resolved a structural weakness in learned security controllers. A trust score is probabilistic and can be wrong under distribution shift, collusion, severe delay, or an unseen attack. If it directly authorizes physical action, a single high-confidence error can have a safety consequence. The shield changed the architecture from "trust the detector" to "use the detector for performance, but constrain the final action using physical safety evidence."
This did not create a formal control barrier function or certified invariant set. The shield was a simple runtime override evaluated empirically. The paper therefore avoided claiming certified safety, formal string stability, or deployment readiness.