arXiv:2607.00334 reports 99.6% anomaly detection in three-agent UR5 cell with GEAR runtime

The paper defines GEAR as a set of gears (Observe, Suggest, Plan, Execute, Intervene) under utility-gated dispatch and event-driven fallback. Single-agent proofs establish monotonic stability and equivalence to a gear-constrained MDP. Multi-agent extension maps runtime evidence into the four SMART governance states and applies swarm Lyapunov analysis plus rendezvous control to guarantee zero collisions under stated assumptions.

Evaluation used fault magnitudes drawn from the NIST Degradation Measurement of Robot Arm Position Accuracy dataset. The three-agent cell recorded 99.6% detection against 2.1% for the baseline, 3.5x lower latency, and an explicit physical-workspace safety certificate. Execution gears function as micro-permissions that separate action control from the higher-level SMART governance layer.

Prior CPS safety work relied on static rule sets or offline verification. GEAR shifts enforcement to verifiable runtime state machines that remain active during continuous operation. This architecture directly addresses behavioral instability and continuity loss that static policies cannot contain once agents leave human oversight loops.

Operational integration requires embedding gear authority inside existing robot controllers and exposing per-agent state to the SMART consensus layer. Future deployments will test whether the formal zero-collision bound holds when sensor noise exceeds the NIST calibration envelope.