Hierarchical Temporal Runtime Assurance for Controlled Agentic AI Systems: Safety Shielding, Auditable Action Repair, and Bounded Recovery
TL;DR
This study introduces MARIS-TRA, a hierarchical temporal runtime assurance extension of Controlled Agentic AI Systems that provides quantitative robustness for speed, arena, pairwise separation, restricted-zone, and bounded-recovery requirements.
Abstract
Agentic artificial intelligence requires safeguards that remain effective across trajectories rather than only at individual decisions. This study introduces MARIS-TRA, a hierarchical temporal runtime assurance extension of Controlled Agentic AI Systems. A compact, auditable fragment of Signal Temporal Logic (STL) provides quantitative robustness for speed, arena, pairwise separation, restricted-zone, and bounded-recovery requirements. The final architecture uses invariant hard-safety formulas to define feasibility, while recovery/liveness is monitored separately and may trigger escalation. Across a 5750-episode main campaign, MARIS-TRA achieved 100.00% realized hard-safety window satisfaction, 100.00% hard-safety episode satisfaction (450/450; Wilson 95% CI 99.15–100.00%), zero collision episodes, and 2.55 ms mean latency in the core comparison. Under the primary strict numerical semantics, an independent CBF-QP baseline achieved 88.89% hard-safety episode satisfaction; all 50 strict failures were very small arena-boundary overshoots in the boundary-stress scenario, with no collision or separation failures, and the post hoc tolerance sensitivity reached 100% at epsilon = 10−4 normalized simulator units. An additional 8640-episode targeted validation examined recovery hysteresis, model mismatch, AHO control flow, and safety–recovery conflicts. Under confirmatory high-density testing, safety-only shielding preserved hard safety in 810/810 episodes, whereas joint-hard enforcement produced 28/810 separation-safety failures (3.46%; Wilson 95% CI 2.40–4.95%) without collisions. Actuation-noise and delay experiments further show that the formal result is a conditional predicted-trace certification rather than a disturbance-robust guarantee on future receding-horizon execution. The empirical claims are, therefore, limited to the evaluated continuous-action multi-agent setting, while the architecture remains policy-separable and auditable.