Original title: Hierarchical Temporal Runtime Assurance for Controlled Agentic AI Systems: Safety Shielding, Auditable Action Repair, and Bounded Recovery
Tymoteusz I. Miller · September 2026
Autonomous AI agents need safety rules that protect them over entire journeys, not just during single choices. To solve this, researchers built MARIS-TRA. It is a system that monitors agents using a clear, inspectable set of mathematical logic rules. These rules measure safety margins for speed, arena edges, distance between agents, restricted zones, and recovery limits. In this setup, unbending safety rules decide what actions are allowed, while recovery goals are tracked on their own so they can ask for help if needed. The researchers tested the system across 5750 test runs. MARIS-TRA achieved 100.00% realized hard-safety window satisfaction, 100.00% hard-safety episode satisfaction, and zero collision episodes. It also ran quickly, averaging 2.55 ms in latency. The team discovered that keeping hard safety separated from recovery targets prevented safety failures, whereas enforcing both as strict requirements caused separation errors.
Imagine you are packing a suitcase with fragile items. You have a rule: never put something on top of the eggs (hard safety). But you also want to fit everything (recovery). If you focus only on not crushing the eggs, you might leave half the items out. If you focus only on fitting everything, you might break the eggs. Here, the system enforces the 'never crush' rule absolutely, while separately adjusting packing order to maximize fit. In the paper, hard-safety formulas like 'no collisions' are monitored independently from recovery actions that try to bring agents back to desired positions.
Agentic AI systems — like autonomous vehicles or robots — need guarantees that they never violate critical safety rules. This system provides an auditable, hierarchical safety shield that perfectly prevented collisions and other violations in extensive tests. It separates inviolable safety from recovery to avoid conflicts, bringing certifiable safety closer to real-world deployment.
The empirical results are limited to continuous-action multi-agent settings. The formal guarantee is a conditional predicted-trace certification, not a disturbance-robust guarantee: it does not account for future uncertainties like actuator noise. Also, while hard-safety was perfect in the main campaign, joint-hard enforcement (combining safety with recovery) produced some separation-safety failures (3.46%) in targeted tests.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Hierarchical Temporal Runtime Assurance for Controlled Agentic AI Systems: Safety Shielding, Auditable Action Repair, and Bounded Recovery — Tymoteusz I. Miller — Machine Learning and Knowledge Extraction — doi:10.3390/make8090293
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper.