Simulating Action-Bound AI Safety: Pre-Commitment Monitoring, Strict Gating, and Authority Throttling in a Toy Benchmark

摘要

This paper presents a toy simulation benchmark and cross-language replication check for Action-Bound AI Safety. It evaluates pre-commitment monitoring, strict binary commitment gating, authority throttling, and cost-aware throttled gating in a simplified robotic-arm setting. The benchmark compares Python multi-seed robustness results with a C++17 replication. The results show that strict binary gating can reduce unsafe commitment but produces high hard false-positive burden, while authority throttling and cost-aware throttled gating preserve most of the safe-stop benefit while sharply reducing unnecessary hard stops. The results should be interpreted as a simulation-based consistency check under transparent toy assumptions, not as real-world robotic validation or proof of deployed-system safety.