Trust Is Not a Score: Runtime Assurance Contracts for High-Risk AI Agents.

arXiv:2609.39717v2 Announce Type: replace Abstract: Benchmarks, audits, and agent protocols describe performance, permissions, and repair, but not how observed evidence should change an agent's authority during a consequential task. We propose a Runtime Assura