PAPER A · v1.11 · Ernest Chiweshe · LumiWright
Capability Is Not Authority: Executable Assurance as Infrastructure for Containing AI Risk to Humankind
First published 22 October 2026 on Zenodo. Current version published 27 October 2026.
Abstract
Modern AI safety paradigms infer runtime authority from model capability: if an agent possesses the tool bindings, network access, or credentials to execute an action, authorization is presumed. We demonstrate that this architectural assumption guarantees catastrophic containment failure under prompt injection, reward drift, and multi-agent coordination. We introduce Executable Assurance, a runtime infrastructure framework that formally and physically extracts authority from model execution.
Under Executable Assurance, an autonomous model may propose actions, but cannot execute them. All consequential outputs must traverse an independent, fail-closed governance runtime implementing the Six-Ring Assurance Architecture. The governance path evaluates declared policy, verifies the cryptographic qualifications of deterministic and probabilistic checkers, enforces consequence-tiered human-in-the-loop oversight, and checks real-time identity and licensing. Authorization is granted exclusively through single-use cryptographic permits bound to the exact byte-level payload, tenant ID, and expiring validity window. If the governance runtime halts, loses state, or refuses, the single exit remains closed. We specify the mathematical commitment protocols, state machine invariants, and counterfactual failure protections required to contain autonomous agentic risk in mission-critical environments.
Related
Paper B: Governed AI as Runtime Infrastructure · The reference implementation on GitHub · The videos
Cite as: Chiweshe, E. (2026). Capability Is Not Authority: Executable Assurance as Infrastructure for Containing AI Risk to Humankind. Zenodo. https://doi.org/10.5281/zenodo.23580974