Authority Contraction and Refusal as Safety Invariants in Autonomous Systems
Year: 2026
Type: article
Abstract: Autonomous systems, autonomous agents, AI agents, and distributed systems increasingly operate under changing conditions of coordination, uncertainty, degradation, partial observability, and failure. As machine autonomy increases, system resilience and fault tolerance require more than maintaining technical capability: a system must also determine whether it retains the decision authority and execution authority required to continue acting. This work examines authority contraction and refusal as safety invariants for autonomous systems, agentic AI, distributed systems, and other forms of machine autonomy. A system may remain technically capable of execution after the conditions supporting legitimate authority have changed. Continued capability therefore does not necessarily imply continued authorization or authorized execution. The paper develops the Stable Authority Boundary (SAB) as a pre-execution control boundary for determining whether recognized authority remains sufficient before consequential action proceeds. SAB separates technical capability from legitimate authority and requires applicable authority state, scope, conditions, permissions, and supporting evidence to remain valid at execution time. Under degraded conditions, authority may contract rather than remain static. Loss of communication, stale state, changing mission conditions, degraded coordination, partial system failure, uncertainty, or loss of required evidence may reduce the set of actions a system is legitimately permitted to perform. Authority contraction therefore connects directly to runtime governance, policy enforcement, runtime authorization, runtime assurance, execution control, fault tolerance, graceful degradation, operational resilience, and system resilience. Within this architecture, refusal is a correct execution outcome when sufficient authority cannot be established. An autonomous agent or distributed system that refuses, holds, constrains, delays, or escalates an action because recognized authority is absent, stale, ambiguous, or insufficient is not necessarily malfunctioning. It may instead be correctly enforcing the governing safety boundary. This distinction is particularly important in systems designed to continue operating through communications loss, degraded coordination, partial failure, or incomplete observability. Conventional resilience mechanisms may preserve availability or technical capability while the authority supporting particular actions has already narrowed. A resilient autonomous system must therefore be capable not only of continuing operation, but also of reducing its permitted action space as authority degrades. The central principle is: As operating conditions degrade, legitimate authority may contract before technical capability disappears. When sufficient execution authority cannot be established, refusal is a valid and necessary system behavior. Discovery language: This work applies to autonomous systems, autonomous agents, AI agents, agentic AI, distributed systems, machine autonomy, AI governance, runtime governance, decision authority, execution authority, runtime authorization, runtime assurance, authorized execution, policy enforcement, fault tolerance, graceful degradation, degraded operation, operational resilience, authority contraction, safe refusal, and execution control. BLOCK VECTOR Research ArchitectureThis publication forms part of the BLOCK VECTOR research architecture. The interactive Research Map provides navigation across the related work: https://blockvectortech.com/research-map.html Archival Research Map — Version 1.0: https://doi.org/10.5281/zenodo.22102620 Individual publications within the architecture link to related works through DOI relationships, while the Research Map provides a consolidated route through the current publication set.
Author: Forbes, David
Language: English
Topic: Access Control and Trust
Subfield: Sociology and Political Science
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7124428436