Authority Contraction and Refusal as Safety Invariants in Autonomous Systems

David Forbes · 2026 · 18 citationsRead the paper

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, autonomous agents, distributed systems, AI governance, runtime governance, operational resilience, and system resilience. A system may remain technically capable of execution even 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 evaluating recognized authority before consequential action proceeds. SAB separates technical capability from legitimate authority and requires applicable authority state, scope, conditions, permissions, and 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, or loss of required evidence may reduce the set of actions a system is legitimately permitted to perform. This connects authority management directly to runtime governance, policy enforcement, runtime authorization, runtime assurance, execution control, fault tolerance, graceful degradation, and operational resilience. Within this architecture, refusal is a correct system behavior when execution authority cannot be established. An autonomous agent or distributed system that refuses, holds, constrains, or escalates an action because recognized authority is absent or uncertain is not necessarily failing; it may be enforcing the required safety boundary. The architecture is relevant to autonomous systems, autonomous agents, AI agents, agentic AI, distributed systems, machine autonomy, AI governance, runtime governance, system resilience, operational resilience, fault tolerance, degraded operation, graceful degradation, policy enforcement, runtime authorization, runtime assurance, decision authority, execution authority, authorized execution, authority contraction, safe refusal, and execution control. The central principle is: as operating conditions degrade, legitimate authority may contract before technical capability disappears; when authority cannot be established, refusal is a valid and necessary execution outcome. BLOCK VECTOR Research ArchitectureThis Research Map is the navigation surface for the BLOCK VECTOR publication architecture. The interactive version is maintained at: 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 back to this map, while this map provides direct routes to the current publication set.

1 idea Seedlabs derived from this research