Work

Sovereign Agent Leases: Measuring and Enforcing Withdrawal of Delegated AI Authority in Coalition Analytical Systems

Year: 2026
Type: article
Abstract: PUBLICATION METADATA REVISION — VERSION 1.2Prepared 15 November 2026. This edition adds a standardized author record, the approved human–AI collaboration acknowledgment, and a full 50-term publication keyword index. The PDF contains 30 source-edition pages plus two supplement pages (32 pages in total). The source research content, evidence cutoff, and limitations are preserved. The PDF and companion JSON contain the same verified keyword set. SOURCE-EDITION RESEARCH DESCRIPTIONRevoking an AI agent’s credentials does not necessarily stop its consequences. Child agents, queued work, cached decisions, and downstream services may continue to act after the originating contributor believes permission has ended. In coalition analytical systems, that gap affects control over what may still be computed, retained, delegated, or released. Sovereign Agent Leases turns this problem into an inspectable engineering contract: whose authority governs an operation, which contributor dependencies must survive delegation, where withdrawal must be enforced, and what evidence can support a claim that the affected work has stopped. Written for senior intelligence leaders, system architects, AI engineers, developers, and assurance teams, this technical working paper develops five connected elements: • Contributor-specific leases with explicit scope, immutable expiration, selected delegation lineage, and root-wide descendant limits.• Enforcement at the irreversible effect commit, addressing the gap between accepting a request and executing it later.• Strict and bounded withdrawal profiles, with conditional stopping limits tied to credential lifetime, status freshness, clock uncertainty, and final commit delay.• Separate measurements for mediation coverage, lineage capture, unauthorized effects, state-access containment, closure delay, retained exposure, and useful availability.• Adversarial failure cases and matched comparisons against competent existing authorization systems, allowing an equivalence or negative result. The release contains the reviewed 30-page paper, seven technical figures, twenty primary references, and a reproducibility package with Python source, calculation data, synthetic traces, a finite-state exploration, and twenty-six executable scenario checks. The established human–AI collaboration acknowledgment is retained in the manuscript. The work addresses coalition intelligence, national-security AI, Zero Trust architecture, secure agentic workflows, and accountable multi-agent autonomy. Its intended readership includes DARPA, IARPA, NSA, Five Eyes research communities, and engineers responsible for deploying and evaluating delegated AI systems. These audience references do not imply sponsorship or endorsement. Evidence status: the numerical figures are analytical calculations or synthetic replay, and the executable model is an abstraction. The paper does not claim validated operational performance, a new cryptographic primitive, or demonstrated superiority over a strong baseline. Its contribution is a coalition-focused specification and falsifiable evaluation method that readers can inspect, challenge, and implement. AUTHOR AND PUBLICATION RECORDCreator: Naveed Islam ButtORCID: 0009-0000-2837-3149Category: Public working paperEdition: Version 1.2, prepared 15 November 2026Source edition: Version 1.1; published November 14, 2026 HUMAN–AI COLLABORATION ACKNOWLEDGMENTI proudly acknowledge the role of human–AI collaboration in developing this work. Working with ChatGPT (OpenAI) has helped me expand my strategic thinking, examine competing explanations, and connect my proposals to relevant scientific and technical research. My experience, objectives, and judgment guide the work, and I retain responsibility for its claims and conclusions. I regard transparent, rigorous collaboration with AI as a constructive method for advancing ideas that may benefit humanity. SOURCE-EDITION SUPPLEMENTARY MATERIALSRetained reproducibility packages and animations are source-edition materials. The revised PDF and companion publication-metadata JSON provide the standardized author record, acknowledgment, and keyword index for this edition.
Language: English
Open Access status: green
OpenAlex ID: W7233596176
FWCI:
Cites:
Cited by:
Related to: