Sovereign Evaluation Leases: Revocable Authorization for Reward Models and Automated Judges

A national contribution can outlive its permission through reward models, synthetic preferences, cached judgments, and successor evaluators. This paper makes the control problem precise: what must be checked before a derived result becomes usable after withdrawal? ABSTRACT Automated evaluators can turn contributed preferences, rubrics, and judgments into training signals that persist across reward models, synthetic datasets, and successor judges. A contributor may later withdraw permission for further use while its earlier influence remains embedded in existing artifacts. This paper specifies Sovereign Evaluation Leases (SEL): time-bounded, contributor-scoped authorization checked against protected provenance and current policy at each governed promotion boundary. A promotion makes a candidate artifact authoritative or reusable within the controlled system. The architecture combines established usage-control, information-flow, provenance, and distributed-authorization mechanisms. It introduces no cryptographic primitive and makes no claim of neural forgetting. A conditional safety argument identifies the assumptions needed to deny dependent promotions after a withdrawal has committed to the policy registry. An executable authorization model compares SEL with an independently expressed, functionally matched conventional baseline and four diagnostic ablations. Across 100,000 synthetic requests, SEL and the strong baseline produce identical decisions: 68,388 accepted operations and zero unauthorized commits within the modeled boundary. Enumeration of a specified 16,384-case Boolean abstraction also finds no disagreement. Additional diagnostics check generated ancestry, event order, commit-time expiry, inherited deadlines, and fixed delivery commitments. Removing individual controls exposes failures, including a cache bypass invisible to update-only metrics. A lineage-omission stress test shows that policy correctness cannot compensate for hidden dependencies. These results establish consistency of the modeled authorization rules; they do not establish implementation security, model quality, privacy, or operational superiority. INCLUDED MATERIALS Version 1.1 contains the final 30-page technical working paper with nine figures, sixteen tables, eight equations, and nineteen references, plus a reproducibility package containing the editable manuscript, source code, data tables, figures, contract diagnostics, publication metadata, and integrity checksums. RESEARCH AUDIENCE AI researchers, systems engineers, coalition-intelligence architects, AI-assurance specialists, and defense research readers, including DARPA, IARPA, and Five Eyes technical communities. This is independent research without agency affiliation or endorsement. HUMAN–AI COLLABORATION ACKNOWLEDGMENT I 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.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-15
DOI
https://doi.org/10.5281/zenodo.24261145
Primary Topic
Access Control and Trust
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Sovereign Evaluation Leases: Revocable Authorization for Reward Models and Automated Judges

Naveed Islam Butt
Zenodo (CERN European Organization for Nuclear Research)
Access Control and Trust
article

Sovereign Evaluation Leases: Revocable Authorization for Reward Models and Automated Judges

Naveed Islam Butt
article en

Abstract

A national contribution can outlive its permission through reward models, synthetic preferences, cached judgments, and successor evaluators. This paper makes the control problem precise: what must be checked before a derived result becomes usable after withdrawal? ABSTRACT Automated evaluators can turn contributed preferences, rubrics, and judgments into training signals that persist across reward models, synthetic datasets, and successor judges. A contributor may later withdraw permission for further use while its earlier influence remains embedded in existing artifacts. This paper specifies Sovereign Evaluation Leases (SEL): time-bounded, contributor-scoped authorization checked against protected provenance and current policy at each governed promotion boundary. A promotion makes a candidate artifact authoritative or reusable within the controlled system. The architecture combines established usage-control, information-flow, provenance, and distributed-authorization mechanisms. It introduces no cryptographic primitive and makes no claim of neural forgetting. A conditional safety argument identifies the assumptions needed to deny dependent promotions after a withdrawal has committed to the policy registry. An executable authorization model compares SEL with an independently expressed, functionally matched conventional baseline and four diagnostic ablations. Across 100,000 synthetic requests, SEL and the strong baseline produce identical decisions: 68,388 accepted operations and zero unauthorized commits within the modeled boundary. Enumeration of a specified 16,384-case Boolean abstraction also finds no disagreement. Additional diagnostics check generated ancestry, event order, commit-time expiry, inherited deadlines, and fixed delivery commitments. Removing individual controls exposes failures, including a cache bypass invisible to update-only metrics. A lineage-omission stress test shows that policy correctness cannot compensate for hidden dependencies. These results establish consistency of the modeled authorization rules; they do not establish implementation security, model quality, privacy, or operational superiority. INCLUDED MATERIALS Version 1.1 contains the final 30-page technical working paper with nine figures, sixteen tables, eight equations, and nineteen references, plus a reproducibility package containing the editable manuscript, source code, data tables, figures, contract diagnostics, publication metadata, and integrity checksums. RESEARCH AUDIENCE AI researchers, systems engineers, coalition-intelligence architects, AI-assurance specialists, and defense research readers, including DARPA, IARPA, and Five Eyes technical communities. This is independent research without agency affiliation or endorsement. HUMAN–AI COLLABORATION ACKNOWLEDGMENT I 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.

Zenodo (CERN European Organization for Nuclear Research)
Access Control and Trust
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon