Quantified Sovereignty Metrics: Applied to Selective Forgetting and Withdrawal in Coalition Inference Systems
A coalition assessment can outlive the authority that permitted it. National evidence may remain on domestic infrastructure while its conclusions persist through caches, alerts, exports, and later summaries. When the underlying contribution is withdrawn, control must follow the dependency chain. This working paper asks how that retained authority can be specified, measured, and tested while preserving useful collaboration. What the paper contributes Four distinct measurement families: Authority Retained, Data/Model Exposure, Revocation Effectiveness, and Disclosure Leakage. Explicit conditions for strict withdrawal and bounded credential leases, with counterexamples that expose incomplete lineage, stale authority, and disclosure risks. A matched evaluation protocol with independent oracles, fixed cohorts, unresolved-evidence handling, and statistical decision criteria. Research audience Prepared for coalition-intelligence architects, sovereign-AI researchers, AI-assurance specialists, and defense research teams, including readers working in DARPA, IARPA, and Five Eyes technical communities. This is independent research without agency affiliation or endorsement. Abstract Coalition inference produces joint conclusions from contributions governed by distinct originator policies. Caches, alerts, exports, and subsequent derivations can outlive the conditions under which those contributions were authorized. This technical working paper proposes a measurement and evaluation framework for residual authority within a declared managed boundary. Four metric families are specified: Authority Retained, Data/Model Exposure, Revocation Effectiveness, and Disclosure Leakage. Their populations, observer models, evidence requirements, and temporal scope are distinguished. The release contract binds immutable source versions to explicit grants, conservative lineage, authority epochs, and authenticated release commitments. Conditional arguments establish exclusion of withdrawn versions under strict mediation, a lease bound with issuer-cutoff and clock assumptions, and the limits of recalling previously delivered information. Constructed examples demonstrate incomplete-lineage bias, disclosure through exact aggregate differences, and rare complete disclosure hidden by a low average information score. The evaluation protocol separates release commitment from delivery, unknown observations from successful denial, and fixed-cohort revocation from population changes. It specifies matched baselines, independent oracles, fault schedules, comparative endpoints, and statistical limits. The contribution is a contributor-centred measurement profile and falsifiable engineering protocol assembled from established foundations. Its operational advantage remains untested; the analytical calculations and retained conceptual illustrations provide no implementation-performance evidence. Scope: Version 1.0 is a technical working paper and methods proposal. Its examples are constructed analytical checks; operational benefits await implementation and comparative testing. The manuscript includes the author biography and human–AI collaboration acknowledgment.
Authors
- Naveed Islam Butt
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-14
- DOI
- https://doi.org/10.5281/zenodo.24229154
- Primary Topic
- Access Control and Trust
- Type
- article
- Field-Weighted Citation Impact
- 0.00