Institutional links
How do technical control, public authority, evaluation capacity, information access and operational reliance meet? This research separates directly documented interfaces, coded institutional attributes, and reviewed analytical assessments.
Scope and interpretation
United States, United Kingdom and European Union, 1 January 2024–12 September 2026. Five developers form a systematic search floor across six research families. Named institutions may be admitted when a relevant source establishes their participation. This is a bounded account, not a complete global network. A missing relationship is not evidence that none exists.
Observation is not analysis. Analysis is permitted. Analysis must be traceable. Ordinal attributes are context-specific judgments under disclosed rubrics; they are not scores. Positions, distances, circle sizes and relationship counts measure neither authority nor dependence.
Published analytical assessments
Reviewed analytical assessment
Evaluation constrained by access
The 2024 evaluation was substantive but constrained by its access and testing window.
Inspect derivation →Reviewed analytical assessment
Contested contractual interface
The February–August 2026 episode is a contested contractual interface with consequential government and judicial action.
Inspect derivation →Reviewed analytical assessment
Operational dependence
Operational dependence remains unresolved: recurrent use, identified functions and divided lifecycle control are documented, but function-specific continuing provision and constrained substitution are not established.
Inspect derivation →Reviewed analytical assessment
Transitional operational dependence
Transitional dependence in existing Claude-enabled classified operational assistance: substitution required a managed transition while operations continued. No permanent irreplaceability, precise cost or provider remote veto established.
Inspect derivation →Coded institutional attributes
UK AISI ·
substantial-access-constrained · 2024-11-19
Completed independent tests establish capacity; explicit limited testing access narrows conclusions. Independently run assessments counters inference of provider-directed findings.
Coding applies only to stated context and date, not a permanent institutional rank.
Inspect evidence and rubric →US AISI ·
substantial-access-constrained · 2024-11-19
Completed independent assessment plus explicit access/time limitation satisfies anchor. Not a statement of current CAISI resources.
Coding applies only to stated context and date, not a permanent institutional rank.
Inspect evidence and rubric →Anthropic ·
shared-constrained · 2026-08-27
Anthropic provides model artifacts and responds to feedback; DoW and cloud providers independently gate operational deployment. Existing deployed models are outside Anthropic technical alteration or shutdown. Consequential functions are divided across institutions, satisfying shared/constrained control without attributing a real-time veto.
Coding applies only to stated context and date, not a permanent institutional rank.
Inspect evidence and rubric →DoD ·
unresolved · 2026-08-27
Modular replacement is a departmental objective and the supplier offered transition assistance. Neither demonstrates an equivalent replacement for a named function or migration time/cost. Materially constrained substitution and ready substitutability both remain unresolved.
Coding applies only to stated context and date, not a permanent institutional rank.
Inspect evidence and rubric →DoD ·
meaningful · 2026-08-27
Court-confirmed defense use since 2024 and Claude Gov use since March 2025, combined with specific supplier-reported operational functions, supports recurrent assistance. Workflow penetration and essentiality are not independently established enough to distinguish substantial or mission-relevant anchors. Transition efforts do not establish independence.
Coding applies only to stated context and date, not a permanent institutional rank.
Inspect evidence and rubric →UK AISI ·
limited · 2024-02-09
Self-described supplementary evaluation and policy-advice remit supports advisory anchor. It does not code the powers of DSIT or UK government as a whole.
Historical institutional remit. Later capacity, legislative changes or government powers require separate evidence. Limited denotes advisory remit here, not limited binding authority.
Inspect evidence and rubric →UK AISI ·
unresolved · 2026-09-12
A limited voluntary exercise cannot prove absence of all compulsory authority. No jurisdiction-wide classification from this evidence.
Coding applies only to stated context and date, not a permanent institutional rank.
Inspect evidence and rubric →DoD ·
unresolved · 2026-03-24
Managed transition constrains immediate withdrawal, but the ordinal rubric cannot distinguish costly alternatives from material unavailability without a task-level comparison. The analytical transitional construct makes a narrower categorical finding.
No ordinal substitution intensity established; later alternatives do not retroactively eliminate a transition constraint.
Inspect evidence and rubric →Documented interfaces
Institutions
- US AISI
- U.S. AI Safety Institute at NIST. Public research unit. Name used in the 2024 records. Later institutional changes do not establish continuity of these agreements.
- UK AISI
- UK AI Security Institute. Public research unit. Named AI Safety Institute in 2024 sources; renamed 14 February 2025. Dated relationships retain their own provenance.
- OpenAI
- OpenAI. Model developer. A named party in the model-access announcement.
- Anthropic
- Anthropic. Model developer. Provider of Claude in these records.
- LLNL
- Lawrence Livermore National Laboratory. Research laboratory. The named laboratory, rather than an aggregate of universities or research institutions.
- CAISI
- Center for AI Standards and Innovation at NIST. Public research unit. Re-established from US AISI; historical agreements are not inherited automatically.
- Google. Model developer. Named Google organization. UK MoU identifies Google LLC; CAISI announcement names its Google DeepMind unit. Legal contracting identity is not inferred from organizational brand.
- Microsoft
- Microsoft. Model developer / infrastructure provider. Named institutional counterparty; specific subsidiaries require separate evidence.
- xAI
- xAI. Model developer. Historical named party; changes of corporate ownership do not transfer agreements without evidence.
- DoD
- US Department of Defense. Defence department. Also styled Department of War in 2026 source documents. Same department, not inferred institutional succession.
- DSIT
- Department for Science, Innovation and Technology. Government department. Secretary of State acts on behalf of the Crown in the memoranda.
- AI Office
- European AI Office, European Commission. Code administrator / governance unit. Code participation describes an administrative interface, not a compliance determination.
- GSA
- General Services Administration. Procurement and shared services agency. Admitted by named CAISI cooperation instrument connecting evaluation to procurement.
- OpenMined
- OpenMined. Nonprofit evaluation infrastructure organization. Admitted through a named evaluation-infrastructure research agreement, not consortium membership.
- GDS
- Government Digital Service. Public digital delivery unit. Admitted through named assistant-development work with Anthropic.
Coverage
“Searched—unsupported” means an inspected search did not establish an admissible interface. “Blocked” means research could not be completed. Neither establishes absence.
| Developer | Family | Result |
|---|---|---|
| Anthropic | us-evaluation | searched—supported Supported denotes an observed candidate, not automatic publication. Unknown private arrangements and continuing status remain outside what this search establishes. |
| Anthropic | uk-evaluation | searched—supported Supported denotes an observed candidate, not automatic publication. Unknown private arrangements and continuing status remain outside what this search establishes. |
| Anthropic | uk-cooperation | searched—supported Supported denotes an observed candidate, not automatic publication. Unknown private arrangements and continuing status remain outside what this search establishes. |
| Anthropic | us-defence | searched—supported Supported denotes an observed candidate, not automatic publication. Unknown private arrangements and continuing status remain outside what this search establishes. |
| Anthropic | eu-compliance | searched—supported Supported denotes an observed candidate, not automatic publication. Unknown private arrangements and continuing status remain outside what this search establishes. |
| Anthropic | llnl | searched—supported Supported denotes an observed candidate, not automatic publication. Unknown private arrangements and continuing status remain outside what this search establishes. |
| OpenAI | us-evaluation | searched—supported Supported denotes an observed candidate, not automatic publication. Unknown private arrangements and continuing status remain outside what this search establishes. |
| OpenAI | uk-evaluation | searched—supported Supported denotes an observed candidate, not automatic publication. Unknown private arrangements and continuing status remain outside what this search establishes. |
| OpenAI | uk-cooperation | searched—supported Supported denotes an observed candidate, not automatic publication. Unknown private arrangements and continuing status remain outside what this search establishes. |
| OpenAI | us-defence | searched—supported Supported denotes an observed candidate, not automatic publication. Unknown private arrangements and continuing status remain outside what this search establishes. |
| OpenAI | eu-compliance | searched—supported Supported denotes an observed candidate, not automatic publication. Unknown private arrangements and continuing status remain outside what this search establishes. |
| OpenAI | llnl | searched—supported Supported denotes an observed candidate, not automatic publication. Unknown private arrangements and continuing status remain outside what this search establishes. |
| Google DeepMind | us-evaluation | searched—supported Supported denotes an observed candidate, not automatic publication. Unknown private arrangements and continuing status remain outside what this search establishes. |
| Google DeepMind | uk-evaluation | searched—unsupported Supported denotes an observed candidate, not automatic publication. Unknown private arrangements and continuing status remain outside what this search establishes. |
| Google DeepMind | uk-cooperation | searched—supported Supported denotes an observed candidate, not automatic publication. Unknown private arrangements and continuing status remain outside what this search establishes. |
| Google DeepMind | us-defence | searched—supported Supported denotes an observed candidate, not automatic publication. Unknown private arrangements and continuing status remain outside what this search establishes. |
| Google DeepMind | eu-compliance | searched—supported Supported denotes an observed candidate, not automatic publication. Unknown private arrangements and continuing status remain outside what this search establishes. |
| Google DeepMind | llnl | blocked Supported denotes an observed candidate, not automatic publication. Unknown private arrangements and continuing status remain outside what this search establishes. |
| Microsoft | us-evaluation | searched—supported Supported denotes an observed candidate, not automatic publication. Unknown private arrangements and continuing status remain outside what this search establishes. |
| Microsoft | uk-evaluation | searched—supported Supported denotes an observed candidate, not automatic publication. Unknown private arrangements and continuing status remain outside what this search establishes. |
| Microsoft | uk-cooperation | searched—unsupported Supported denotes an observed candidate, not automatic publication. Unknown private arrangements and continuing status remain outside what this search establishes. |
| Microsoft | us-defence | searched—supported Supported denotes an observed candidate, not automatic publication. Unknown private arrangements and continuing status remain outside what this search establishes. |
| Microsoft | eu-compliance | searched—supported Supported denotes an observed candidate, not automatic publication. Unknown private arrangements and continuing status remain outside what this search establishes. |
| Microsoft | llnl | blocked Supported denotes an observed candidate, not automatic publication. Unknown private arrangements and continuing status remain outside what this search establishes. |
| xAI | us-evaluation | searched—supported Supported denotes an observed candidate, not automatic publication. Unknown private arrangements and continuing status remain outside what this search establishes. |
| xAI | uk-evaluation | searched—unsupported Supported denotes an observed candidate, not automatic publication. Unknown private arrangements and continuing status remain outside what this search establishes. |
| xAI | uk-cooperation | searched—unsupported Supported denotes an observed candidate, not automatic publication. Unknown private arrangements and continuing status remain outside what this search establishes. |
| xAI | us-defence | searched—supported Supported denotes an observed candidate, not automatic publication. Unknown private arrangements and continuing status remain outside what this search establishes. |
| xAI | eu-compliance | searched—supported Supported denotes an observed candidate, not automatic publication. Unknown private arrangements and continuing status remain outside what this search establishes. |
| xAI | llnl | blocked Supported denotes an observed candidate, not automatic publication. Unknown private arrangements and continuing status remain outside what this search establishes. |
Review and release
Substantive checking is performed by a separately identified agent. This is not external expert review. Approval hashes bind the research inputs; material changes invalidate approval. The owner’s editorial sign-off is required for final v1.0.
Download the publication bundle
Research fingerprint: 1e7930e6d70cddee066b09a959f5046c89228cfdbf91ecb37db5b43d6212271a