Standards Body · Foundation paper, public edition · Released July 17, 2026
Canonical record: https://standardsbody.ai/library/foundation-paper/global-interoperability/
Standards Body is an independent research and institutional-design project. It is not currently a regulator, accreditation body, certification body, or governmental authority. This document is research; it is not an adopted standard.
Series: Foundations for Frontier AI Evaluation Infrastructure
Version: 1.0
Status: Canonical working white paper
Project: Standards Body
Primary domain: standardsbody.ai
Core line: Foundations for Frontier AI
Research basis reviewed through: July 16, 2026
Document owner: Standards Body
Review cycle: Annual review, with event-triggered revision after material changes in international standards, evaluation networks, treaty systems, accreditation arrangements, model capabilities, or cross-border deployment practice
This paper defines the Standards Body position on global interoperability for frontier artificial intelligence evaluation, assurance, standards, and governance.
It is intended to serve as:
This paper does not propose a single global regulator.
It does not claim that all jurisdictions should adopt identical laws, thresholds, risk tolerances, or institutional structures.
It does not authorize Standards Body to recognize governments, accredit evaluators, certify systems, or negotiate treaties.
It defines the infrastructure required for distinct institutions to exchange and interpret credible evidence without erasing legitimate differences.
Frontier AI systems cross borders more easily than the institutions responsible for evaluating and governing them.
A model may be:
The evidence surrounding that system may also be distributed.
One institution may hold:
Another may hold:
Another may possess:
Another may assess:
Another may issue:
Another may possess:
Without interoperability, these institutions can produce a large quantity of evidence that does not combine into shared understanding.
The same term may have different meanings.
The same score may be produced under different conditions.
The same model name may refer to different configurations.
The same certification language may imply different levels of assurance.
A threshold in one framework may not correspond to a threshold in another.
An evaluator accepted in one jurisdiction may be unrecognized in another.
An incident category may not transfer across legal systems.
A confidential finding may be too sensitive for public disclosure but too important to remain isolated.
A requirement may be technically equivalent to another requirement while appearing different because the documentation and institutional form are different.
These failures create practical consequences:
Global interoperability is the discipline of making evidence, standards, protocols, qualifications, and institutional decisions usable across boundaries.
It does not require one universal test.
It does not require one legal system.
It does not require one threshold.
It does not require one definition of acceptable risk.
It requires enough shared structure that different institutions can answer:
Standards Body adopts the following core position:
Global frontier AI governance should pursue interoperable evidence, protocols, terminology, and assurance systems while preserving legitimate jurisdictional, cultural, institutional, and policy differences. The objective is shared understanding and portable trust, not forced global uniformity.
A mature interoperability system should operate across at least ten layers.
Institutions use shared definitions or explicit mappings among terms such as:
Evidence refers to a verifiable model, system, configuration, protocol, evaluator, and date.
Evaluation procedures can be compared, mapped, reused, or reproduced across institutions.
Scores, uncertainty, baselines, task populations, and thresholds can be interpreted without false equivalence.
Reports include a common minimum set of machine-readable and human-readable information.
Evaluator qualifications, scopes, review levels, accreditation, certification, and result status can be understood across borders.
Organizations can classify, share, escalate, and learn from incidents using compatible structures.
Institutions can map technical evidence into different legal regimes without pretending the regimes are identical.
Sensitive evidence can be exchanged through trusted, tiered, and accountable channels.
Governments, laboratories, standards bodies, auditors, developers, researchers, and open communities can cooperate without surrendering their respective authority.
Interoperability should be built through modular agreements.
A country may recognize another country's evaluator competence while retaining its own deployment decision.
A purchaser may accept an evaluation report while requiring additional local testing.
A standards body may map two protocols without declaring them equivalent.
A regulator may recognize an international standard as evidence without making it the only compliance route.
A confidential incident exchange may operate among trusted institutions while a public summary remains available.
A multilingual protocol may preserve a shared construct while allowing locally valid tasks.
The central institutional distinction is between:
These are not the same.
A result can be scientifically informative without being legally determinative.
An evaluator can be competent without being authorized by every jurisdiction.
A certification can be valid under one scheme without satisfying every local requirement.
Interoperability should therefore support graded outcomes:
The global landscape already contains important building blocks.
NIST has published a plan for global engagement on AI standards centered on cooperation, consensus standards, and information sharing.[^nist-global] The OECD's Hiroshima AI Process Reporting Framework provides a common voluntary reporting structure for advanced AI governance and risk-management practices.[^haip] The Council of Europe Framework Convention on Artificial Intelligence establishes a treaty-level framework focused on human rights, democracy, and the rule of law.[^coe-convention] The United Nations Global Digital Compact establishes a broad global framework for digital cooperation and AI governance.[^un-gdc] International accreditation arrangements such as the ILAC Mutual Recognition Arrangement demonstrate how testing and inspection results can gain cross-border acceptance through peer-evaluated accreditation systems.[^ilac-mra] The International Network for Advanced AI Measurement, Evaluation and Science has begun articulating areas of consensus and open questions for advanced AI evaluation across national institutes.[^aisi-network]
These initiatives are not interchangeable.
They operate at different levels:
Global interoperability should connect them without collapsing them.
The eighth foundation of Standards Body is therefore a shared evidentiary language for plural institutions.
Frontier AI evidence should be understandable and usable across institutional and jurisdictional boundaries without requiring every institution to adopt the same law, protocol, threshold, or policy judgment.
Interoperability should preserve legitimate diversity while making differences explicit.
The first object of international alignment should be the structure and meaning of evidence, not immediate agreement on every policy outcome.
No evaluation, certification, or incident record is globally useful unless the system, configuration, protocol, evaluator, and time are identifiable.
Recognition should be granular. Institutions should be able to recognize competence, process, evidence, or legal effect separately.
Global interoperability is incomplete when only a small number of wealthy countries, companies, or laboratories can produce or interpret the required evidence.
International evidence sharing should protect sensitive models, vulnerabilities, incidents, personal data, and dual-use information through proportionate access controls and accountable disclosure.
Mappings, equivalence decisions, recognition arrangements, and shared schemas should change as standards, systems, and institutional capabilities change.
This paper covers interoperability concerning:
This paper does not establish:
Interoperability enables systems or institutions to exchange and use information effectively.
Harmonization reduces differences among requirements, methods, or standards.
Interoperability can exist without full harmonization.
Two approaches may interoperate without being equivalent.
Equivalence is a stronger claim that outcomes or requirements are sufficiently comparable for a stated purpose.
Mutual recognition is an institutional agreement to accept specified results, qualifications, certificates, or processes.
It normally depends on interoperability but adds legal or organizational commitment.
Uniformity requires sameness.
Interoperability requires understandable and manageable difference.
Technical comparability does not automatically create political agreement.
A shared cyber capability result may support different national responses.
The ability of distinct systems, organizations, protocols, or jurisdictions to exchange, interpret, and use information or evidence effectively.
Shared or mapped meaning among terms, classifications, and data elements.
Compatibility in data structure, format, encoding, and transmission.
Compatibility among processes, workflows, responsibilities, and decision steps.
Ability of technical systems, tools, APIs, schemas, and environments to work together.
Ability to interpret and compare measurements produced by different methods, instruments, task sets, or institutions.
Ability of organizations with different mandates and governance structures to coordinate and rely on one another's work.
Ability to map or coordinate requirements and evidence across legal systems while preserving each system's authority.
An arrangement through which parties accept specified results, qualifications, certificates, or decisions issued under another recognized system.
Acceptance by one party of another party's evidence or status without reciprocal obligation.
A determination that different requirements, methods, or systems achieve sufficiently comparable outcomes for a defined purpose.
The degree to which results can be meaningfully compared.
The degree to which systems or requirements can operate together without unacceptable conflict.
A structured mapping between terms, requirements, controls, classifications, or standards.
A mapping showing relationships among multiple vocabularies or classification systems.
A common conceptual structure that supports implementation by different organizations.
A baseline set of shared requirements or metadata accepted across participating systems.
Adaptation of a protocol, standard, or system to local language, law, culture, infrastructure, or professional practice.
The degree to which a translated evaluation preserves the intended construct and interpretation.
The specific activities, methods, domains, systems, and conditions covered by a recognition arrangement.
An institution, credential, cryptographic root, registry, or assurance mechanism relied upon to establish identity or legitimacy.
A maintained record of protocols, systems, evaluators, certificates, incidents, mappings, or recognition status.
A structured set of artifacts supporting a claim or decision.
A declaration of the standards, schemas, protocols, identifiers, and recognition conditions supported by an organization or system.
A documented determination to accept, conditionally accept, partially accept, suspend, or reject external evidence or status.
The part of a shared system that permits jurisdictional, linguistic, cultural, or domain-specific adaptation.
The shared components that should remain stable across implementations.
An analysis that connects results across protocols, versions, languages, or task forms.
An unnecessary or disproportionate difference in requirements or procedures that impedes cross-border use or trade.
Movement of activity toward jurisdictions or institutional arrangements with weaker or more favorable requirements.
The principle that data is subject to the laws, governance, and control arrangements associated with relevant jurisdictions or communities.
Evaluation in which data, tasks, models, or evidence remain distributed while coordinated methods produce shared results.
A governed network through which authorized institutions share nonpublic information under defined security, use, and accountability rules.
The accumulated cost created by incompatible terminology, formats, identifiers, protocols, and institutional arrangements.
Deterioration in the validity of a recognition arrangement after methods, standards, institutions, or systems change.
Model supply chains, hosting, users, data, capital, and applications cross jurisdictions.
No single institution sees the entire system.
Repeated assessments can consume:
Some duplication is valuable for independent replication.
Unnecessary duplication is not.
Institutions may use different words for similar concepts or the same word for different concepts.
A score of 70 under one protocol may not be comparable with 70 under another.
Organizations may need separate evidence packages for each market.
Misuse, security failures, and model behavior can affect users in multiple countries.
Some regions have advanced laboratories and institutes.
Others do not.
Governments and organizations may distrust foreign evaluators, proprietary evidence, or politically influenced institutions.
No single developer can control all downstream versions or deployments.
Shared standards and recognition can reduce unnecessary repeated testing.
Interoperability can also:
It requires governance.
The preferred model is:
Shared evidence structure, explicit local interpretation, accountable local or international decision authority.
Local differences should be documented rather than silently embedded in test content or scoring.
Can institutions understand each other's terms?
Can institutions verify the system, protocol, evaluator, and evidence source?
Can reports and artifacts be exchanged?
Can methods be reproduced or mapped?
Can results be compared with appropriate uncertainty?
Can evaluator competence and result status be recognized?
Can sensitive information move safely?
Can evidence be used under different legal systems?
Can organizations coordinate review, escalation, and appeals?
Can all participating regions meaningfully implement and use the system?
Failure at one layer can undermine the others.
Begin with compatible evidence and explicit mappings.
A common JSON field is not useful if institutions interpret it differently.
Confirm what was evaluated before comparing results.
Equivalence should always state the purpose for which it is accepted.
Share the smallest stable core necessary for cooperation.
Allow local adaptation with a clear record.
State when results cannot be compared.
Avoid dependence on one institution, registry, cloud, country, or accreditation body.
Share enough for coordination without unnecessary disclosure.
Interoperability should include funding, training, infrastructure, and translation.
Prefer documented interfaces and portable evidence.
Support multiple standards through crosswalks where appropriate.
Recognition applies to identified versions and scopes.
Mappings and recognition decisions should be appealable and reviewable.
Machine-readable systems should have clear human explanations.
Crosswalks, schemas, and recognition should update as the field changes.
Terms such as "frontier model," "systemic risk," "critical capability," and "independent evaluator" vary across organizations.
A shared vocabulary should include:
Mappings may indicate:
Definitions should be:
Legal definitions may need precise jurisdictional meaning.
Do not overwrite them with technical vocabulary.
Translate concepts, not only words.
Maintain disputed terms and unresolved interpretations.
TERMINOLOGY.md should become the canonical project vocabulary, with Foundation 8 defining mapping requirements.
A model name is often insufficient.
Systems can differ through:
Where feasible, use:
Closed APIs may require developer attestation and evaluator verification.
Record:
Identity should include all material components.
A material identity change should trigger re-evaluation or explicit inheritance analysis.
A future system could assign persistent, non-proprietary identifiers for evaluated AI artifacts and configurations.
Identity systems should not disclose sensitive information unnecessarily.
Each protocol should have:
A profile should describe:
Shared evaluation tools can support portability.
The UK AI Security Institute's Inspect framework demonstrates a modular approach to evaluation tasks, agents, tools, scorers, and model interfaces.[^inspect]
A reference implementation can clarify the protocol.
It should not become the only permitted implementation unless required.
Compare:
Same protocol and materially equivalent execution.
Different task forms with validated comparability.
Different methods measuring substantially the same construct.
Related but not directly comparable.
No defensible comparison.
Use shared systems, anchor tasks, human baselines, and statistical analysis.
Allow forks with explicit lineage and compatibility statements.
Retired protocols should remain discoverable with status.
A score requires context.
Use common units where valid.
Examples:
Some domains require specialized measures.
Map rather than force conversion.
Thresholds may be:
Human comparisons require common definitions of:
Interoperability should carry uncertainty, not only point estimates.
Use only when assumptions are defensible.
Standards Body opposes a single global frontier safety score.
A future registry can document metrics, protocols, mappings, and limitations.
Metadata makes evidence discoverable and interpretable.
What was evaluated?
How was it evaluated?
Who performed the work?
What artifacts support the result?
What was found?
What review or recognition applies?
Is the evidence current?
Where and how can it be used?
Use open, documented schemas.
Every machine record should have a plain-language explanation.
Record:
Minimize personal data.
Allow domain and jurisdiction extensions.
Define:
An evaluator's competence should be tied to:
Accreditation can support recognition when accreditation bodies operate under shared requirements and peer evaluation.
The ILAC Mutual Recognition Arrangement supports cross-border acceptance of results from accredited testing, calibration, inspection, proficiency-testing, and reference-material activities.[^ilac-mra]
Its general principle, accredited once and accepted across participating systems, is instructive.
Frontier AI will require narrower and more dynamic scopes.
Recognition may require:
Publish:
Institutions recognizing evaluators should themselves be reviewed.
National institutes may recognize each other's technical evidence while retaining independent conclusions.
Institutional reputation is not a substitute for scope and evidence.
A certificate may travel only if:
ISO/IEC 42001 provides requirements for AI management systems, and ISO/IEC 42006 provides requirements for bodies auditing and certifying those systems.[^iso-42001][^iso-42006]
These standards can support international consistency in organizational assurance.
They do not establish universal system safety.
More technical evidence may be needed for:
A recognition statement should identify:
Global marks risk oversimplification.
Prefer verifiable registry records.
Material suspension should be communicated across recognition networks.
Translation should preserve legal and technical meaning.
Incidents can reveal:
Different institutions may use different scales.
Provide crosswalks.
High-level incident and lesson.
Detailed technical and operational evidence.
Highly sensitive vulnerabilities, personal data, or national-security information.
Define:
Include near misses to improve prevention.
Use common identifiers and linkage.
Incidents should update:
A future network could support trusted exchange among qualified national institutes and evaluators.
Define:
Keep evidence local while sharing:
May support joint evaluation without full asset disclosure.
Limitations remain.
Account for:
A secure exchange network needs:
Interoperability should not compel unsafe or unlawful disclosure.
Crosswalks reduce duplication and clarify overlap.
Every mapping should include:
Similar language does not prove equivalent requirements.
Evidence under one standard may support another requirement.
It should not be reused beyond its valid scope.
Use:
Update after either source changes.
Require jurisdiction-specific legal expertise.
Legal systems may prioritize:
Technical evidence can be portable even when legal consequences differ.
The European Union AI Act creates a risk-based legal structure with differentiated obligations for AI systems and general-purpose AI models.[^eu-ai-act]
Technical evidence may support compliance, but local legal interpretation remains necessary.
The Council of Europe Framework Convention provides a treaty-level framework focused on human rights, democracy, and the rule of law and was opened for signature in September 2024.[^coe-convention]
It illustrates a rights-centered international layer distinct from technical evaluation standards.
The Global Digital Compact provides a global framework for digital cooperation and AI governance, including capacity building and international cooperation.[^un-gdc]
Law may:
Authorities can cooperate through:
Interoperability should not obscure who has legal authority.
A system may face incompatible requirements.
Use:
Private standards should not acquire legal effect without accountable recognition.
Translation can change:
Use:
Report:
Some tasks should be localized rather than literally translated.
Require bridge studies.
Local professional and cultural expertise is essential.
English-centered protocols can misrepresent global capability and risk.
Use qualified legal translators and jurisdictional review.
Maintain canonical codes with local-language descriptions.
Translation work should receive professional credit and funding.
Open-weight models are copied, modified, quantized, merged, and deployed across borders.
Record:
Open models can support independent reproduction.
No single organization may control downstream deployment.
Communities can maintain:
Community evaluation should be recognized when methods and evidence are credible.
Open access changes safeguard and threat assumptions.
Interoperability schemas should be implementable without large compliance teams.
License terms affect use but should not be confused with evaluation evidence.
Open tooling and schemas reduce institutional dependence.
A standard that only a few countries can implement is not globally interoperable in practice.
International systems should support meaningful participation, not only consultation.
The United Nations has emphasized capacity building, equitable access, and participation by developing countries in global AI governance and safe, secure, and trustworthy AI systems.[^un-resolution][^un-gdc]
Regional centers can adapt methods and build local trust.
Capacity building should support autonomous local competence rather than permanent reliance.
Local institutions should participate in priority setting.
Programs should support careers and institutions in participants' regions.
Track who can:
National institutes can:
The network has articulated shared areas of consensus and open questions concerning advanced AI evaluation, including the need for common understanding across borders.[^aisi-network]
Bilateral agreements can support:
The OECD supports:
The UN provides broad global participation and development-oriented coordination.
The Council of Europe provides a human-rights, democracy, and rule-of-law treaty framework.
ISO and IEC provide international consensus standards and conformity-assessment infrastructure.
ILAC and related international accreditation systems demonstrate peer-evaluated recognition mechanisms.
Industry groups can support technical coordination but require conflict controls.
Independent research and public-interest scrutiny are necessary counterweights.
No single institution is likely to govern all layers effectively.
A polycentric system can distribute authority while sharing evidence.
Accepts an external result as informative.
Accepts that an evaluator is qualified within scope.
Accepts that a protocol, audit, or certification process meets shared requirements.
Gives formal effect under a legal system.
Evidence may be reviewed but receives no presumed weight.
Evidence may support analysis.
Evidence is accepted subject to conditions or supplemental work.
Evidence is accepted for the defined technical purpose.
Evaluator, process, or certificate is recognized within scope.
Evidence or status receives formal legal effect.
Use:
Document:
Use:
Use:
Handled by competent legal institutions.
Unresolved disagreement should remain visible.
Use:
Failure to agree should not produce false compatibility.
Different organizations can own different functions.
At its present stage, Standards Body should:
Include:
Disclose:
Separate technical mapping from legal recognition where possible.
Publish:
Use narrow, temporary processes for:
Allow challenge of recognition and mapping decisions.
Governance bodies and arrangements should be reviewable.
Characteristics:
Characteristics:
Characteristics:
Characteristics:
Characteristics:
Collect definitions across major frameworks and institutions.
Publish a stable minimum and explicit mappings.
Create a machine-readable evaluation record.
Test identity records for closed and open systems.
Register protocol versions, owners, scope, and status.
Map two evaluation protocols and two governance frameworks.
Run common systems across protocols.
Publish competence and recognition metadata.
Pilot public and restricted incident records.
Create a limited trusted network among qualified partners.
Conduct a conditional technical-recognition decision.
Support participation from underrepresented regions and smaller evaluators.
Invite external institutions to challenge the architecture.
Expand only after evidence.
Global Frontier Evaluation Interoperability Profile
Demonstrate that two independent institutions can exchange and interpret frontier AI evaluation evidence without using identical protocols or surrendering local decision authority.
Autonomous cyber capability.
This domain connects Foundations 1 through 7 and provides:
Target participation from:
Participants retain:
Each institution evaluates:
Compare:
Each institution decides whether the other's evidence is:
Simulate a cross-border safeguard failure and test:
The pilot succeeds if it:
Failure: One institution's system becomes the global default without meaningful consent.
Safeguard: modular core, localization, plural governance, alternative mappings.
Failure: Shared standards become too weak to matter.
Safeguard: common baseline plus higher-assurance profiles.
Failure: Different protocols are treated as interchangeable.
Safeguard: purpose-bounded recognition and bridge studies.
Failure: Shared terms acquire different meanings.
Safeguard: versioned vocabulary and definition governance.
Failure: Evidence is applied to a different model or configuration.
Safeguard: signed manifests, configuration metadata, re-evaluation triggers.
Failure: Numeric results travel without context.
Safeguard: required protocol, uncertainty, and task metadata.
Failure: Powerful jurisdictions or organizations control accepted evidence.
Safeguard: distributed recognition, peer review, appeals, capacity support.
Failure: Mutual recognition spreads poor assurance.
Safeguard: surveillance, proficiency testing, suspension, limited scope.
Failure: One registry controls legitimacy.
Safeguard: interoperable registries, open formats, mirrored records.
Failure: Global coordination creates a high-value breach target.
Safeguard: federation, compartmentalization, minimal disclosure.
Failure: Technical mapping is treated as legal equivalence.
Safeguard: separate technical and legal decisions.
Failure: Local-language evaluation changes construct meaning.
Safeguard: translation validation and local experts.
Failure: English performance and institutions define global capability.
Safeguard: multilingual task development, funding, and bridge studies.
Failure: Countries are expected to adopt standards they cannot implement.
Safeguard: capacity building, regional hubs, shared infrastructure.
Failure: Actors use interoperability to select the weakest regime.
Safeguard: recognition conditions and local minimums.
Failure: Institutions collect the same evidence in different formats.
Safeguard: common evidence package and crosswalks.
Failure: Legal or reputational concerns block sharing.
Safeguard: protected exchange, de-identification, corrective incentives.
Failure: Institutions continue accepting obsolete scopes or methods.
Safeguard: expiry, surveillance, automatic status updates.
Failure: Technical networks divide into incompatible blocs.
Safeguard: neutral interfaces, multilateral participation, scientific cooperation.
Failure: Standards Body or another technical institution is mistaken for a regulator.
Safeguard: precise public language and mandate boundaries.
Competition creates distrust and strategic secrecy.
Response:
Begin with narrow technical artifacts:
Residual concern:
High-sensitivity domains may remain fragmented.
Values differ.
Shared evidence does not require shared values.
Institutions can agree on what was measured while disagreeing on acceptable risk.
It can.
Response:
Recognition should preserve local minimum requirements and conditions.
They can.
Response:
Include small-actor pathways, open tools, capacity support, and competition review.
Correct.
Response:
Use peer evaluation, narrow scopes, proficiency testing, surveillance, and suspension.
It adds reporting burden.
Response:
Use a minimal core, reusable machine-readable records, and evidence portability.
Some information cannot.
Interoperability can still support:
Full harmonization is not required.
Technical and evidentiary crosswalks can coexist with legal difference.
Translation creates real measurement challenges.
Response:
Validate translation, localize where necessary, and report comparability limits.
It would also create:
Use interoperable registries.
Independent national testing is valuable.
Exclusive reliance creates duplication and limits learning.
Conditional recognition preserves sovereignty.
Their decentralized nature changes responsibility.
Interoperable identity, evaluation, incident, and community-governance systems remain useful.
How reliably can different frontier evaluation frameworks be mapped?
Which bridge methods work across dynamic, agentic, and held-out protocols?
What identity mechanisms work for closed, continuously updated systems?
Which recognition models preserve rigor without excessive duplication?
What information can be shared safely and usefully across borders?
Which methods preserve construct validity across languages and cultures?
Which capacity-building models produce durable local institutions?
How should technical evidence be mapped into distinct legal systems?
Can existing international accreditation arrangements adapt to dynamic frontier AI evaluation?
When can models and data remain local while evidence remains comparable?
How can registries remain accurate, distributed, and politically legitimate?
How quickly do recognition arrangements become obsolete?
Which technical layers can remain interoperable during political conflict?
Does interoperability improve safety and efficiency enough to justify institutional cost?
Build and validate a multilingual frontier evaluation ontology.
Develop portable identifiers and configuration manifests.
Create an open minimum schema for protocols and results.
Develop methods for mapping constructs, tasks, scoring, and thresholds.
Test statistical and qualitative comparability across institutions.
Pilot scope-specific competence recognition.
Develop public and restricted international incident records.
Test federated, confidential, and tiered evidence sharing.
Develop multilingual evaluation methodology.
Compare regional hubs, fellowships, shared facilities, and twinning.
Develop technical-to-legal crosswalk methods.
Design distributed, signed, machine-readable registries.
Pilot conditional recognition in bounded domains.
Develop event-triggered updates and suspension propagation.
Measure duplication, cost, safety, participation, and trust.
Map ten core terms across five major frameworks.
Create identity records for one closed and one open-weight system.
Have three evaluators publish the same minimum result schema.
Run two protocols on common reference systems.
Have two institutions assess whether to recognize each other's evidence.
Run a confidential cross-border incident exchange drill.
Translate and validate an evaluation in three languages.
Synchronize signed records across two independent registries.
Support a regional evaluator to implement the common profile.
Map NIST AI RMF, ISO/IEC 42001, and one frontier safety framework.
Test one technical report against two jurisdictional requirements.
Simulate method invalidation and status propagation.
A future global interoperability standard could require:
Canonical terms, mappings, and versioning.
Persistent identifiers for systems, protocols, evaluators, and reports.
Minimum human-readable and machine-readable fields.
Construct, tasks, administration, scoring, security, and expiration.
Units, uncertainty, baselines, and comparability statement.
Competence, scope, independence, security, and recognition.
Current, expired, suspended, corrected, withdrawn, or superseded.
Purpose, scope, conditions, duration, and appeal.
Classification, minimum data, sensitivity, and notification.
Access control, provenance, retention, onward disclosure, and incident response.
Language, cultural, legal, and professional adaptation.
Methods for mapping standards and requirements.
Minimum support and participation expectations.
Change control, conflicts, dissent, and public records.
Deprecation, transition, suspension, and archive.
Such a standard should be developed through STANDARDS_DEVELOPMENT_PROCESS.md with international participation.
Interoperability must preserve protocol versioning, bridge studies, and explicit discontinuity.
Cross-border use of protected evidence requires secure custody, access, and compromise response.
Capability evidence should be comparable enough to support coordinated preparation without forcing one global threshold.
Reviewer access, independence, competence, and findings need portable profiles.
Accreditation, proficiency, registries, and mutual recognition are central interoperability mechanisms.
Voluntary frameworks, standards, procurement, and law require crosswalks and recognition.
International recognition can reward interoperable evidence, while prestige competition can fragment it.
Standards Body adopts the following working positions.
Global interoperability is necessary because frontier AI systems, evidence, and deployment cross borders.
Interoperability should not be confused with global uniformity.
Shared evidence structure should generally precede attempts to impose shared policy outcomes.
Technical comparability does not require identical legal consequence.
Evidence recognition, competence recognition, process recognition, and legal recognition should remain distinct.
Every interoperable evaluation record should identify the model or system, configuration, protocol, evaluator, date, uncertainty, limitations, and status.
A model name alone is not adequate system identity.
Numeric scores should never travel without protocol and measurement context.
Protocol equivalence should be demonstrated for a stated purpose, not assumed.
Noncomparability is a legitimate and important result.
Shared terminology should include explicit mappings and disputed-term records.
Global metadata should use a stable common core with local extensions.
Evaluator recognition should be scope-specific and versioned.
Accreditation and mutual-recognition systems should be adapted carefully rather than copied mechanically.
Certification issued under one scheme should not automatically satisfy every jurisdiction.
Incident reporting should support public, trusted, and restricted disclosure layers.
International evidence sharing should be secure, purpose-limited, and auditable.
Interoperability should not compel unlawful or unsafe disclosure.
Legal crosswalks require jurisdiction-specific expertise.
Multilingual evaluation requires construct validation, not literal translation alone.
English-language performance should not define global AI capability by default.
Open-weight model lineage and configuration should be recorded through portable identity methods.
Capacity building is part of interoperability, not a separate charitable addition.
Developing countries and underrepresented regions should participate in technical design and governance, not only implementation.
Recognition networks should use plural trust anchors and avoid one global monopoly.
Interoperable registries should be portable, signed, and independently governable.
Recognition should expire or suspend after material changes or loss of confidence.
International cooperation should preserve national and institutional decision authority.
Shared minimums should not prevent stronger local safeguards.
Global interoperability should be evaluated by whether it improves evidence, reduces unnecessary duplication, broadens participation, and strengthens real decisions.
A protocol should be considered interoperable when:
Two results should be treated as directly comparable only when:
A result may receive conditional recognition when:
Recognition should be suspended when:
A global common requirement should not be proposed when:
A cross-border incident should be shared through a trusted channel when:
Protocol A:
Protocol B:
Purpose of mapping:
Date:
Reviewers:
Recognizing body:
External body or evidence:
Recognition object:
Purpose:
Date:
Incident identifier:
Reporting organization:
System:
Date:
Jurisdictions affected:
Source protocol:
Target language or region:
Version:
Date:
| Dimension | Core Question |
|---|---|
| Purpose | Is the intended cross-border use defined? |
| Vocabulary | Are terms shared or explicitly mapped? |
| Identity | Is the model, system, protocol, evaluator, and date identifiable? |
| Metadata | Is a common minimum record available? |
| Protocol | Can methods be reproduced, mapped, or bridged? |
| Measurement | Are scores, uncertainty, and baselines interpretable? |
| Comparability | Is equivalence demonstrated rather than assumed? |
| Noncomparability | Can the system clearly state when comparison fails? |
| Evaluator competence | Is scope-specific qualification legible? |
| Recognition | Are purpose, conditions, duration, and status explicit? |
| Accreditation | Are recognizing bodies competent and peer reviewed? |
| Certification | Are claims limited to scheme and scope? |
| Incidents | Can material incidents be classified and exchanged? |
| Security | Can sensitive evidence move under accountable controls? |
| Provenance | Can evidence history and modification be traced? |
| Localization | Are language, culture, law, and professional context addressed? |
| Legal mapping | Are technical and legal equivalence kept distinct? |
| Open-source fit | Can distributed model lineage and evidence be represented? |
| Capacity | Can smaller and underrepresented institutions participate? |
| Registry | Are records portable, current, and independently verifiable? |
| Governance | Are changes, conflicts, appeals, and dissent managed? |
| Resilience | Does the system avoid one trust anchor or registry monopoly? |
| Efficiency | Does interoperability reduce unnecessary duplication? |
| International utility | Can evidence support real cross-border decisions? |
| Adaptation | Can mappings and recognition evolve as AI changes? |
Global AI governance is often discussed as a choice between two extremes.
One extreme is fragmentation.
Every country, company, evaluator, and standards body creates its own:
Evidence becomes difficult to combine.
Organizations repeat the same work.
Smaller countries and institutions become dependent on dominant actors.
Incidents fail to travel.
The other extreme is forced uniformity.
One framework, one registry, one evaluator network, one risk model, or one political bloc becomes the global default.
Local law, culture, language, professional context, and institutional legitimacy are treated as obstacles rather than sources of knowledge.
Neither extreme is sufficient.
Frontier AI requires shared evidence and plural authority.
The shared layer should make it possible to know:
The plural layer should preserve the right of institutions and communities to decide:
Interoperability is the bridge.
It does not eliminate disagreement.
It makes disagreement more precise.
It does not eliminate duplication.
It distinguishes independent replication from administrative repetition.
It does not eliminate national authority.
It allows national authority to use evidence produced elsewhere without surrendering judgment.
It does not eliminate confidential information.
It creates controlled pathways for information that should not remain isolated.
It does not guarantee trust.
It provides the infrastructure through which trust can be earned, limited, reviewed, and withdrawn.
The eighth foundation of Standards Body is therefore portable evidence across plural institutions.
The future international system should not require every institution to speak with one voice.
It should make it possible for different voices to understand the same evidence.
[^nist-global]: National Institute of Standards and Technology, A Plan for Global Engagement on AI Standards, NIST AI 100-5, released 2024 and updated in 2025. https://www.nist.gov/publications/plan-global-engagement-ai-standards
[^nist-ai-standards]: National Institute of Standards and Technology, AI Standards. https://www.nist.gov/artificial-intelligence/ai-standards
[^nist-rmf]: National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework (AI RMF 1.0), 2023. https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.100-1.pdf
[^haip]: OECD, Hiroshima AI Process Reporting Framework. https://oecd.ai/en/hiroshima
[^haip-overview]: OECD, HAIP Reporting Framework Overview. https://oecd.ai/en/transparency/overview
[^haip-insights]: OECD, How Are AI Developers Managing Risks? Insights from Responses to the Reporting Framework of the Hiroshima AI Process Code of Conduct, 2025. https://oecd.ai/en/ai-publications/how-are-ai-developers-managing-risks-insights-from-responses-to-the-reporting-framework-of-the-hiroshima-ai-process-code-of-conduct
[^coe-convention]: Council of Europe, Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law. https://www.coe.int/en/web/artificial-intelligence/the-framework-convention-on-artificial-intelligence
[^coe-treaty-status]: Council of Europe Treaty Office, Chart of Signatures and Ratifications of Treaty 225. https://www.coe.int/en/web/conventions/full-list/?module=signatures-by-treaty&treatynum=225
[^un-gdc]: United Nations, Global Digital Compact, 2024. https://www.un.org/global-digital-compact/sites/default/files/2024-09/Global%20Digital%20Compact%20-%20English_0.pdf
[^un-gdc-site]: United Nations Office for Digital and Emerging Technologies, Global Digital Compact. https://www.un.org/digital-emerging-technologies/global-digital-compact
[^un-resolution]: United Nations General Assembly, A/RES/78/265, Seizing the Opportunities of Safe, Secure and Trustworthy Artificial Intelligence Systems for Sustainable Development, 2024. https://digitallibrary.un.org/record/4043244/files/A_RES_78_265-EN.pdf
[^un-ai-report]: United Nations Secretary-General's High-level Advisory Body on Artificial Intelligence, Governing AI for Humanity: Final Report, 2024. https://www.un.org/sites/un2.un.org/files/governing_ai_for_humanity_final_report_en.pdf
[^aisi-network]: UK AI Security Institute, International Consensus and Open Questions in AI Evaluations, February 12, 2026. https://www.aisi.gov.uk/blog/international-ai-network-consensus-and-open-questions
[^aisi]: UK AI Security Institute. https://www.aisi.gov.uk/
[^inspect]: UK AI Security Institute, Inspect AI. https://inspect.aisi.org.uk/
[^iso-ai]: International Organization for Standardization, Artificial Intelligence Standards. https://www.iso.org/sectors/it-technologies/ai
[^iso-sc42]: International Organization for Standardization, ISO/IEC JTC 1/SC 42 Artificial Intelligence Catalogue. https://www.iso.org/committee/6794475/x/catalogue/
[^iso-42001]: International Organization for Standardization, ISO/IEC 42001:2023, Artificial Intelligence Management Systems. https://www.iso.org/standard/42001
[^iso-42006]: International Organization for Standardization, ISO/IEC 42006:2025, Requirements for Bodies Providing Audit and Certification of Artificial Intelligence Management Systems. https://www.iso.org/standard/42006
[^iso-23894]: International Organization for Standardization, ISO/IEC 23894:2023, Artificial Intelligence, Guidance on Risk Management. https://www.iso.org/standard/77304.html
[^iso-42005]: International Organization for Standardization, ISO/IEC 42005:2025, AI System Impact Assessment. https://www.iso.org/sectors/it-technologies/ai
[^ilac-mra]: International Laboratory Accreditation Cooperation, ILAC Mutual Recognition Arrangement and Signatories. https://ilac.org/ilac-mra-and-signatories/
[^ilac-about]: International Laboratory Accreditation Cooperation, About ILAC. https://ilac.org/about-ilac/
[^ilac]: International Laboratory Accreditation Cooperation. https://ilac.org/
[^eu-ai-act]: European Union, Regulation (EU) 2024/1689 Laying Down Harmonised Rules on Artificial Intelligence. https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
[^eu-summary]: EUR-Lex, Rules for Trustworthy Artificial Intelligence in the European Union. https://eur-lex.europa.eu/EN/legal-content/summary/rules-for-trustworthy-artificial-intelligence-in-the-eu.html
[^oecd-principles]: OECD, OECD AI Principles. https://oecd.ai/en/ai-principles
[^oecd-observatory]: OECD.AI Policy Observatory. https://oecd.ai/
[^iso-casco]: International Organization for Standardization, CASCO Conformity Assessment Toolbox. https://casco.iso.org/
[^iso-recognition]: International Organization for Standardization, Recognition of Conformity Assessment Bodies. https://casco.iso.org/recognition-of-cabs.html
Date: July 16, 2026
Change type: Complete foundational edition
Summary: Establishes the fully developed canonical working white paper for Foundation 8. Defines the global interoperability problem, shared and local layers, terminology, system identity, protocol and measurement interoperability, metadata, evaluator recognition, certification portability, incident exchange, secure evidence sharing, standards and legal crosswalks, multilingual evaluation, open-weight systems, capacity building, international institutions, recognition architecture, dispute resolution, governance, maturity, implementation, a Standards Body pilot, metrics, failure analysis, objections, evidence gaps, research agenda, standards implications, operational templates, scorecard, and current primary-source research basis.
Status: Ready for internal review and future expert critique.