Trustworthy AI Agent Standards Family — External Standards Alignment

Abstract

This document defines the external-standards alignment model for the Trusted AI Kernel (TAK), Global AI Agent Identification and Governance (GAID), and Job-Specific Intelligence (TAK-JSI) standards family.

The standards family does not replace organization-level AI governance, model evaluation, identity and authorization protocols, credential formats, software testing, or conformity assessment. It composes those bodies of work around an operational assurance subject that is not consistently represented elsewhere: an enduring AI agent identity operating through a materially versioned configuration, qualified for a bounded job and constrained at runtime by live authority, data, oversight, and evidence controls.

The principal augmentation is the end-to-end assurance chain:

enduring agent identity
  -> versioned operating profile
  -> job- and context-specific qualification
  -> runtime authority and autonomy enforcement
  -> attributable action evidence
  -> surveillance, drift detection, and revalidation

This document is informative. Normative requirements remain in the three canonical standards.

1. Scope

This document:

• identifies the standards and active standards-development work most relevant to the family
• states whether each external work is adopted, profiled, augmented, or merely adjacent
• identifies the specific gap filled by TAK, GAID, or TAK-JSI
• prevents the family from redefining protocols or conformity models that already have authoritative owners
• provides a standards-body contribution map for future liaison, incubation, and formal proposals
• establishes terminology for implementation statements and cross-standard conformance claims

This document does not:

• confer endorsement by any cited standards body
• claim conformance to an external standard without the evidence that standard requires
• turn an active draft, community-group report, or project authorization request into a published consensus standard
• define a new credential transport, cryptographic primitive, directory protocol, occupational taxonomy, or certification-body accreditation model

The execution sequence, readiness gates, contribution packages, and go/no-go criteria for external engagement are maintained separately in the informative Standards Contribution Roadmap. This document remains the source of truth for technical overlap, synergy, augmentation, and non-duplication boundaries.

2. Alignment Vocabulary

The following relationship terms are used throughout this document:

An implementation statement should not use conforms to, certified to, or an equivalent claim for an external standard when the implementation has only mapped to or profiled part of that standard.

3. Layered Standards Architecture

The family occupies a composition layer between broad governance and protocol implementation:

4. NIST Alignment

4.1 NIST AI Risk Management Framework

The NIST AI RMF 1.0 provides the broad GOVERN, MAP, MEASURE, and MANAGE risk functions.

The family:

• maps-to GOVERN through accountable issuers, owners, data stewards, qualification authorities, and runtime policy ownership
• maps-to MAP through the TAK-JSI job, activity, stakeholder, data, deployment, and harm context
• maps-to MEASURE through qualification plans, representative evaluations, uncertainty, evidence provenance, operational surveillance, and regression detection
• maps-to MANAGE through TAK action gates, autonomy ceilings, escalation, suspension, revocation, and revalidation
• augments the framework by binding those activities to a stable agent identity and a specific, fingerprinted operating profile

An AI RMF implementation is organization- and system-oriented. A TAK-JSI qualification is deliberately narrower: it answers whether one identified operating profile has demonstrated fitness for one declared job and context.

4.2 NIST AI Agent Standards Initiative

The NIST AI Agent Standards Initiative organizes current U.S. work around industry-led standards, community protocols, agent identity, and security evaluation.

The family supplies candidate material for the initiative’s stated gap-analysis and measurement work:

• TAK: testable runtime requirements for authority, tools, data, oversight, delegation, evidence, and autonomous operation
• GAID: requirements for identity continuity, operating-profile binding, claim status, and attributable receipts
• TAK-JSI: comparable evaluation without reducing fitness to a model-wide or context-free benchmark score

NIST engagement is treated as pre-standardization, measurement, and implementation-validation work. Formal consensus text is expected to progress through the relevant standards-development organization.

5. ISO/IEC Alignment

5.1 ISO/IEC JTC 1/SC 42

ISO/IEC JTC 1/SC 42 is the horizontal ISO/IEC committee for artificial intelligence. The relevant work allocation is:

The family is not proposed as a replacement for SC 42’s horizontal corpus. It is a compositional profile and potential new-work contribution spanning trustworthiness, testing, lifecycle, data, and conformity assessment.

5.2 Management, Risk, Impact, and Oversight

5.3 Quality, Testing, and Data

5.4 Conformity Assessment

TAK-JSI qualifies an AI agent operating profile. It does not certify a human and does not appropriate the terminology or accreditation model of personnel certification.

A TAK-JSI implementation is expected to preserve the normative distinction among:

• self-assertion
• organization attestation
• independent assessment
• accredited certification

Only the last category implies operation under a recognized accreditation and certification scheme.

6. IEEE Alignment

6.1 IEEE P3709

IEEE P3709 is an active project for a framework and technical requirements for agentic AI.

TAK is suitable as a requirements and conformance contribution addressing:

• the agent harness as a control plane distinct from the probabilistic planner or model
• principal-to-action authority continuity
• least-authority tool invocation and delegation narrowing
• policy-enforced data and provider eligibility
• qualification-aware action and autonomy ceilings
• proactivity that cannot widen authority, competence, or oversight
• evidence-earned autonomy progression and mandatory regression
• material-change detection and validation continuity
• reconstructable, attributable execution evidence

The intended relationship is augments, not replacement: P3709 provides the broader agentic-AI framework, while TAK supplies implementable runtime-governance requirements.

6.2 IEEE P3833

IEEE P3833 addresses proactive AI agents in multimodal human-computer interaction.

TAK contributes the cross-domain distinction that:

requested proactivity != delegated authority != demonstrated qualification != permitted autonomy

P3833’s multimodal interaction scope remains externally owned. TAK augments proactivity controls by defining the hard ceilings and evidence conditions that a user-experience preference cannot override.

7. W3C and 1EdTech Alignment

7.1 Agent Identity Registry Protocol Community Group

The W3C Agent Identity Registry Protocol Community Group is developing verifiable agent identity infrastructure, including agent credentials, trust negotiation, lifecycle management, and integration with MCP, A2A, OAuth/OIDC, and SPIFFE.

GAID:

• adopts externally governed credential and cryptographic mechanisms
• profiles agent credentials with AIDoc, operating-profile, qualification, and receipt semantics
• augments the identity layer by separating the enduring subject from mutable operational state
• augments lifecycle status with material-change and revalidation semantics
• leaves DID methods, credential proof formats, and trust negotiation to their owning standards

7.2 Agent Declaration and Assurance Community Group

The W3C Agent Declaration and Assurance Community Group is developing declarations of agent identity, ownership, software versions, models, operating boundaries, conformance profiles, and runtime bindings.

This is the closest active overlap with GAID. The preferred relationship is a coordinated profile:

GAID must not create a parallel universal manifest where an interoperable W3C manifest can carry the same semantics.

7.3 Agent Trust Protocol Community Group

The W3C Agent Trust Protocol Community Group addresses verifiable identity, trust scoring, privacy-preserving interaction, and conformance testing.

GAID can provide verified inputs to trust decisions, but it deliberately does not define a universal scalar trust score. Trust remains contextual: a valid identity or qualification for one job does not imply general trustworthiness or authorization for another.

7.4 Verifiable Credentials and Open Badges

The W3C Verifiable Credentials Data Model 2.0 and 1EdTech Open Badges 3.0 already define portable, signed claim envelopes with issuer, subject, evidence, validity, expiry, status, and verification patterns.

GAID is intended to profile those formats for portable qualification and assurance claims rather than define a competing credential envelope. It augments them with AI-agent-specific semantics:

• enduring agent subject and operating-profile fingerprint
• job/activity/data/risk scope
• qualification scheme and evidence references
• maximum evidence-supported autonomy tier
• material-change and pending-revalidation state
• explicit separation of a credential claim from live runtime authorization

8. IETF, OpenID, and Workload-Identity Alignment

8.1 IETF WIMSE

The IETF Workload Identity in Multi-System Environments working group defines workload identity architecture, identifiers, credentials, and authentication mechanisms. Current AI-agent work includes:

• WIMSE Applicability for AI Agents
• AI Agent Authentication and Authorization
• Signed Authorization-Evidence Records
• Cross-Organizational Delegation for Workload and Agent Identity

These drafts are works in progress and are not cited as completed standards.

The family relationship is:

• WIMSE, SPIFFE, OAuth, GNAP, mTLS, HTTP Message Signatures, and proof tokens authenticate and authorize workloads across protocol boundaries
• GAID binds the enduring AI subject and governed operating state to those workload identities
• TAK determines whether an authenticated request is permissible under current authority, qualification, data, and oversight constraints
• GAID and TAK add qualification and policy-decision references to attributable action evidence
• protocol-level authorization remains necessary but is not sufficient evidence of job competence

8.2 OpenID Foundation AIIM

The OpenID Foundation AI Identity Management Community Group provides a venue for AI-agent identity use cases, threat modeling, modular roles and scopes, and interoperability with open identity standards.

GAID is positioned as an AI-agent assurance and operating-state profile for established OpenID/OAuth deployments. It does not replace OAuth authorization servers, OpenID providers, tokens, proof-of-possession, or enterprise identity policy.

9. Protocol and Evidence Carriers

The family defines the semantics that these carriers transport. Implementations should preserve the native identifiers and verification material of adopted protocols instead of translating them into opaque, unverifiable prose.

A2A’s enterprise guidance covers transport security, standard web authentication, authorization, data minimization, observability, audit, and API management. Its core leaves identity outside A2A payload semantics and leaves authorization boundaries implementation-defined; its multi-tenancy identifier is opaque routing, not agent identity. Consequently, GAID profiles A2A through its standard extension points rather than changing the A2A task or transport model. The profile adds the enterprise identity concerns A2A does not normatively define: canonical agent subjects, private-to-public aliases, full source-side multi-agent custody, public-only boundary projection, and signed commitments to withheld participation.

10. Distinctive Augmentation

The family addresses a gap left when adjacent standards are implemented independently.

10.1 Identity Is Not Operating State

An agent can retain its enduring identity while its model, instruction bundle, tools, retrieval sources, memory policy, provider routing, qualification, or autonomy posture changes. GAID binds both without conflating them.

10.2 Capability Is Not Qualification

A model benchmark, system card, declared skill, successful demonstration, or broad capability evaluation can contribute evidence. None alone establishes that the complete operating profile is qualified for a job under representative tools, data, policy, and consequence boundaries.

10.3 Qualification Is Not Authorization

TAK-JSI establishes a ceiling on permissible autonomy. TAK still requires live principal authority, route and workflow grants, tool policy, data eligibility, and mandatory oversight for every action.

10.4 Proactivity Is Not Autonomy

Proactivity expresses initiative and interaction preference. Autonomy expresses permitted execution latitude after hard constraints and evidence have been evaluated. Increasing proactivity cannot raise the autonomy ceiling.

10.5 Validation Does Not Survive Material Change Silently

Qualification and assurance claims are bound to a fingerprinted operating profile. A material change places affected claims into revalidation unless the applicable scheme contains evidence that the change is immaterial.

10.6 Evidence Closes the Lifecycle

Action receipts and operational surveillance feed subsequent qualification and autonomy decisions. The standards family therefore connects pre-deployment evaluation to runtime behavior and post-deployment revalidation rather than treating testing as a one-time gate.

11. Standards-Contribution Profile

A contribution derived from this family should contain:

• a neutral problem statement and declared scope
• a gap analysis against the specific committee’s published and active work
• normative terminology with explicit exclusions
• a clause-level requirements draft
• a machine-readable information model
• conformance assertions and test vectors
• at least two independent implementation or pilot contexts
• privacy, security, safety, accessibility, and internationalization considerations where applicable
• patent, copyright, contribution-license, and reference-implementation terms

The complete artifact matrix and staged readiness criteria are defined in the Standards Contribution Roadmap.

The preferred contribution allocation is:

For ISO/IEC, a formal new-work proposal should follow the ISO proposal process through the relevant national body or other authorized proposer and should identify a project leader, market need, affected stakeholders, relationship to existing work, and an initial draft or outline.

12. Implementation and Claiming Guidance

An implementation using this alignment model should:

• identify the exact version or dated draft of each external specification used
• distinguish published standards from active drafts and community-group work
• avoid implying endorsement by a cited organization
• avoid claiming that a GAID, credential, badge, or qualification grants live authority
• avoid representing organization-attested assessment as accredited certification
• publish a crosswalk from family conformance assertions to adopted external requirements
• preserve external identifiers, credential status, proof material, and evidence provenance
• document deviations and the project-specific reason for each deviation
• allow other conforming protocol carriers where the semantic and verification requirements are preserved

13. Open Standardization Questions

The following questions require multistakeholder resolution:

• which W3C manifest or credential vocabulary should carry the canonical public AIDoc projection
• whether a globally resolvable GAID requires a new identifier scheme or can remain a semantic layer over existing URI, DID, WIMSE, SPIFFE, and enterprise identifiers
• which material-change classes can preserve qualification without full reassessment
• how qualification schemes establish representative job and data conditions across sectors
• which evidence can be public, selectively disclosed, auditor-only, or prohibited from disclosure
• how independent assessors and certification schemes demonstrate competence and impartiality
• how cross-organizational relying parties discover trusted issuers, schemes, and status services
• how sector regulators constrain autonomy independently of technical qualification
• how interoperability tests demonstrate that badges and manifests are not mistaken for live authorization

These questions are intentionally not resolved by unilateral DPF implementation. They identify where standards-body consensus is required.

14. Summary

The family complements a mature but segmented standards landscape. Its contribution is the composition contract that makes an AI agent’s identity, current operating state, job qualification, runtime authority, autonomy posture, evidence, and revalidation status mutually consistent and verifiable.

The standards should therefore progress through coordinated contributions:

• TAK as runtime-governance requirements
• GAID as an identity and assurance semantic profile
• TAK-JSI as a job-specific operating-profile qualification scheme

This positioning preserves existing protocol and conformity-assessment authorities while filling the operational gap between organization-level governance, model evaluation, identity, and consequential autonomous action.