This document specifies how an agent trust record is displayed, written and exported. A trust record describes one AI agent: its identity, interface, mandate and the parties who build, run and authorise it, together with a trust assertion over its identity, mandate and policy. Agent type is described with the Agentic AI Ontology [[AIAO]], and the record terms are aligned with the KR for Trust ontology [[KRT]].

DRAFT. This is a draft for discussion in the AI Knowledge Representation Community Group. It has not been adopted by the group. The fields, views and vocabulary alignment described here are open for comment, and the reference visualiser [[TRUST-RECORD-VIS]] is offered for testing. After review, the group will decide whether to publish a revised version as a Community Group Report.

To enter comments, suggestions, sections or evaluation results for this visualiser specification, open an issue or PR. Posts to the group's mailing list, public-aikr@w3.org, should carry [trustvisualiser] in the subject.

Implementation reports, test results and reruns are discussed separately from the model: open an issue with [trustvisualiser-impl] at the start of the title. Each report is listed in Implementation reports.

Scope

This specification covers how an agent trust record is shown, written and exported. Verification of the claims in a record is out of scope; that work belongs to [[KRT]] and the trust-layer technical note [[KR-OF-TRUST]].

Reference implementations:

Data model

A trust record describes one agent with thirteen declared fields in four groups, followed by a trust assertion over three of them. Fields marked proposed have draft descriptions open for review.

GroupFieldDescription
IdentityAgent name or identifierThe name, DID or other identifier the agent is known by.
IdentityAgent typeOne class per AIAO dimension, plus cognitive capabilities.
InterfaceProtocol (proposed)How it talks to other agents and services, e.g. MCP, A2A, an HTTP API.
InterfaceWhere it lives and operatesURL, DID address or other.
InterfaceKey publisher (proposed)Who publishes the key set a verifier trusts: the agent's own origin, or an issuer or network the verifier has chosen.
MandateMandate / functionsWhat the agent does.
MandateAuthority (proposed)Who granted it the power to act, and within which limits.
MandatePolicy (proposed)The rules it operates under, e.g. a published usage or data policy.
PartiesDeveloperWrites the agent.
PartiesDeployerRuns the agent.
PartiesUserAuthorises use as per the mandate, under the policy.
PartiesOwner (proposed)Who owns the agent.
PartiesFunder (proposed)Who paid for the agent.

The trust assertion states that the agent's identity, mandate and policy hold:

Identity
Name and location resolve to the agent.
Mandate
Its actions stay inside the declared functions and authority.
Policy
It runs under the declared policy.

A record states what was declared. Checking a declaration against what the agent did is outside this specification.

Visualisation requirements

A conforming visualiser MUST provide the three views below.

  1. Model: shows the four groups and the trust assertion side by side, each field with its description, and a diagram of the parties around the agent (developer builds, deployer operates, user authorises per mandate under policy, trust checks).
  2. Write a record: a form with one input per field. Agent type offers one class per AIAO dimension and a list of cognitive capabilities; any of them MAY be left blank. A preview card SHOULD show the record as it is written.
  3. Register: a table of all records, one row per agent, in the field order of the data model. An invented example row MAY be included and MUST be marked as an example.

A conforming visualiser MUST export records as:

Records SHOULD be kept in the viewer's browser only, with nothing sent to a server.

Each view SHOULD link to the Agent Ontologies Explorer [[AGENT-ONTO-EXPLORER]], so readers can look up an AIAO class from a record.

Accessibility of the visualisation

A conforming visualiser MUST be usable with a keyboard and a screen reader, and SHOULD meet [[WCAG21]] level AA.

The reference implementation provides:

Barriers found by users are reported as issues tagged [trustvisualiser]. Automated checks find only some barriers; testing with people who use screen readers or magnification is planned.

Alignment with AIAO and KRT

AIAO describes what kind of agent is involved; KRT records whether a relying party could act on a claim the agent presented. The draft mapping below is preliminary and open for review.

Trust record or AIAO termKRT term (v0.1)Relation
Agent (AIAO)krt:AgentAIAO classes describe the kind of agent.
Functional role (AIAO)krt:MandateA role names the kind of work; a mandate states the scope granted.
Activity (AIAO)krt:ActionAIAO activity classes, once defined, can type KRT actions.
Deployerkrt:OperatorEquivalent.
Userkrt:PrincipalEquivalent.
Policykrt:DeclaredPolicyEquivalent.
DevelopernoneGap; proposed as an addition to KRT.

The visualiser's type picker uses AIAO v1.1 (eight dimensions). AIAO version 2.3 describes six dimensions and models functional roles as roles an agent plays. The two need to be brought into line.

Implementation reports

Implementations of this specification, and test results that relate to it, are listed here. Each entry links to its evidence and carries one of three status levels:

Documented
Described by the contributor; not run.
Run by the contributor
Run by the contributor, with the results and the exact version published.
Independently rerun
Rerun by someone other than the contributor, with the results published.
ImplementationContributorScopeStatusEvidence
No reports yet.

Questions for evaluation

Comments are especially welcome on these points:

  1. Are the thirteen fields enough to describe an agent you build, run or rely on? Which are missing or redundant?
  2. Should the descriptions of Protocol, Authority and Policy be adopted as drafted?
  3. Should Developer be added to KRT, or recorded outside it?
  4. Should "Where it lives and operates" accept several values, such as a URL and a DID together?
  5. Which AIAO version should the type picker follow: v1.1 (eight dimensions) or v2.3 (six dimensions, roles)?
  6. Do the three views make a record easy to read for someone outside the group? Try writing a record for an agent you know and report what was unclear.
  7. Are CSV, JSON and Turtle the right exports, and is the Turtle output usable with KRT?
  8. Does the visualiser work for you with a keyboard, a screen reader, larger text or high contrast? Please report any barrier as an issue.

Example records for real agents are especially useful for testing. The chair will collect the comments, revise this draft and the visualiser, and bring a revised version to the group for a decision on publication.

Contributions

The following contributions have been made to this draft and to the visualiser. Further contributions are invited; see the status section for how to send them.

ContributorContributionStatus
Larry Lewis (W3C Agent Trust Protocol CG)Review of the trust record model: separate declared, verified, permitted and observed states; declared functions from granted authority; identifiers from service endpoints; test the role mappings against deploymentsPartly implemented: "declared, not verified" labels and an "unknown" verification state in the visualiser. Other points open for discussion as issues
Yanfeng Liu (Nanjing University, V19)Reviewer for audit and accountability; reproducible examples for evidence freshness, human approval and revocationListed as reviewer; examples offered
Francesco Marinoni MorettoCommerce identity rows with pinned sources; who publishes the key set a verifier trusts (agent's own origin, or an issuer or network the verifier chose)Contributed: rows confirmed with corrections; key-publisher column proposed and added as a proposed field; draft alignment issue between UCP and Web Bot Auth confirmed and reported to UCP (issue 1080)
Kenne Ives (AgentAvow)did:web example and signed tool-definition test vectors; pinning the key from a source independent of the issuerOffered; not yet run