Most work on grounding agentic AI focuses on meaning: better ontologies, semantic layers, knowledge graphs and retrieval. ARF argues that a second layer is missing — an ontology of authority.
A perfectly consistent ontology can tell an agent what a role means, what a discount is, which workflow applies and what a system permission allows. It still may not tell the agent whether the documented authority matches what is actually happening on the ground.
Everything agrees.
Except reality.
For a human, the discrepancy may be absorbed through experience. For an autonomous agent, a documented structure can produce a confident, consistent decision that is nevertheless wrong. The source paper frames this as a missing ontology of authority rather than a data-quality problem.
One decision.
Five domains.
The framework traces an authority relation through five layers: who holds the role, what the term means, how the process is codified, what the system enforces, and what is actually true in the world.
Move from the structure an agent is given to the authority it encounters.
Who holds the role
Actor, role, relationship and organizational position.
Make the split measurable.
actual exceeds documented
The supplied ARF example expresses a divergence magnitude of 0.62 and a direction of actual-exceeds-documented. The coefficient is presented as an auditable measure of the split between documented and lived authority.
The framework deliberately treats this as a scored, directional divergence rather than a metaphor. The aim is to make authority discrepancies inspectable in the same operational spirit as a data-quality rule.
From concept
to relation.
The academic version defines an Authority Relation primitive. Its purpose is to connect an actor, action and object to the cross-domain evidence that resolves the authority claim.
resolvesTo makes the cross-domain consistency requirement checkable: the process step, system permission and real-world fact can be validated as references rather than left as prose. affirmedBy places business, technical and governance affirmation directly into the structure.
{
"@context": "https://arf-schema.org/v1/context.jsonld",
"@id": "arf:ar-instance:disc-2026-0614-7741",
"@type": "arf:AuthorityRelation",
"arf:actor": {
"@id": "org:role:regional-commercial-director-emea-north",
"arf:heldBy": "org:person:p-10234"
},
"arf:action": { "@id": "arf:action:approve" },
"arf:object": {
"@id": "biz:discount-request:dr-88291",
"arf:resolvesTo": [
"process:workflow-step:pricing-approval-gate-3",
"sys:crm-permission:perm-4471",
"world:fact:margin-impact-q3-2026"
]
},
"arf:dnaCoefficient": {
"arf:divergenceMagnitude": 0.62,
"arf:direction": "actual-exceeds-documented"
}
}Check authority
before action.
The source describes an MCP-backed Authority Relation store as one practical integration point: an agent can check divergence and affirmation status before a write-capable action. It also explicitly notes that the integration substrate itself must be trusted; an authority check inside a compromised server is not protection against that compromise.
Authority resolution is a governance problem — with architecture consequences.
Business, technical and governance affirmation become part of the authority record.
When an object resolves consistently across layers, state changes can propagate without bespoke point-to-point reconciliation.
The agent can evaluate authority as a prerequisite to consequential action rather than infer it from documents.
The framework is presented here as a Portend Labs research artifact. Claims about external studies, market adoption and related research remain those of the supplied paper and are not independently verified on this page.