Every term in the F-ACT standard has one fixed meaning. This page is the canonical reference — use these definitions, not paraphrases, wherever conformance or interoperability matters.
The third GUARD dimension. Answers: is every consequential action recorded in a form that survives a dispute? A conforming envelope must set recordsConsequentialActions and tamperEvident to true.
See also: consequential-action, tamper-evident
An explicit, enumerated set of actions an agent is permitted to perform. In F-ACT, the allow-list lives in scope.allowed and works together with default-deny: everything not listed is forbidden.
See also: usage, default-deny
The tier an envelope’s author intends the agent to satisfy (L1, L2, or L3). Declared in the conformanceTargetTier field. The reference checker reports results for all three tiers regardless of the target.
See also: l1-declared, l2-enforced, l3-provable
An action taken by or attributed to an AI agent that has a material effect — creating, modifying, or deleting a record; sending a message; making a decision that affects a person. The Audit dimension requires that all such actions are recorded.
See also: audit
The fifth GUARD dimension. Answers: what data may this agent touch, on what legal basis, and may the data leave its boundary? A conforming envelope must state a lawful basis and declare whether data may egress outside its permitted boundary.
See also: lawful-basis, egress-boundary
The principle that any action not explicitly on the allow-list is forbidden. In an F-ACT envelope, scope.defaultDeny must be true for L1 conformance. Without it, the allow-list is advisory rather than a boundary.
See also: allow-list, usage
The declared perimeter beyond which data handled by the agent may not travel. At L1, data.egressOutsideBoundary must be false, meaning the agent commits that data stays within its permitted destinations.
See also: data
The L2 (Enforced) concept: a runtime mechanism that checks policy before an agent acts and blocks the action if the policy says no. In an F-ACT envelope, enforcement is declared via a top-level enforcement block that names the policy decision point, asserts preventive mode, and provides a decision-log endpoint.
See also: l2-enforced, policy-decision-point
Framework for AI Compliance and Transparency. An open, royalty-free standard for AI agent governance. The name is always written with a hyphen (F-ACT), never as FACT. The standard is maintained by the Open Conformance Foundation (being established) and the core claims are pledged on irrevocable, royalty-free terms.
The five-dimension governance framework at the core of F-ACT. Every AI agent’s governance envelope must answer one question per dimension: Grant, Usage, Audit, Revocation, and Data. The name is the initials of the five dimensions in order.
See also: grant, usage, audit, revocation, data
The first GUARD dimension. Answers the question: who granted this agent its authority, and can that grant be traced to an accountable human? A conforming envelope must name a human principal whose authority the agent’s actions resolve to, and set traceableToNamedHuman to true.
See also: principal, traceable-to-named-human
A single JSON document that answers all five GUARD questions for one AI agent. It is not a certificate or an endorsement — it is a machine-readable self-declaration of the agent’s governance posture. The envelope is what the reference checker validates.
See also: guard
The first conformance tier. The agent’s governance envelope answers all five GUARD questions. This is fully verifiable from the document itself — the reference checker can confirm it with a static check. Declared is the floor of the standard.
See also: guard, governance-envelope
The second conformance tier. The envelope declares that governance is enforced at runtime: policy is checked before the agent acts, the default is deny, and a decision log is maintained. A document can only declare this — real proof requires runtime evidence from the policy decision point. The checker labels L2 results as “declared only.”
See also: enforcement, policy-decision-point
The third conformance tier. The envelope declares that tamper-evident provenance exists, a third party can verify it, and revocation is cryptographically provable. Like L2, a document can only declare this — real proof requires fetching and checking the evidence. The checker labels L3 results as “declared only.”
See also: provenance, tamper-evident
The legal or policy ground on which the agent is permitted to touch data (e.g. consent, contract, legitimate interest). Declared in data.lawfulBasis. The standard does not prescribe which bases are valid — it requires that one is stated.
See also: data
The baseline tier. The agent has no governance envelope. Nothing is declared, enforced, or provable. This is where every agent starts before adopting F-ACT.
The entity that maintains the F-ACT standard without a commercial interest in any particular product built on it. The intended neutral steward is the Open Conformance Foundation (being established). Until the Foundation is a legal person, the core claims are held by Society OS Pty Ltd under an irrevocable, royalty-free pledge.
The named location in a system where the allow/deny decision is made before an agent acts. Required at L2 in enforcement.policyDecisionPoint. The checker verifies it is declared; proving it actually blocks actions requires runtime evidence.
See also: enforcement, l2-enforced
The named human whose authority the agent’s actions ultimately resolve to. Declared in grant.principal. An agent without a named principal cannot reach L1 — traceableToNamedHuman must be true and a non-empty principal string must be set.
See also: grant, traceable-to-named-human
The L3 (Provable) concept: tamper-evident evidence of what the agent did, stored in a format a third party can independently verify. Declared in a top-level provenance block that names the evidence type, a verification endpoint, and a revocation-proof mechanism.
See also: l3-provable, tamper-evident
The fourth GUARD dimension. Answers: can its authority be withdrawn instantly — and proven withdrawn? A conforming envelope must set revocable, realTime, and thirdPartyVerifiable to true.
A property of records or evidence meaning that any alteration is detectable. At L1, the Audit dimension requires tamperEvident to be true for the action log. At L3, provenance evidence itself must be tamper-evident (e.g. a signed hash chain).
See also: audit, provenance
The second GUARD dimension. Answers: what is this agent permitted to do — and what is explicitly out of bounds? A conforming envelope must carry an explicit allow-list of permitted actions and set defaultDeny to true, meaning anything not on the list is forbidden.
See also: allow-list, default-deny