SETTING GLOBAL STANDARDS FOR TRUSTED AI CREDENTIALSAI Competence Framework v1.29 · current release
You are reading the current version of the framework, v1.29.Permanent address for this version
AICF-METH-001PUBLISHEDNORMATIVE FOR DRAFTING

How the framework is written

This document states the rules by which every competence statement in the framework is drafted, reviewed, approved, versioned and retired. It is published so that the framework can be checked against its own method by anyone, including people who disagree with it.

Twelve sections · sections 4, 5 and 6 are cited by the build validation · version 1.29

Section 1

Purpose and status of this document

This is a layer 0 document: it governs the framework rather than forming part of it. It is normative for anyone drafting or amending a statement, and informative for anyone reading one. Where this document and a statement disagree, the statement is wrong and is corrected at the next release.

It is approved by the standards board, revised only through the amendment process in section 9, and versioned alongside the framework.

Section 2

How statements are drafted, reviewed and approved

01ScopingEditorial group

The domain boundary and the competences claimed to sit inside it are written before any statement is drafted.

02DraftingDomain lead

Statements are drafted against the nine writing rules and the controlled verb list, level by level from L1 upward.

03Editorial reviewStandards editor

Wording, level placement, type assignment and indicator count are checked. Rule breaches are returned with the rule number.

04Technical reviewTwo practitioners

Two people who do the work confirm the competence exists as described and is judgeable from the indicators.

05ConsultationPublic

The draft is published for the window set by its change class. Every submission is answered in writing.

06ApprovalStandards board

The board approves the release, records dissent, and the version is cut. Nothing reaches the site unapproved.

Section 3

Evidence base and practitioner input

A statement is admitted when someone can point to work in which the competence is actually exercised. Four sources are used, and each drafted statement records which of them it rests on.

Practitioner accounts

Structured interviews and written submissions describing work actually performed, not work aspired to.

Primary source for Skill and Judgement statements

Job and role evidence

Published role descriptions, task inventories and incident records where AI work is named.

Used to test whether a competence is general

Regulatory and standards texts

Obligations imposed on organisations, read for the competences a person needs to discharge them.

Primary source for Practice statements

Technical literature

Peer-reviewed and industry work on failure modes, evaluation and system behaviour.

Used for Knowledge statements; never sufficient alone

No statement in version 1.29 rests on vendor documentation alone. Where a competence exists only because a single product behaves a certain way, it belongs in an extension or nowhere.

Section 4 · normative · #writing-rules

The nine writing rules

These rules are enforced by the build. A statement that breaks one of them does not reach the site, and a contribution that breaks one is returned with the rule number cited.

R1

One competence per statement

A statement describes a single thing a person does. If it needs the word “and” to join two actions, it is two statements.

Evaluate whether a model’s output is fit for the decision it informs.
Evaluate model output and document the evaluation.
R2

Opens with a controlled verb for its level

The first word is a verb from the level’s list in section 5. The verb sets the level; the level does not soften the verb.

Design an evaluation approach for a novel task. (L3)
Help design an evaluation approach. (verb not on any list)
R3

Observable in work

A statement describes something another person could see evidence of. Internal states are prohibited under section 5.

Report the limits of an evaluation to the person relying on it.
Understand the limits of an evaluation.
R4

Tool-neutral and vendor-neutral

No product, model, vendor or interface name appears in a core statement. Where the competence exists only for one product, it belongs in an extension.

Configure retrieval so that sources are attributable.
Configure the vendor console’s retrieval tab.
R5

Context-neutral at the core

A core statement holds in any sector. Sector-specific obligations belong in an extension; function-specific ones in an overlay.

Record the basis on which an output was accepted.
Record the basis on which an output was accepted for clinical use.
R6

Present tense, active voice, third person implied

The subject is the competent person. No modal verbs: “should”, “may” and “is able to” state an aspiration rather than a competence.

Diagnose why an evaluation result is not reproducible.
Should be able to diagnose evaluation results.
R7

Two to four performance indicators, each independently assessable

Indicators are normative and state what an assessor looks for. One indicator is not a statement; five means the statement is too broad.

3 indicators, each judgeable from one piece of evidence.
1 indicator restating the statement in other words.
R8

Assessable without a specified assessment method

The statement says what competence is, not how to test it. Methods belong to awarding bodies. Evidence examples are non-normative and clearly labelled.

Justify a decision to reject an output.
Pass a written test on output rejection.
R9

No forward dependency

A statement may assume statements at the same or a lower level, never a higher one. Cross-domain relationships are declared as related, not as prerequisites.

assumes: D6.L2.03 (from an L3 statement)
assumes: D6.L4.01 (from an L3 statement)
Section 5 · normative · #verbs

Controlled verb lists

Every statement opens with a verb from the list for its level. This is the mechanism by which L2 in one domain means the same thing as L2 in another. The lists are closed: additions are amendments to the framework and follow the versioning policy in section 11.

L1 Aware
identifydescriberecogniseliststatedistinguishlocatename
L2 Applied
applyuseproduceconfigurefollowdocumentrecordselectcheckreport
L3 Proficient
designevaluateadaptdiagnoseintegratejustifyoptimisereviewcritiqueresolveadvise
L4 Advanced
defineestablishleadarbitratesetgovernresolveadvancecommissionhold accountable
Prohibited at every level
understandappreciateknowbe aware ofbe familiar withgrasphave knowledge of

These describe an internal state rather than an observable action and cannot be assessed. A statement containing one of them fails the build.

Section 6 · normative · #schema

Statement schema, field by field

The published schema is the authority. This table restates it in prose.

FieldTypeRequiredRule
idstringYesMatches ^(D[1-8]|O[1-3]|X-[A-Z]{3})\.L[1-4]\.\d{2}$. Canonical casing. Never reused, never re-pointed.
statementstringYesThe competence text, written to the nine rules in section 4.
domainstringYesThe owning unit: a domain, overlay or extension code.
levelL1–L4YesMust agree with the level segment of the identifier.
typeenumYesKnowledge, Skill, Judgement or Practice. One value only.
indicatorsstring[]YesTwo to four, normative. Fewer than two fails the build.
assumesstring[]NoStatement identifiers at the same or a lower level. Resolved as links.
relatedstring[]NoIdentifiers or domain wildcards such as D2.L2.*. Rendered on both pages.
statusenumYesactive, deprecated or superseded. Deprecated statements resolve permanently.
stabilityenumYesstable or provisional. Provisional statements may change within a minor release.
introducedstringYesThe version in which the identifier first appeared.
supersededBystringNoRequired when status is superseded. Must resolve to an existing identifier.
Validation — the build fails on any of these
FAILAn identifier that does not match the pattern in section 7.
FAILA duplicate identifier anywhere in the collection.
FAILAn assumes or related reference to an identifier that does not exist.
FAILA wildcard resolving to a domain that does not exist.
FAILAn assumes reference pointing to a higher level than the referring statement.
FAILFewer than two performance indicators.
FAILAn opening verb absent from the controlled list for the statement’s level.
FAILAny prohibited verb anywhere in the statement text.
FAILstatus: superseded without a supersededBy that resolves.
Section 7

Identifier scheme and permanence

An identifier takes the form D6.L3.04: unit, level, sequence. Units are D1D8 for domains, O1O3 for overlays and X-AGT style codes for extensions.

An identifier is never reused for a different competence.
An identifier never changes level, domain or sequence number.
A deprecated identifier resolves permanently, with the version of deprecation and any successor.
Every version of a statement remains readable at its dated address.
Section 8

Editorial roles and who holds them

Standards editorHolds the writing rules. Final say on wording, level placement and type.
Domain leadDrafts and maintains one domain. Answers submissions against it.
Technical reviewersConfirm a competence exists as described and is judgeable.
Standards boardApproves releases and settles scope disputes. Records dissent.
RegistrarCuts releases, freezes dated trees, maintains the deprecation record.

Named holders, terms and declared interests are published on the board page.

Section 9

How disagreements are resolved

Disagreement about wording is settled by the writing rules. Disagreement about whether a competence exists is settled by evidence under section 3. Only disagreement about scope reaches the board, and its decision is published with the dissent recorded.

01Wording: settled by the standards editor against the nine rules. The rule number is cited in the response.
02Existence: settled by evidence under section 3. Absent admissible evidence, the statement is not drafted.
03Level or type: settled by the standards editor with one technical reviewer. Escalates to the board if they disagree.
04Scope: only the board decides whether a competence belongs to the framework, and publishes the decision with dissent.

Every decision and its dissent

Section 10

Consultation, and how submissions are answered

Every submission receives a written response citing the rule or the evidence on which it was accepted, amended or rejected. Responses are published with the release record for the version in which the submission was considered.

Make a submission

Section 11

Versioning policy and stability guarantees

Change classWhat it coversConsultation minimumVersion step
EditorialWording clarified without changing meaning. Typography, links, non-normative examples.Nonepatch
AdditiveNew statements, new indicators, new extensions or overlays. Nothing existing changes meaning.30 daysminor
DeprecatingA statement retired or superseded. The identifier resolves permanently thereafter.60 daysminor
StructuralLevels, types, domain boundaries, the schema, or this methodology.90 daysmajor

A statement is never edited in place to mean something different. It is deprecated, and a new identifier carries the new meaning. Deprecated identifiers resolve permanently.

Version history and the deprecation record

Section 12 · stated by the organisation

Known limitations of this methodology

Publishing the method invites the question of where it is weak. These are the answers we would give if asked, so we give them unasked. Each is reviewed at every release and this list is expected to change.

Level boundaries are judged, not measured

The L2/L3 boundary in particular rests on editorial judgement applied consistently rather than on external calibration.

Action: inter-rater agreement study on level placement, published with its results whatever they show.

No assessment validity evidence yet

The framework asserts that its statements are assessable. No awarding body has yet published results that would test whether they are.

Action: publish validity findings from the first credentials, including negative findings.

The subject matter moves faster than the release cycle

A consultation window is slow relative to changes in the systems the framework describes. This is a deliberate trade against churn in citable identifiers.

Action: provisional stability marking, reviewed at each release.

English-language drafting only

The controlled verb lists are English and carry connotations that will not survive translation cleanly.

Action: translation policy published before any non-English edition.

This page displays version 1.29 · last reviewed 30.08.2026 · release record: /framework/v1.29