SETTING GLOBAL STANDARDS FOR TRUSTED AI CREDENTIALSAI Competence Framework v1.29 · current release
D1

AI foundations

PUBLISHED

What AI systems are, how they behave, and what they cannot do.

Statements
Version
v1.29
Last reviewed
30.08.2026
In scope
Categories of AI system and their characteristic behaviours
The distinction between a model, a system built around it, and a product
Non-determinism and output variability as expected properties
Context, memory and their practical limits
Fabrication and confabulation as properties of the technology, not defects to be reported
Capability and limitation patterns, and how to establish them empirically
Cost, latency and their role as design constraints
How the field is changing, and how to keep a working picture current
Out of scope
ExcludedWhere it lives
Building, training or fine-tuning modelsX-MOD
The mathematics of machine learningOut of framework
Research-level understanding of architectureOut of framework
Boundary notes

Where this domain abuts another, and how the line is drawn

Against all other domains

D1 supplies the vocabulary D3 through D8 use. Where another domain needs a concept explained rather than applied, the explanatory statement belongs in D1.

Against X-MOD

If the competence concerns choosing and reasoning about a model, it is D1. If it concerns producing one, it is the extension.

Statements by level

Every identifier is a permanent address. Indicators are normative; they state what would be observed in a person who meets the statement.

L1 Aware

4 statements
D1.L1.01Knowledge

Describe in plain terms what a generative AI system does when it produces output.

Indicators
States that output is generated rather than retrieved
Distinguishes a model from a search index or database
Avoids attributing intent, understanding or knowledge to the system
D1.L1.02Knowledge

State that an AI system may produce different output for the same input, and why that matters.

Indicators
States that repetition is not confirmation
Identifies work where variation between runs is unacceptable
Recognises that a single satisfactory result is not a result
D1.L1.03Judgement

Identify tasks an AI system is unsuitable for.

Indicators
Names at least two tasks in their own work that should not be delegated to a system
States the reason in terms of consequence rather than difficulty
Refers borderline cases rather than deciding them alone
D1.L1.04Knowledge

Describe the difference between a model, an application built on it, and the service that hosts it.

Indicators
Identifies which layer they are actually using
States that behaviour may be set by the application rather than the model
Recognises that a change at any layer can change the output

L2 Applied

6 statements

Select an appropriate system for a defined task from those sanctioned for use.

Indicators
Matches the task to a stated capability rather than to familiarity
Records the choice and the reason for it
Refers tasks no sanctioned system covers

State the limits of a system’s knowledge, including what it cannot access.

Indicators
Establishes whether the system can reach current or internal information
States the consequence of a knowledge cut-off for the task in hand
Supplies missing material rather than relying on recall

Explain accurately to a colleague what a system did and did not do.

Indicators
Describes the contribution of the system and of the person separately
Avoids overstating autonomy or reliability
Corrects a colleague’s incorrect assumption rather than leaving it

Identify where the confidence of an output is unrelated to its correctness.

Indicators
Recognises fluent, assertive output that is unsupported
Checks rather than accepting a stated certainty
Reports the distinction when passing output onward

Record which system, version and configuration produced a given output.

Indicators
Records the identifier of the system and its version
Records configuration that materially affects output
Retains the record with the work product, not separately

Recognise a change in capability or behaviour following a model or version update.

Indicators
Identifies output that has changed without a change in the instruction
Re-checks work that depended on the previous behaviour
Reports the change rather than adapting silently

L3 Proficient

8 statements
D1.L3.01Judgement

Assess whether a class of task lies within the capability of current systems.

Indicators
States what evidence would settle the question, and obtains it
Distinguishes a task that fails from a task that fails unpredictably
Records the assessment and the date it was made

Compare candidate models against the stated requirements of a defined task.

Indicators
Compares against requirements rather than against general reputation
Tests on representative material rather than on examples
States where the comparison is inconclusive

Explain the mechanism behind a failure mode to a non-specialist audience.

Indicators
Explains why the failure occurs, not only that it occurs
Uses no analogy that implies understanding or intent
Enables the audience to anticipate the failure elsewhere
D1.L3.04Judgement

Assess how context length, retrieval and tool access change what a system can do.

Indicators
Distinguishes a limit of the model from a limit of its surroundings
Identifies where added context degrades rather than improves output
States the consequence for the design under consideration
D1.L3.05Judgement

Distinguish limits inherent to the approach from limits of a particular implementation.

Indicators
States which limits a better implementation would remove
Identifies claims that assume an inherent limit has been solved
Revises the position as evidence changes
D1.L3.06Judgement

Advise on the capability, cost and latency trade-off in system selection.

Indicators
Quantifies the trade-off for the specific workload
States what is given up by the cheaper or faster option
Records the basis so the decision can be revisited
D1.L3.07Judgement

Evaluate a vendor’s description of a system against its observable behaviour.

Indicators
Identifies what the description asserts and what it omits
Tests the assertion that matters most to the intended use
Distinguishes an unsupported claim from a supported claim about something else
D1.L3.08Practice

Maintain currency as capability changes, and revise advice already given.

Indicators
Tracks change through primary rather than secondary sources
Identifies advice invalidated by a change and issues a correction
Distinguishes a capability announcement from a capability in general availability

L4 Advanced

4 statements
D1.L4.01Judgement

Establish which classes of work the organisation will and will not apply AI to.

Indicators
States the boundary in terms a practitioner can apply without escalation
Provides a route for cases at the boundary
Reviews the boundary on a stated cycle rather than on request
D1.L4.02Practice

Set the organisation’s position on model selection, hosting and dependence.

Indicators
States the acceptable forms of dependence on a single provider
Provides for substitution, and tests that it is possible
Records the decision and its reasoning for those who must apply it
D1.L4.03Practice

Establish how capability change is tracked and how existing commitments are revised.

Indicators
Assigns named responsibility for tracking
Defines what triggers a review of a live commitment
Publishes the revision rather than absorbing it
D1.L4.04Judgement

Advise a governing body on capability claims, including where a claim cannot be substantiated.

Indicators
Separates what is evidenced from what is projected
States the consequence of the claim being wrong
Declines to endorse a claim that cannot be substantiated
Editorial notes

Known gaps, open questions and contested points

Published because the record is more useful than the appearance of completeness.

The most frequently skipped domain and the most frequent root cause of failure elsewhere.

Statements should resist the temptation to become a technology survey; anything that will be wrong within eighteen months is a candidate for exclusion.

This page displays version 1.29 · last reviewed 30.08.2026