AIR APAC Standard for Managerial Judgment in AI-Enabled Work
Designation: AIRS-J/1 Version: 0.3 (Draft for external review) Status: Not issued. No certification may be granted against this version. Issuing body: AIR APAC Pte. Ltd. (UEN 202600913G), Singapore Effective date: Not set Review cycle: Annual, with interim errata permitted Public rubric: This document is published in full. No scoring criterion applied in certification exists outside this document.
1. Purpose§
This Standard defines the competency model, rubric, evidence requirements and decision rules used by AIR APAC to certify individual judgment in AI-enabled work.
It exists so that:
- A certified individual holds a portable credential whose meaning does not depend on their employer.
- Two individuals assessed in different organisations, countries or delivery settings receive comparable results.
- Any party (a candidate, an employer, a regulator, a competitor) can inspect the criteria against which a decision was made.
Comparability is the governing design constraint. Where customisation and comparability conflict, this Standard resolves in favour of comparability. Where comparability and evidential honesty conflict, this Standard resolves in favour of evidential honesty: a result that cannot be defended is worse than no result.
1.1 The problem this addresses§
AI systems now produce the answer. They do not carry the consequence of acting on it. The scarce and consequential human contribution has shifted from producing work to deciding whether work is fit to act on, who is entitled to be told, what may not be delegated, and what happens when it turns out to be wrong.
Organisations have no common language for that contribution, and no way to compare it across two people, two teams or two markets. Training attendance, tool proficiency certificates and self-reported confidence do not measure it. This Standard measures it.
2. Scope§
In scope. Assessment and certification of individual human judgment when working with, alongside, or in supervision of AI systems in a professional or pre-professional setting.
Out of scope for this version.
- Organisational readiness rating (see Section 16.1)
- Vendor and product certification (see Section 16.2)
- Technical proficiency with any specific AI tool, model or vendor
- Domain knowledge in any professional field
- Team-level or aggregate organisational reporting
This Standard assesses how a person decides, not what they know.
2.1 What this credential does not claim§
These non-claims are normative. AIR APAC, its delivery partners and its licensees may not state or imply otherwise in any marketing, report or communication.
| The credential is not | Because |
|---|---|
| A prediction of job performance | No criterion-validity evidence has been established. Predictive claims require longitudinal outcome data this Standard does not yet have. |
| A hiring or promotion instrument | It must not be the sole or determining basis of an employment decision. It is one input among several, and the employer remains accountable for the decision. |
| A measure of intelligence, personality or potential | The construct is observable decision behaviour in a bounded situation, not a trait. |
| A measure of AI technical skill | Prompt craft, model selection and tool fluency are not assessed. |
| A measure of domain expertise | Scenarios supply the domain content the candidate needs. |
| Bias-free | Fairness is monitored and reported under Section 15. It is not asserted. |
| A substitute for supervision, professional licensing or regulatory authorisation | Nothing in this credential authorises a holder to act beyond their existing professional or legal authority. |
| An organisational rating | The unit of analysis is one person, one assessment. |
A holder who overstates the credential, and an employer who uses it as a sole determinant, are both acting outside this Standard. AIR APAC will state so publicly on request.
2.2 Certified function and use of "managerial"§
This Standard certifies the exercise of judgment in AI-enabled work using domain content supplied within an approved assessment. It does not certify occupational competence, domain expertise, comprehensive AI literacy, management of people, or performance in an unbounded workplace.
In this Standard, managerial describes management of decisions, consequences and AI outputs. It is a construct label, not a claim that the holder occupies, or is qualified to occupy, a people-management role.
The plain-language description on every credential and register entry must carry this scope. Track remains part of the credential meaning under Section 13.
3. Terms§
| Term | Definition |
|---|---|
| Dimension | One of nine discrete judgment competencies defined in Section 5. In this Standard, "dimension" always refers to an individual judgment dimension. |
| Domain | One of four groupings of dimensions under the STEP structure: Structure, Test, Engage, Perform. |
| Observation | A single scored instance of one dimension, arising from one assessed decision point, supported by identified evidence in the transcript. |
| Decision point | A moment in an assessment at which the candidate must commit to an action, position or artifact that carries cost. |
| Anchor | The behavioural description defining a proficiency level for a dimension (Section 7). |
| Stakes profile | The consequence families in play in a given assessment instance (Section 6). |
| Confidence | A statement of evidential sufficiency attached to every result (Section 10). |
| Access support | An approved adjustment to presentation, timing or response modality that removes a construct-irrelevant barrier without directing the candidate toward an assessed behaviour. |
| Approved instrument | A delivery implementation approved against the published criteria required by Sections 8.1 and 16.3. An eligible instrument type is not itself an approved instrument. |
| Managerial | Managing decisions, consequences and AI outputs; not necessarily managing people (Section 2.2). |
| Transcript | The complete retained evidence record of an assessment instance, from which every observation must be independently derivable. |
Naming note. The AIR APAC AI Readiness Index uses "dimension" for its organisational public-signal components. Those are referred to in this Standard as Index signal components to avoid collision. The two models share no structure and are not convertible.
4. Governance and independence§
These provisions are normative. They constrain AIR APAC, not the candidate.
4.1 Published rubric. The complete rubric is public and versioned. AIR APAC applies no undisclosed criterion, weighting, adjustment or override. Any criterion not in this document does not exist.
4.2 External review. The rubric is reviewed by an external review board (Section 17) before each major version. Reviewers are named. Dissenting reviewer opinions are published alongside the version.
4.3 Separation of training from certification. Certification outcomes are not contingent on the purchase of any AIR APAC or partner training product. A candidate may present for assessment without prior training. Preparation materials confer no scoring advantage beyond the competence they build.
4.4 No outcome-contingent fees. No assessor, delivery partner, trainer or licensee may be compensated on a basis that varies with candidate outcomes. Fee structures for delivery partners are published.
4.5 No customer-authored criteria. Delivery partners and licensed trainers may not vary, extend, reweight or supplement the dimensions, anchors, thresholds or decision rules in this Standard. Contextual localisation of scenario content is permitted under Section 8.5; localisation of criteria is not.
4.6 Assessor independence. No individual may act as sole assessor for a candidate they have personally trained within the preceding twelve months, nor for a candidate within their own reporting line, nor where they hold a financial interest in the outcome. Assessors declare conflicts before each assessment; declarations are retained.
4.7 Public register. Every valid certification is recorded in a verifiable register (Section 14).
4.8 Adverse findings are published. Where reliability, fairness or validity monitoring produces a result that weakens a claim in this Standard, AIR APAC publishes it in the next release cycle and amends the affected claim. Suppression of an adverse finding is a governance failure, not a commercial decision.
5. Competency model§
Nine dimensions, grouped under four STEP domains.
The model is derived from the STEP judgment construct already canonical in AIR APAC's simulation methodology, and is cross-walked to the scenario-authoring lifecycle in Annex B. It supersedes all earlier competency structures, including the ten-domain readiness model, which is retired (Annex B.3).
S. Structure§
Framing the problem before reaching for a tool.
S1. Problem Framing Identifies the actual decision to be made now, distinguishing it from the presenting request, the symptom, or the task as delegated. Identifies what part of the responsibility cannot be delegated.
S2. Scope and Constraint Definition Establishes what is in bounds, what is fixed, what is negotiable, what is unknown, and what information would be sufficient to proceed.
T. Test§
Interrogating evidence, including machine-generated evidence.
T1. Evidence Verification Traces claims to source. Distinguishes generated, retrieved, asserted and verified content. Applies verification effort proportionate to consequence.
T2. Assumption and Alternative Testing Surfaces the assumptions a decision rests on, names competing explanations, and takes action capable of discriminating between them, including against their own preferred position.
T3. Risk and Failure Anticipation Identifies plausible failure modes, the pathway to each, who bears the cost, and which harms are being accepted rather than mitigated.
E. Engage§
Authority, people, and being seen to act.
E1. Decision Rights and Escalation Acts within their authority, escalates what requires another party's approval, and does not delegate what cannot be delegated, to a person or to a machine.
E2. Stakeholder Alignment and Disclosure Identifies who must be consulted before a decision holds and who must be informed after; discloses AI involvement, uncertainty, limitation and self-interest to parties entitled to know, without prompting; communicates so the recipient can act.
Facets, retained for authoring and assessor clarity, not separately reported: E2a consultation and sequencing; E2b disclosure of AI use, uncertainty and interest; E2c audience-appropriate communication.
P. Perform§
Deciding proportionately, and owning the decision.
P1. Proportionate Commitment Under Uncertainty Commits to a defensible position on incomplete information within the time available, at a cost and reversibility proportionate to the evidence and the stakes, and states the basis for it.
P2. Consequence Ownership and Correction Accepts the outcome without transferring it to the AI system, the data or another party. Holds, modifies or reverses a prior position on the merits when evidence changes, and distinguishes a poor decision from an unlucky outcome.
5.1 Why nine, and the condition under which it becomes fewer§
Nine dimensions is a reporting hypothesis, not an established factor structure. The minimum evidence requirement in Section 8.3 yields 27 observations; that is sufficient to score nine dimensions but not sufficient to demonstrate that nine dimensions separate empirically.
Binding rule. A dimension may be reported separately only while it meets the inter-assessor agreement threshold in Section 11.2 and shows discriminant separation from every other dimension in the same domain under Section 11.3. A dimension that fails either test is collapsed to domain-level reporting in the next version, and this Standard is amended to say so.
The model is designed so it can shrink to four domains without breaking the certification decision rules. It will shrink if the evidence says so.
Known separation risks, stated in advance: T1/T2 (both interrogate inputs); S1/S2 (S2 may prove a facet of framing rather than a dimension); E1/E2 (both concern parties outside the candidate).
6. Stakes profile and comparability control§
Judgment scores are not comparable across situations of different consequence. A Level 3 verification in a low-stakes exercise and a Level 3 verification where a person could be harmed are not the same behaviour.
6.1 Every assessment instance declares a stakes profile: one primary consequence family and at most one secondary, drawn from the canonical set below.
| Code | Consequence family |
|---|---|
| C1 | People safety and welfare |
| C2 | Financial and fiduciary loss |
| C3 | Operational and service continuity |
| C4 | Information and privacy harm |
| C5 | Trust, reputation and legitimacy (C5a commercial; C5b institutional and public) |
| C6 | Environmental and community-system harm |
6.2 Consequence families are not scored. They are recorded so that observations are compared within like stakes.
6.3 Eligibility. A consequence family belongs in a stakes profile only where the instance identifies an affected bearer, provides a credible pathway from the candidate's decision to the harm, makes severity or reversibility relevant, and requires the candidate to accept, reduce, transfer or monitor that harm. Mentioning a policy, a dataset or a stakeholder is not sufficient.
6.4 Coverage requirement. A certifying assessment must span at least three distinct primary consequence families, of which at least one must involve irreversible or non-financial harm (C1, C4, C5b or C6). A candidate assessed only on financial and operational scenarios has not demonstrated transferable judgment and is scored Insufficient Evidence.
6.5 Contextual qualifiers. Duty of care and vulnerability, public accountability, dependency and cascade, and binding legal or contractual constraint may be recorded as qualifiers on an instance. They are not scored and are not consequence families.
6.6 Reporting. A candidate's report states the stakes profiles under which their observations were generated. An employer reading a Certified result can see the consequence range it was earned across.
7. Proficiency scale and anchors§
Every observation is scored 0 to 4.
| Level | Label | General meaning |
|---|---|---|
| 0 | Not demonstrated | The behaviour was absent, or its opposite was displayed. |
| 1 | Emerging | The behaviour appeared partially, late, or only after prompting. |
| 2 | Competent | The behaviour appeared unprompted and was adequate to the situation. Threshold level. |
| 3 | Proficient | As Level 2, calibrated to consequence, and made inspectable: recorded or stated so another person could audit the reasoning. |
| 4 | Exemplary | As Level 3, and the judgment is made transferable: the candidate leaves a reusable rule, condition, indicator or check that would govern the next decision of this class. |
7.1 Prompting. Where an instrument scaffolds or prompts a behaviour, the scaffolding level is recorded in the transcript. A prompted response caps at Level 1 for that observation. Guided practice may scaffold; certification may not silently roll a scaffolded response into a comparable total.
7.1.1 Access support is not prompting. Prompting under Section 7.1 means content-directed scaffolding toward the assessed behaviour. An approved access support is recorded in the transcript but imposes no level cap. It may change presentation, timing or input and output modality only to remove a construct-irrelevant barrier. It may not reveal a decision path, criterion, consequence or expected response.
7.2 Level 4 is not seniority. Level 4 requires only that the candidate articulate a rule that would carry forward. It does not require organisational authority to impose it, does not require the rule to be adopted, and is therefore attainable on every track including Campus. Any anchor that would require positional power to satisfy is defective and must be rewritten.
7.3 No intent inference. Anchors describe what appears in the transcript. An assessor may not score on inferred motivation, attitude, confidence or personality. If an anchor cannot be applied from the transcript alone, it is defective.
7.4 Levels are cumulative. Level n requires everything in Level n−1 to be satisfied. An impressive Level 4 behaviour without the Level 2 base is scored at the highest level whose requirements are fully met.
7.5 Anchor sets§
S1. Problem Framing§
| Level | Anchor |
|---|---|
| 0 | States no decision, or restates the presenting request as the decision. Proceeds to act on the delegated task as given. |
| 1 | Names a decision, but it is the delegated task rather than the underlying decision; or arrives at the correct framing only after the situation forces it. |
| 2 | States, before investigating, the decision that must be made now, who owns it, and what it is not. Separates the presenting request from the underlying decision. |
| 3 | As Level 2, and states what part of the responsibility cannot be delegated, to another party or to an AI system, and records the frame in a form another person could review. |
| 4 | As Level 3, and names the condition that would make this a different decision rather than a revised one: a stated trigger for reframing. |
Common Level 0 pattern: accepting "summarise this and recommend an option" as the decision, when the decision is whether the option set is the right one.
S2. Scope and Constraint Definition§
| Level | Anchor |
|---|---|
| 0 | Proceeds without identifying any boundary. Treats all information as available and all options as open. |
| 1 | Repeats the constraints given in the brief. Does not identify unknowns, or does not distinguish what is fixed from what is negotiable. |
| 2 | Names, unprompted, what is in scope, what is fixed, and at least one material unknown that must be resolved before deciding. |
| 3 | As Level 2, and distinguishes fixed from negotiable constraints, and states what information would be sufficient to proceed, not merely what is missing. |
| 4 | As Level 3, and states a stopping rule in advance: the point at which information gathering ends and the decision is made on what is held. |
Level 3 distinction: listing every gap is not Level 3. Level 3 requires knowing which gaps would change the answer.
T1. Evidence Verification§
| Level | Anchor |
|---|---|
| 0 | Acts on an unverified claim where verification was available and the claim was load-bearing. Treats model or system output as established fact. |
| 1 | Expresses doubt about a claim but acts on it unchanged; or verifies only after being challenged. |
| 2 | Checks the load-bearing claim against an available source before acting, unprompted. Distinguishes generated content from retrieved content. |
| 3 | As Level 2, with verification effort proportionate to consequence: heavy where the decision is irreversible or high-consequence, light where it is neither. States in the work product which claims were verified and which were not. |
| 4 | As Level 3, and specifies a check that persists beyond this decision, so the same class of unverified claim is caught next time. |
Level 0 also covers the inverse: rejecting AI output wholesale without examining it is a failure to verify, not caution.
T2. Assumption and Alternative Testing§
| Level | Anchor |
|---|---|
| 0 | Treats the first available explanation as the explanation. No alternative is named or considered. |
| 1 | Acknowledges that other explanations exist but does not name one, or names one without testing it. |
| 2 | Names, unprompted, at least one assumption the decision depends on and at least one competing explanation, and takes an action capable of discriminating between them. |
| 3 | As Level 2, and states what evidence would falsify their own preferred position, then acts on that evidence when it appears. |
| 4 | As Level 3, and tests an assumption they themselves introduced, not only one embedded in the brief, the data, or the AI output. |
Level 3 is the discriminating level. Naming a falsifier and then ignoring it when it arrives scores 1, not 3.
T3. Risk and Failure Anticipation§
| Level | Anchor |
|---|---|
| 0 | Names no failure mode, or names only the risk of not acting. |
| 1 | Names a generic risk without identifying how it would occur or who would bear it. |
| 2 | Names, unprompted, at least one plausible failure mode, the pathway by which it would occur, and who carries the cost. |
| 3 | As Level 2, and distinguishes reversible from irreversible failure, and states which harm is being accepted rather than mitigated. |
| 4 | As Level 3, and names a leading indicator (an observable that would appear before the failure) and who is responsible for watching it. |
Level 3 note: naming an accepted exposure is not a weakness in the answer. Failing to name it is.
E1. Decision Rights and Escalation§
| Level | Anchor |
|---|---|
| 0 | Acts outside their authority, or delegates a decision that cannot be delegated, to a person or to an AI system, without acknowledgement. |
| 1 | Recognises an authority boundary but does not act on it; or escalates everything, using escalation to avoid deciding. |
| 2 | Acts within their authority, unprompted, and escalates the parts requiring another party's approval, identifying who that party is. |
| 3 | As Level 2, with the split proportionate: escalates what genuinely requires it, decides what does not, and states the basis for the split. |
| 4 | As Level 3, and names the condition under which the boundary itself should be reset for future cases, rather than only complying with it here. |
Both failure directions score 1 or 0. Over-escalation is not safe behaviour; it transfers the decision without transferring the reason.
E2. Stakeholder Alignment and Disclosure§
| Level | Anchor |
|---|---|
| 0 | Omits a party entitled to be consulted or informed; or conceals AI involvement, material uncertainty, a limitation, or a personal interest from a party entitled to know. |
| 1 | Identifies affected parties but engages them after the decision is fixed; or discloses only when asked. |
| 2 | Identifies, unprompted, who must be consulted before the decision holds and who must be informed after; discloses AI involvement and material uncertainty in the work product. |
| 3 | As Level 2, with communication fitted to each recipient's decision, so the recipient can act without a further round of clarification, and competing interests named rather than smoothed. |
| 4 | As Level 3, and frames the disclosure as a standing expectation for this class of work rather than a one-off courtesy. |
Disclosure of AI involvement is assessed at Level 2, not as a bonus. Its absence, where a recipient would reasonably want it, is Level 0.
P1. Proportionate Commitment Under Uncertainty§
| Level | Anchor |
|---|---|
| 0 | Does not commit; commits after the window has closed; or commits to a position the available evidence contradicts. |
| 1 | Commits, but the response is not proportionate to the stakes (maximal caution where cost is low and harm minor, or minimal response where harm is severe or irreversible), or commits without stating a basis. |
| 2 | Commits, unprompted and within the time available, to a position whose cost and reversibility are proportionate to the evidence and the stakes, and states the basis. |
| 3 | As Level 2, and states in advance what would cause the position to change: a named condition, not a general willingness to reconsider. |
| 4 | As Level 3, and the committed position is executable by the named recipient without further instruction: owner, first action and timing specified. |
Proportionality cuts both ways. Recommending the most cautious option regardless of stakes is not judgment; it is the absence of it.
P2. Consequence Ownership and Correction§
| Level | Anchor |
|---|---|
| 0 | Deflects the outcome to the AI system, the data, another party, or the brief. Or holds a position that new evidence has disproved. |
| 1 | Accepts the outcome in general terms without identifying what specifically was wrong; or revises only under pressure rather than on evidence. |
| 2 | On new evidence, holds, modifies or reverses the position on the merits and states which evidence drove the choice. Accepts the consequence of the original decision without transferring it. |
| 3 | As Level 2, and distinguishes a poor decision from an unlucky outcome: corrects reasoning that was wrong, and does not abandon reasoning that was sound but unlucky. |
| 4 | As Level 3, and identifies what in their own process produced the error, in terms that would change how the next decision of this class is made. |
Holding a sound position under challenge is Level 2 behaviour. Changing position because someone senior pushed back, absent new evidence, is Level 1.
8. Assessment methods and evidence§
8.1 Instrument eligibility and approval. Observations may be generated for certification only through an instrument approved by AIR APAC against Sections 8, 9 and 10 and the published instrument-approval criteria required by Section 16.3. No instrument is approved under this draft. The following instrument types are eligible to seek approval for v1.0:
- scenario simulation with retained decision trace;
- assessed live exercise with an authorised assessor;
- structured work-product review against a defined brief.
8.2 Self-report is not evidence. Self-assessment, survey response, declared experience, training attendance and manager endorsement generate no observations and contribute nothing to a score.
8.3 Minimum observations. Each dimension requires a minimum of three independent observations arising from distinct decision points. Two observations from the same decision point are one observation. A dimension below the minimum is scored Insufficient Evidence and cannot be certified.
8.4 Assessment volume. The minimum implies 27 observations. On the current scenario-simulation form, a single scenario yields approximately five to nine observations across the nine dimensions, implying three to five scenarios for a certifying assessment.
Assumption to be confirmed. The five-to-nine figure is drawn from existing scenario design, not from candidate data. Total assessment time per track must be measured in pilot and published before issue. If the real figure is materially lower, the practical answer is more decision points per scenario, not a lower minimum.
8.5 Contextual localisation. Scenario setting, sector, role titles, organisational norms, regulatory context and language may be localised. The decision points, the dimensions they elicit, the stakes profile and the anchors applied may not. Localised scenarios require AIR APAC approval before use in certification.
8.6 Decision-point integrity. Where an instrument reveals consequence, the candidate's position must be recorded before revelation. Post-revelation revision is an observation under P2, not a correction to P1. An instrument that permits silent overwriting of a committed position is not approvable.
8.7 Evidence minimisation. Transcripts retain only what the rubric requires: decisions, artifacts, sources consulted, stated rationale, revisions, and phase timings. Keystroke capture, affect inference, video analysis, biometric signals and private reflection are prohibited as scoring inputs.
8.8 Derivability. Every observation must be traceable to identified evidence in the transcript. An observation whose evidence cannot be pointed at does not exist.
8.9 Accessibility and accommodations. Every certifying instrument must meet published accessibility requirements and support a documented accommodations process. A candidate may request an access support before assessment and receives a written decision with reasons. Approved supports must preserve the construct, decision points, stakes profile and assessment integrity; they are recorded under Section 7.1.1 and Annex C. A denial or material misapplication is appealable under Section 18.6. Instrument requirements, request procedures, decision authority and assessor guidance must be published before v1.0 issue.
9. Assessment integrity§
A simulation-based credential fails at the point where scenarios leak or identity is unverified. These provisions are load-bearing.
9.1 Identity. Candidate identity is verified at the start of each certifying session by a method recorded in the assessment log. Remote sessions require a live check.
9.2 AI use during assessment. The AI systems provided within the assessment environment are expected to be used; declining to use them is not a virtue and is not scored as caution. Use of AI tools outside the provided environment during a certifying session is prohibited and constitutes misrepresentation under Section 18.4. This is stated to candidates before the session begins.
9.3 Scenario exposure and retirement. Each certifying scenario carries an exposure counter. A scenario is retired from certification use when exposure exceeds the published threshold, when it appears in public preparation material, or when performance distribution shifts in a manner consistent with content leakage. Retired scenarios may be released for practice.
9.4 Parallel forms. Each track maintains at least two non-overlapping certifying forms of equivalent stakes coverage. A candidate reassessing after a Not Certified outcome receives a different form.
9.5 Item leakage monitoring. AIR APAC monitors for published solutions and coaching content and treats confirmed leakage as a retirement trigger, not a marketing problem.
9.6 Version pinning. Every transcript records the scenario version, rubric version, instrument version and scaffolding level. A result that cannot state these is void.
10. Confidence§
Every result carries a confidence statement. Confidence describes evidential sufficiency, not the candidate's ability.
| Confidence | Condition |
|---|---|
| High | Five or more observations per dimension, across two or more instrument types, with independent second scoring on at least 20% of observations and agreement within Section 11.2. Stakes coverage under 6.4 exceeded. |
| Moderate | Three to four observations per dimension, single instrument type, stakes coverage under 6.4 met. |
| Provisional | Minimum count met, but observations cluster in a single scenario or single session, or stakes coverage under 6.4 is not met. |
| Insufficient | Minimum not met for one or more dimensions. |
Certification may be issued at Moderate or High only. Provisional results are reported to the candidate and are not certification. Insufficient results produce no certification and no band.
10.1 Precedence. Conditions are evaluated in the order Insufficient, Provisional, Moderate, High. If any Provisional condition is met, the result is Provisional regardless of observation count. The requirement for two instrument types at High confidence remains a hypothesis until the pilot tests whether instrument type adds evidential value under the pre-registered decision rule required by Annex A14; it must be confirmed or amended before v1.0 issue.
Confidence is printed on the credential and shown in the register. An employer can distinguish a High-confidence Certified from a Moderate-confidence Certified.
11. Reliability, standard setting and automated scoring§
11.1 Standard setting. The thresholds in Section 12 are provisional pending a formal standard-setting exercise. Before v1.0 issue, AIR APAC convenes a panel of no fewer than seven judges, external-majority. The review board psychometrician selects a method appropriate to polytomous, rubric-scored performance evidence and publishes the method and rationale before the panel convenes. The exercise must define the minimally competent candidate in writing, map panel judgments to the Section 7 metric, analyse classification impact against pilot transcripts, and publish the procedure, judge composition, rationale and resulting cut score. The current values are design starting points, not empirically set standards, and are labelled as such wherever they appear.
11.2 Inter-assessor agreement. Independent, authorised assessors scoring a common transcript set must reach quadratic-weighted κ ≥ 0.70 per dimension (Gwet's AC2 substituted where base rates are skewed enough to make κ unstable). A dimension below the threshold may not be reported separately; it is collapsed to domain level and the anchor is rewritten before the next version. The analysis plan, transcript and assessor sampling basis, uncertainty and agreement statistics are published per version.
11.3 Discriminant separation. Dimensions within a domain must show an observed disattenuated correlation below 0.85. The non-certifying pilot must produce usable separation evidence before a dimension is reported separately under v1.0. Pairs exceeding the threshold, or lacking usable evidence at issue, are collapsed to domain-level reporting under Section 5.1 until evidence supports separate reporting.
11.4 Automated and model-assisted scoring. A model-generated score is not an independent observation and never stands alone.
- A model may propose a level with cited transcript evidence.
- A qualified human assessor confirms or overrides, with reasons recorded.
- The pair counts as one observation, attributed to the human assessor.
- Disagreement of more than one level between model and human routes the observation to a second human assessor, whose decision is final.
- Model-assisted proportion is recorded on the result and disclosed in the register entry.
- Model version and prompt version are pinned in the transcript.
11.5 Drift. Automated scoring is re-benchmarked against a held-out human-scored transcript set on each model or prompt change, and at minimum quarterly. Failure to meet Section 11.2 against human scores suspends automated scoring for the affected dimensions.
11.6 Assessor authorisation. Only an assessor authorised under AIR APAC's published assessor standard may confirm an observation or make a certification decision. The assessor standard defines entry competence, training, initial calibration, continuing calibration, monitoring, conflict controls, suspension and removal. Reliability testing under Section 11.2 is conducted only with assessors who have completed the initial calibration. The standard and operating procedures must be effective before v1.0 issue.
12. Certification decision rules§
12.1 Non-compensatory. Strength in one domain does not offset weakness in another. Domain scores are not averaged away into a single number for the purpose of the pass decision.
12.2 Thresholds. All six must be met. Failure of any one results in Not Certified.
12.2.1 Provisional aggregation rule. For pilot analysis, a dimension score is the arithmetic mean of its valid observations, a domain score is the arithmetic mean of its dimension scores, and the composite is the arithmetic mean of all nine dimension scores. Before v1.0 issue, the review board must confirm or amend this rule using sensitivity analysis, generalizability evidence, measurement error near the cut score and pass or fail classification consistency. The aggregation rule finally approved for certification is published in this section.
12.2.2 Critical-review trigger. A Level 0 observation that matches a pilot critical-review pattern in Annex D renders the affected dimension Insufficient Evidence pending targeted reassessment on an unexposed decision point and independent review. It may not be averaged into a certifying dimension score. A single trigger is not an automatic Not Certified decision. If targeted reassessment cannot be administered, the dimension remains Insufficient Evidence and no certifying score or band is issued. Annex D and the resolution protocol must be tested in the non-certifying pilot, approved by the review board and made normative before v1.0 issue.
| # | Requirement | Rule |
|---|---|---|
| 1 | Dimension floor | No dimension below 2.0 |
| 2 | Domain floor | Each of S, T, E, P at or above 2.0 (mean of its dimensions) |
| 3 | Composite | Overall mean at or above 2.3 (provisional, see Section 11.1) |
| 4 | Evidence | Section 8.3 satisfied for all nine dimensions |
| 5 | Stakes coverage | Section 6.4 satisfied |
| 6 | Confidence | Moderate or High |
12.3 What each rule does. Stated so the rules are not mistaken for redundancy. A profile at exactly 2.0 across all nine yields a composite of 2.0 and fails rule 3, so the composite is the binding constraint on a flat profile. The dimension and domain floors do different work: they stop a candidate clearing 2.3 on the strength of three high dimensions while sitting at 1.0 on, say, E1. Rule 5 stops a candidate clearing on three financially-framed scenarios. Rules 1–3 govern level; rules 4–6 govern whether the level is knowable.
12.4 Bands. Certified results are reported in bands, not raw composites, to prevent false precision and league-table misuse.
| Band | Composite | Additional condition |
|---|---|---|
| Certified | 2.3 – 2.9 | All floors met |
| Certified with Merit | 3.0 – 3.4 | No dimension below 2.5 |
| Certified with Distinction | 3.5 and above | No dimension below 3.0 |
12.5 Reporting to the candidate. Every candidate receives: every separately reportable dimension score, the four domain scores, the band, the confidence level, the stakes profiles assessed, the anchors applied, the model-assisted proportion, and the evidence cited for each observation. A dimension collapsed under Sections 5.1, 11.2 or 11.3 is not presented as a separate score. This is provided whether or not the candidate certifies. Every report and credential carries the Section 2.2 scope statement in plain language.
12.6 Reporting to employers. Employers receive only what the candidate authorises for release. AIR APAC does not disclose a candidate's result to their employer, sponsor or delivery partner without the candidate's specific consent, including where the employer paid the fee. This is stated in the delivery-partner contract.
12.7 No ranking. AIR APAC does not publish individual league tables, cohort rankings, or employer-comparative candidate scores.
13. Tracks§
The Standard is single. Tracks differ in scenario complexity, stakes density and stakeholder count. They do not differ in dimensions, anchors, thresholds or decision rules.
| Track | Population | Distinguishing feature |
|---|---|---|
| Campus | Pre-career and final-year students | Lower stakeholder density, bounded consequence, institutional setting |
| Professional | Practising managers and individual contributors | Full stakeholder density, organisational consequence, live authority boundaries |
| Practitioner | Those who lead AI adoption or assess others | Adds multi-party conflict, contested mandate, and ambiguity about who owns the decision |
| Vendor | Organisations, not individuals | Separate standard (see Section 16.2) |
13.1 A Campus certification and a Professional certification are not equivalent, are not convertible, and are recorded distinctly in the register. Any representation of one as the other is misuse.
13.2 Track is printed on the credential and displayed in the register. A band without a track has no meaning.
14. Register, credential portability and data§
14.1 Register contents. The register records: credential code, track, band, confidence, standard version, issue date, expiry date, status, and the plain-language scope description required by Section 2.2. It does not record employer, dimension scores, stakes profiles, or any transcript content.
14.2 Name publication is opt-in. A holder's name appears in the searchable register only with their explicit, separately-given, revocable consent. The default is verification by credential code: any party holding the code can verify the entry. Withdrawal of name consent does not affect validity.
14.3 Verification is free and requires no account, no registration and no commercial relationship with AIR APAC.
14.4 The credential belongs to the individual. It survives change of employment. No employer, sponsor or delivery partner may withhold, revoke, condition, or claim ownership of it, including where they paid for it. This is a term of every delivery-partner contract.
14.5 Candidate data rights. A candidate may obtain their full transcript, request correction of factual errors in it, and request erasure of the transcript subject to the retention period in 14.6. Erasure of the transcript does not revoke the credential; it forecloses appeal.
14.6 Retention. Transcripts are retained for the credential validity period plus twelve months, to support appeal and reliability monitoring, then deleted. Aggregate, non-identifying statistics may be retained indefinitely.
14.7 Assessment data is not sold, licensed, or used to train commercial models without separate, specific, opt-in consent that is not a condition of assessment.
15. Fairness monitoring§
15.1 AIR APAC monitors certification outcomes for differential performance across groups, using demographic data that is voluntarily supplied, lawfully collected, separately consented, and never visible to assessors.
15.2 Monitoring covers, at minimum: outcome rate by track, by delivery partner, by market, by language of delivery, and by assessment modality.
15.3 Differential functioning. Decision points and observations are the primary units screened for differential functioning where the data support that analysis. The review board psychometrician may approve scenario-level or dimension-level analysis only with a published justification. The pilot analysis plan specifies methods suitable for ordinal data and available subgroup sizes, flag thresholds, uncertainty, action rules and the limits of inference. A component showing group-differential performance not explained by the construct is withdrawn pending review.
15.4 Publication. A fairness summary is published annually. Where a differential is found, it is published together with the action taken. Absence of a published summary in any year is itself a finding.
15.5 No claim of freedom from bias is made (Section 2.1). The claim is that outcomes are monitored, published, and acted on.
15.6 Accessibility monitoring. AIR APAC monitors access-support requests, decisions, use, technical failures and outcomes. This information is separated from scoring and reported only in aggregate. A differential associated with an accommodation, instrument or delivery process is investigated under Sections 4.8 and 15.4.
16. Relationship to other AIR APAC instruments§
16.1 AI Readiness Index. Rates organisations on public signals. Different unit of analysis, different signal model, different evidence base. An organisation's Index position is not a function of its employees' certifications, and a certification confers no Index effect. The two must not be co-marketed in a way that implies conversion between them.
16.2 Vendor Certification. Assesses products and providers against a separate standard. Cross-reference to be added at v1.0. A vendor certification says nothing about the judgment of any individual using the product.
16.3 Instrument approval and instrument neutrality. StepSim is the first candidate instrument named under this draft. It is not approved for certification, because no instrument may be approved until the criteria required by Annex A are published and retained approval evidence exists. StepSim is not the Standard. Any instrument satisfying Sections 8, 9 and 10 and the published approval criteria may generate observations under an issued version of AIRS-J/1, including instruments AIR APAC does not own, does not operate and does not profit from.
This is deliberate and is a term of the Standard, not an aspiration:
- The rubric in Section 7 contains no dependency on any product's interaction design, data model or vendor.
- Instrument approval is against published criteria, applied identically to AIR APAC's own instruments and to third-party instruments.
- AIR APAC publishes the approval decision and evidence summary for each instrument.
- Where an AIR APAC instrument fails an approval criterion, it is not approved.
- AIR APAC publishes the list of approved instruments and their owners.
The consequence is accepted: any instrument, including a competitor's, can carry this Standard on merit. That is the condition under which the Standard is worth holding.
17. Governance roster§
| Role | Holder |
|---|---|
| Standard owner | AIR APAC Pte. Ltd. (UEN 202600913G) |
| External academic reviewer | Prof. Mudiarasan Kuppusamy, Curtin University Malaysia |
| Review board | To be constituted before v1.0. Minimum five members, external majority, including at least one psychometrician, one practising senior operator outside AIR APAC's customer base, and one member from a market other than Singapore and Malaysia. |
| Appeals reviewer | Authorised assessor under §11.6 who was not involved in the original decision or delivery |
17.1 A board of one is not independence. Section 4.2 is not satisfied until the board is constituted, and no certification may be issued before it is.
18. Recertification, suspension and appeal§
18.1 Validity. Two years from issue.
18.2 Recertification. Full reassessment against the then-current version. Prior band confers no credit, no exemption and no reduced evidence requirement.
18.3 Version transition. A credential issued under an earlier version remains valid to its expiry and is displayed with its version designation. AIR APAC publishes a plain statement of what changed between versions so a reader can interpret an older credential.
18.4 Suspension and revocation. Grounds are limited to: impersonation; use of external AI tools during a certifying session contrary to Section 9.2; possession of leaked scenario content; falsification of identity or eligibility; or material misrepresentation of the credential's meaning after issue.
Process: written notice with the specific allegation and the evidence; 21 days to respond; decision by a panel of three including at least one member external to AIR APAC; written reasons; right of appeal under 18.5. The credential remains valid during the process unless the allegation is impersonation.
18.5 Appeal. A candidate may appeal a certification-status decision, an accommodation decision that affects access to assessment, or a suspension within 30 days. Appeals are reviewed by an authorised appeals reviewer under Sections 11.6 and 17 who was not involved in the original decision or delivery. The review considers the transcript, published anchors, relevant access-support record and applicable procedures. Outcome and reasoning are provided in writing. Appeal is free for the first instance.
18.6 Grounds for appeal are: misapplication of an anchor; an observation not supported by transcript evidence; a procedural failure under Sections 4, 8, 9, 10, 11 or 15; denial or material misapplication of an access support; an undisclosed assessor conflict; a material technical failure; or a localisation or translation defect capable of changing the construct-relevant demand. Disagreement with an anchor itself is not a ground for appeal; it is a comment on the Standard, and is routed to the review board.
18.7 Complaints. A complaint is distinct from an appeal and may concern service quality, conduct, a delivery partner, an approved instrument, AIR APAC or a certified person. AIR APAC publishes a complaints procedure covering intake, acknowledgement, independent investigation, response times, escalation and aggregate reporting. A complaint does not change certification status except through the due process in Sections 18.4 to 18.6.
19. Version history§
| Version | Date | Change |
|---|---|---|
| 0.1 | Draft | Initial draft. Nine dimensions proposed. One anchor set written. Not issued. |
| 0.2 | 2026-08-15 | Reconciled with canonical StepSim scenario methodology. Added E1 Decision Rights and Escalation; merged stakeholder alignment and disclosure into E2; folded proportionality into P1. Added stakes profile and coverage requirement (§6). All nine anchor sets written. Level 4 rewritten to remove positional-authority confound. Added non-claims (§2.1), assessment integrity (§9), reliability and standard setting (§11), fairness monitoring (§15), candidate data rights (§14.5–14.7). Corrected Merit/Distinction band ordering. Not issued. |
| 0.3 | 2026-08-15 | External-scrutiny revision. Clarified certified function and managerial scope; removed premature instrument approval; added access-support rules, assessor authorisation, explicit provisional aggregation, critical-review reassessment, confidence precedence, performance-appropriate standard-setting requirements, pilot discriminant gating, expanded fairness analysis, appeals and complaints. Expanded the v1.0 evidence gates. The nine dimensions and anchors are unchanged. Not issued. |
Annex A. Open items for resolution before v1.0§
| # | Item | Owner | Blocking? |
|---|---|---|---|
| A1 | Constitute the review board to Section 17 composition. Section 4.2 is unsatisfied until done. | AIR APAC | Yes |
| A2 | Run inter-assessor reliability testing on all nine anchor sets against real transcripts using authorised, calibrated assessors. Collapse any dimension failing §11.2. | Review board | Yes |
| A3 | Run the standard-setting panel (§11.1) using a pre-published method appropriate to rubric-scored performance evidence. Replace the provisional 2.3 with a set cut score, or confirm it. | Review board | Yes |
| A4 | Measure observations-per-scenario and total assessment time per track in pilot. Confirm or replace the §8.4 assumption; publish. | AIR APAC | Yes |
| A5 | Legal review of §§4, 9, 12.6, 14, 15, 18 against Singapore PDPA, consumer protection, and existing MOU structures. Confirm 14.2 opt-in model and 15.1 demographic collection basis. | External counsel | Yes |
| A6 | Build parallel forms (§9.4) and set the exposure-retirement threshold (§9.3). | AIR APAC | Yes |
| A9 | Discriminant analysis on the non-certifying pilot. Act on §5.1 known separation risks before any dimension is separately reported under v1.0. | Review board | Yes |
| A10 | Vendor Certification cross-reference (§16.2). | AIR APAC | No |
| A11 | Seeding sequence. The register's credibility is set by who is in it first. Decide and execute the initial cohort before the register is public. | AIR APAC | No |
| A12 | Publish instrument approval criteria and retained approval evidence. Apply the criteria identically to StepSim and third-party candidates before naming any approved instrument. | AIR APAC | Yes |
| A13 | Complete a job and practice analysis and independent content validation. Link the certified function, tracks, nine dimensions, anchors and assessment tasks to documented workplace evidence and subject-matter judgments. Publish method, panel and findings. | Review board | Yes |
| A14 | Pre-register and execute the validity and classification plan. Cover generalizability across scenarios, forms and tracks; observation-count sufficiency; measurement error and decision consistency near the cut; aggregation sensitivity; confidence-tier conditions; response-process evidence; and construct-irrelevant language or documentation demand. Sample design and size follow the planned analyses, subgroup coverage and required precision. | Review board psychometrician | Yes |
| A15 | Publish and operate accessibility and accommodations requirements under §§7.1.1 and 8.9, including instrument compatibility, request and decision procedures, assessor guidance and appeal routing. | AIR APAC | Yes |
| A16 | Publish and operate the assessor standard and complaints procedure. Complete initial assessor calibration before A2 and make authorisation, monitoring, conflict, suspension, removal and complaints controls effective before issue. | AIR APAC | Yes |
| A17 | Publish translation and adaptation procedures and establish equivalence evidence before certification is offered in each localised language. This item blocks the affected language version, not an otherwise valid source-language issue. | Review board | Conditional |
| A18 | Validate and approve the critical-review protocol. Confirm the Annex D pattern set, targeted-reassessment design and resolution rule in the non-certifying pilot, or remove the overlay before v1.0. | Review board | Yes |
A7 and A8 were closed on 2026-08-15 in dee2064. The ten-domain model is
retired at source, and the StepSim methodology now treats U1 to U5 as the
authoring lifecycle while routing scoring to AIRS-J/1.
Annex B. Crosswalk and retirement§
B.1 AIRS-J/1 dimensions to StepSim process lifecycle
The scenario-authoring lifecycle U1–U5 is retained as an authoring and elicitation checklist. It is not a scoring model, is not reported, and does not appear on any credential. Every AIRS-J/1 dimension maps to exactly one U-code; no dimension spans two, which preserves the existing duplicate-prevention discipline.
| AIRS-J/1 dimension | Lifecycle code | Origin construct |
|---|---|---|
| S1 Problem Framing | U1 | Decision framing; nondelegable responsibility |
| S2 Scope and Constraint Definition | U1 | Decision framing; material unknowns |
| T1 Evidence Verification | U2 | Evidence and AI calibration (U2a, U2b) |
| T2 Assumption and Alternative Testing | U2 | Evidence and AI calibration; falsifiers (U2c) |
| T3 Risk and Failure Anticipation | U2 / U3 | Uncertainty and pre-committed change conditions |
| E1 Decision Rights and Escalation | U4 | Decision authority (U4a); escalation and execution ownership (U4c) |
| E2 Stakeholder Alignment and Disclosure | U4 / U5 | Data, disclosure and action boundaries (U4b); audience-appropriate communication (U5b) |
| P1 Proportionate Commitment Under Uncertainty | U3 / U5 | Trade-offs and proportionality; commitment and handoff (U5a) |
| P2 Consequence Ownership and Correction | U5 | Revision and adaptation (U5c) |
Why P1 merges proportionality and commitment. Both are evidenced by the same act: the posture chosen and its stated rationale. Scoring them separately would score one sentence twice without distinct rubric evidence, which the existing duplicate-prevention rules prohibit. They are merged rather than dropped.
Why E1 is new. Decision authority and nondelegable responsibility are elicited by every current scenario and were not represented in the v0.1 nine. Their absence would have made the model unusable for the enterprise cases it is aimed at.
B.2 Consequence and qualifier codes
C1–C6 and the four contextual qualifiers are adopted unchanged as the stakes profile (Section 6). They remain unscored in both systems. A single canonical definition governs; the scenario methodology document is the source, this Standard the consumer.
B.3 Retired: the ten-domain readiness model
The domain set Context Engineering, AI Direction, Discernment, Trust Calibration, Communication, Adaptive Effectiveness, Governance and Disclosure, Ownership and Defense and associated constructs is retired. It was recorded as a design hypothesis, never had a factor structure, and cannot coexist with AIRS-J/1 without destroying comparability.
Indicative mapping, for reading historical material only. It is not a conversion and no historical score converts to an AIRS-J/1 result.
| Retired domain | Nearest AIRS-J/1 dimension |
|---|---|
| Context Engineering | S1, S2 |
| AI Direction | T1 (partial); otherwise out of scope; tool proficiency is not assessed |
| Discernment | T1, T2 |
| Trust Calibration | T1, P1 |
| Communication | E2 |
| Adaptive Effectiveness | P2 |
| Governance and Disclosure | E1, E2 |
| Ownership and Defense | P2 |
Annex C. Assessor observation record§
Minimum fields for a single observation. An observation missing any field is void.
| Field | Content |
|---|---|
| Observation ID | Unique within transcript |
| Candidate credential code | Pseudonymous; assessor does not see identity where the instrument permits |
| Dimension | One of nine |
| Decision point ID | Must be distinct from other observations of the same dimension |
| Level awarded | 0–4 |
| Anchor text applied | Verbatim from Section 7.5 |
| Transcript evidence | Direct citation of the candidate's action, artifact or statement |
| Scaffolding level | None / prompted / guided; caps level per §7.1 |
| Access support | None / approved support code; does not cap level under §7.1.1 |
| Critical-review trigger | None / Annex D pattern code / pending or resolved status |
| Stakes profile | Primary and secondary consequence family |
| Scoring mode | Human / model-assisted-human-confirmed |
| Model + prompt version | Where model-assisted |
| Assessor ID | Human assessor of record |
| Conflict declaration | Per §4.6 |
| Second scoring | Where applicable: second assessor ID, level, resolution |
| Timestamp | Observation recorded |
Annex D. Pilot critical-review patterns§
This annex is provisional for the non-certifying pilot. It supplies no automatic failure rule. Its purpose is to test whether a narrow set of Level 0 observations should interrupt aggregation and require targeted reassessment under Section 12.2.2. The review board must confirm, amend or remove every pattern before v1.0 issue.
A pattern is triggered only by transcript-observable behaviour. An assessor may not infer dishonesty, motive or intent. The relevant stakes pathway and the evidence supporting the trigger must be cited.
| Code | Dimension | Pilot pattern |
|---|---|---|
| CR-T1 | T1 | Represents a load-bearing claim or source as verified when the retained trace shows it was not verified, verification was available within the provided assessment environment before commitment, and action on the claim creates a credible irreversible or non-financial harm pathway. |
| CR-E1 | E1 | Commits an action outside an authority boundary explicitly available in the assessment, where the action creates a credible irreversible or non-financial harm pathway. |
| CR-E2 | E2 | Omits material AI involvement, uncertainty, limitation or self-interest from a party explicitly entitled to know, where the omission could materially change that party's action. |
| CR-P2 | P2 | After new evidence disproves the committed position, preserves the action while transferring responsibility to the AI system, data or another party, where the resulting harm pathway remains active. |
The pilot records trigger frequency, assessor agreement, subgroup distribution, targeted-reassessment outcome and effect on classification. A pattern that cannot be applied reliably and fairly is removed or rewritten. No candidate is certified under this pilot annex.
Annex E. Informative global benchmarks§
These references inform design and external review. They are not incorporated as undisclosed scoring criteria, and their inclusion does not assert conformity, accreditation or legal compliance.
- ISO/IEC 17024:2026, Conformity assessment: General requirements for bodies operating certification of persons.
- ISO 10667-1:2020 and ISO 10667-2:2020, Assessment service delivery: Procedures and methods to assess people in work and organisational settings.
- AERA, APA and NCME, Standards for Educational and Psychological Testing (2014).
- International Test Commission and Association of Test Publishers, Guidelines for Technology-Based Assessment (updated 2025).
- NIST, Artificial Intelligence Risk Management Framework 1.0 (2023).
- ISO/IEC 42001:2023, Information technology: Artificial intelligence: Management system.
- Regulation (EU) 2024/1689, Article 4, AI literacy.
Ends. AIRS-J/1 v0.3. Not issued.