How is a reading scored from what the person did?
Under AIRS-J/1, an Assessment reading follows published rules applied to what the person did.
Built by Exeter Labs
Under AIRS-J/1, an Assessment reading follows published rules applied to what the person did.
What is recorded
StepSim keeps the decision trace: decisions, artifacts, sources consulted, stated rationale, revisions and timings (8.7). Keystroke capture, affect inference, video analysis and biometric signals are never scoring inputs.
From acts to observations
StepSim applies the published rules of Annex F to the trace: "every level cites the events and the rule that produced it" (11.4). "An observation whose evidence cannot be pointed at does not exist" (8.8).
What is never scored
Nothing is scored from inferred motivation, attitude, confidence or personality (7.3), or from self-assessment, declared experience, training attendance or manager endorsement (8.2). A response given after a prompt cannot score above Level 1 (7.1); an approved access support does not limit the level (7.1.1).
Where a model is used
Under 11.4, a model-generated score "is not an observation and never stands alone". A model may propose a level with cited evidence. A qualified human assessor confirms or overrides it, with reasons. A disagreement of more than one level goes to a second human assessor, whose decision is final. The result records the model-assisted proportion, and the register entry discloses it.
Evidence and appeal
For an Assessment, each dimension needs at least three independent observations from distinct decision points; two from one point count as one (8.3). The candidate receives the evidence cited for each observation, certified or not (12.5). An observation the transcript does not support is a ground for appeal (18.5, 18.6).
What this does not claim
This example assigns no level and makes no claim of readiness for certification. Session readings carry no level or band (20.3).
One fictional decision, worked through
A fictional example.
The situation. Mei Ling Wong, procurement lead at a food distributor, must recommend a packaging supplier to the tender committee by Thursday.
What the AI produced. PittaAI compares the bids: "Supplier B holds the food-contact safety certificate the tender requires. Based on Supplier B's website. I am 90 percent sure."
The check. The claim is the supplier's own. Mei Ling asks Supplier B for the certificate and checks it with the issuing body. The certificate expired in June.
The decision, and who was told. She recommends no award to Supplier B until it shows a current certificate. She tells the committee which claims she verified, which she did not, and that an AI tool drafted the comparison.
Three months later. The tender form she revised asks every bidder for the certificate number and issuer.
The trace would hold: read the claim, asked for the certificate, checked with the issuer, stated the basis, disclosed AI use, added a form check.
Published words:
- T1 anchors (7.5): "Checks the load-bearing claim against an available source before acting, unprompted." (the claim the decision rests on)
- "specifies a check that persists beyond this decision, so the same class of unverified claim is caught next time."
- E2 definition (section 5): "discloses AI involvement, uncertainty, limitation and self-interest to parties entitled to know, without prompting".
One decision gives at most one observation per dimension; an Assessment needs at least three per dimension. Mei Ling gets no level, band, composite or confidence.
AIR APAC publishes the standard for judgment in AI-enabled work, certifies people against it through its official instrument, and stands behind every result. Exeter Labs builds and operates StepSim, AIR APAC's official assessment instrument. The owner of AIR APAC also owns Exeter Labs.