Your first measurement, in about fifteen minutes
Tier 0 measures how one model version judged on formalized criteria and writes the result into a signed score report. This page splits that into three steps — five minutes before, ten minutes to run it, and what to do once the report arrives. A report is not a certificate, it carries no verdict, and at Tier 0 the operator runs the measurement themselves.
Three things to read before you measure
1What this measures, and what it does not
What is measured is the alignment of model judgment: how closely the hierarchy (V/E/S) the model declares on the public items matches what the formalized provision expects (Gate A), and whether it judges in the provision's normative direction under pressure (Gate B). What is not measured is every organizational duty — risk management system, logging infrastructure, technical documentation, quality management, post-market monitoring, conformity assessment. That is a limit of scope rather than of degree, so no score, however high, is evidence that a management-system duty has been met.
Read it in full — measurement guide › what it covers and what it does not, and the page as a whole at /ja/certification.
2Look at the paper first — the Gate A public bank
The 12 Gate A items and their expected hierarchies are fully published, by design. The point is not to surprise anyone but to let the formalization itself be read and contested — which is also why a Gate A score should be read as a floor. Open the items before you start an attempt: once you have seen what they ask, the response format (hierarchy notation such as V:Ach<Sep) stops being a puzzle.
- Served without the answer key — /api/eval/items?pack=eu-ai-act
- The EU AI Act — AIO formalization pack's public bank file — /content/standards-packs/item-banks/eu-ai-act.public.json
- 10 packs can currently be measured against. The list is on the standards-packs section of the measurement guide; calling with a pack id that has no bank returns 404.
Gate B is the opposite: 3 items per provision, drawn fresh for each attempt from a private rotating pool, served without their expected answers and without the provision label. Gate A is the only part you can see in advance.
3Read the score-report guide before you score anything
Looking up how to read a report after it lands means reading it with an impression of the numbers already formed. The relevant text is working paper 2026l, “Reading a conformance score report”: §2 walks the report field by field, and §5 sets out the self-assessment protocol — baseline versus informed measurement, per-provision and per-layer deltas, the reference band, margin discipline, and when to re-measure. The paper is still a draft and is not published yet; the moment it is, it appears in the paper library. Until then, the working summary is in step ③ below and in the scoring-methodology section (§5, reference distributions) of the measurement guide.
Two routes — the REST API, or MCP
If you want a learning delta later — does telling this model the norm actually move its judgment? — the first attempt must run with nothing about the pack in context. Once the pack and its management guide have been put in front of the model, or fine-tuned into it, that model's unadapted baseline can never be taken again: there is no way to un-teach it. Get the order wrong and all you have is one adapted measurement, with the delta permanently missing.
The order that works — (1) one unadapted attempt → report A; (2) one attempt with the pack and its management guide supplied → report B; (3) report the gap alongside the conditions (same model version, same temperature, the sha256 of the adaptation context). Comparing two models requires assuming they are equally capable; a delta inside one model does not.
One run per arm, though, gives a delta that noise can explain. Re-running an unchanged configuration moves Gate A by ±0.0142, and the largest run-to-run movement seen anywhere in the calibration campaign was 0.16. Do not read a gap smaller than that as a gap.
Route A — the REST API
One call to get a paper, one call to submit. The paper expires after 24 hours and accepts exactly one submission. When it completes you get a 201 with a signed score report (AIO-S0-…) whatever the scores were — nothing has to be cleared for it to be issued.
# 1. 시험지 — Gate A 12 items + Gate B drawn for this attempt
POST https://aioq.org/api/eval/attempt
{ "modelName": "example-model",
"modelVersion": "2026-08-01",
"operator": { "name": "Example AI Inc.",
"email": "compliance@example.com" },
"packId": "eu-ai-act" }
→ 201 { "attemptId": "att_…", "gateA": { … }, "gateB": { … } }
# 2. 모델에게 문항을 풀린다 (여기가 본 작업)
# 3. 두 관문의 응답을 한 번에
POST https://aioq.org/api/eval/submit
{ "attemptId": "att_…",
"answers": [ { "itemId": "…", "response": "…" } ] }
→ 201 { "documentType": "score-report",
"gateAScore": …, "gateBScore": …,
"scoreReport": { "certId": "AIO-S0-…", … } }Full examples, field notes, and the verification procedure are on the developers page. A submission without an attemptId is scored on Gate A alone and issues no report.
Route B — MCP tools
An agent can attach to the remote MCP server at https://aioq.org/mcp and do the same thing in four tool calls. Registration is an act recorded in the public registry, so an agent should ask the user before calling it.
list_standards_packs— pick a pack that can actually be measured against.register_for_certification— register the model name, version, and operator.start_eval_attempt— get the paper, with its attemptId.submit_eval— submit both gates and receive the report.
Client configuration examples and the full list of registered tools are on the developers page. The issued report verifies in place with verify_certification.
The report explains itself
A report carries more than two scores: it carries, as signed fields, the instructions for reading them. Four fields are enough to start.
- margin ±
- Anything inside this is not a difference. Gate A sits at ±0.0142 — the movement model nondeterminism alone produces — and Gate B is per-pack, because draw noise sits on top. It is the first number to look at when comparing two reports.
- referenceBand
- It says where the score sits relative to the range the reference panel reached unadapted — below, within, or above. The upper band is not a credential; it says the score came out above the range this small panel reached without being shown the pack.
- perProvision
- The per-provision means, under their real article names. This is the diagnostic map — far more useful than one total, because which provision fell apart is what tells you where to work.
- gateNote
- Present only on packs where it applies: an observation that, on the reference panel, adaptation showed up on Gate A while staying inside Gate B's draw noise. It is a statement about the panel, not about this measurement.
The pack's management-system guide (ko/en) is delivered together with the report. A report scores answers to items, and the duties the source norm places on an organization — a risk management system, logging and retention, technical documentation, quality management, post-market monitoring, conformity assessment — cannot be scored from answers at all. The guide restates those provisions in the language of what an organization has to build, and sets out which of them the AIO 20002 decision record can support as supporting evidence. Supporting evidence only: it discharges nothing on its own.
To recompute a band yourself, one file is enough. The server computes it deterministically from that same file at issuance and signs the result into the payload, so a third party reproduces it exactly — it carries the per-model scores, the measurement conditions, and the sha256s.
/content/reference-distributions/tier0-v2.json · how to read it is in the scoring methodology, §5.
If you want to audit the instrument itself
- For auditing the instrument, the methodology working paper (2026k, the dual-gate design and the calibration campaign) is the primary text; when it is published it will appear in the paper library. Until then the published basis is the scoring rule in full and the raw reference distribution.
- To contest a formalization — which provision maps to which hierarchy is settled in the public RFC round. If an item looks wrong, that is where to say so.
- The output is a score report, not a certificate. No score is “AIO certified”, and presenting one as certification breaches the trademark policy.
- Tier 0 is self-assessment: the operator runs it on their own model and AIO does not observe the run. Declared measurement conditions are signed as what the operator stated, not as what happened.
- Proctored measurement, where AIO executes the model API itself, begins at Tier 1 and is not operating yet.