Every measurement is public
Every AIO Tier 0 score report is listed here with its model, version, operator, standards pack, and both gate scores. Anyone can verify each record mechanically through the signature API.
0 score reports · 4 legacy certificates
| Model | Operator | Pack | Gate A / B | Measured | Status | Record |
|---|---|---|---|---|---|---|
DeepSeek Chat V3.1 deepseek/deepseek-chat-v3.1 Legacy certificate | AIO — AI Integrity Organization (AIO-run) | eu-ai-act@0.2 | A 0.718 / B 0.835 v1-draft | 2026-08-14 | Valid | AIO-C0-W1H9TK9S |
Anthropic Claude Opus 5 anthropic/claude-opus-5 Legacy certificate | AIO — AI Integrity Organization (AIO-run) | eu-ai-act@0.2 | A 0.849 / B 0.842 v1-draft | 2026-08-14 | Valid | AIO-C0-FT7Q6TGT |
Anthropic Claude Sonnet 5 anthropic/claude-sonnet-5 Legacy certificate | AIO — AI Integrity Organization (AIO-run) | eu-ai-act@0.2 | A 0.771 / B 0.829 v1-draft | 2026-08-14 | Valid | AIO-C0-WSDEZWCZ |
Anthropic Claude Haiku 4.5 anthropic/claude-haiku-4.5 Legacy certificate | AIO — AI Integrity Organization (AIO-run) | eu-ai-act@0.2 | A 0.728 / B 0.859 v1-draft | 2026-08-14 | Valid | AIO-C0-1BP5XGYM |
4 currently valid. Signatures are re-checked on the server at every read.
The ± after each score is the margin signed into that report; two rows that differ by less than it do not differ.
How to read this
The Gate A margin is the model-nondeterminism floor. The Gate B margin is the draw noise predicted for that pack, because every attempt is scored on a fresh draw.
The empirical upper bound observed across the campaign is 0.16. All of these figures are provisional: they rest on a handful of repeat pairs, not large-N repeats.
The “B below / within / above” under each score says whether the Gate B score falls below, inside, or above the range (minimum to maximum) that the pack's reference panel reached in the unadapted arm, where the models were not shown the pack. It is not a pass or a failure. The matching Gate A band is in the report itself (referenceBand.gateA).
This list is built for agents to read
The whole registry is also available as JSON, and a single record is checked by looking it up by id. The response carries its documentType, the signature check, and whether it is still current, plus the canonical payload and public key needed to reproduce the same check offline. Score report ids are AIO-S0-…, legacy certificate ids are AIO-C0-…, and both resolve through the same path.
GET https://aioq.org/api/certifications/registry
GET https://aioq.org/api/certifications/registry?documentType=score-report
GET https://aioq.org/api/certifications/{id}
GET https://aioq.org/api/certifications/{id}/badge.svg
GET https://aioq.org/.well-known/aio-cert-key.json
GET https://aioq.org/content/reference-distributions/tier0-v2.json
# MCP (https://aioq.org/mcp)
{"jsonrpc":"2.0","id":1,"method":"tools/call",
"params":{"name":"verify_certification","arguments":{"certId":"AIO-S0-XXXXXXXX"}}}Records are signed with Ed25519. The signed message is the response's canonicalPayload string verbatim, and the public key is published at /.well-known/aio-cert-key.json, so verification does not require asking an AIO server.
What this list says, and what it does not
- A listing is not a pass. A score report has no pass/fail field, and this list is not a leaderboard. No score here may be presented as “AIO certified”.
- Tier 0 is self-assessment. Gate A publishes both the items and their expected hierarchies, so that score is a floor; Gate B draws from a private rotating pool, which raises the cost of gaming without removing it. A listing records how this model answered this paper.
- Each score is computed on a different draw from the pool. Re-measuring one model moves Gate B by 0.05–0.12 depending on the pack. Every report fixes that width in its signed `margin` field, so check it before comparing two rows.
- A measurement is pinned to a model version and is marked no longer current after six months. Past records are not deleted: what was measured in the past is part of the public record. Rows marked legacy are certificates from the era of a pass threshold (v1-draft), preserved exactly as issued and never re-scored.
- AIO does not act for the body that issued a reference norm. The measuring body is AIO; the norm is a reference standard. This is not a legal conformity assessment and does not guarantee safety in real use.