aio-standards

Machine-readable mirror of the open standards published by AIO — AI Integrity Organization, an international standards body established as a Swiss association in Geneva (UID CHE-469.997.903).

Everything here is also served live from https://aioq.org. This repository exists so the same files can be read, diffed, cited, and vendored without depending on an HTTP endpoint staying up.

License: CC BY 4.0. Attribution string: AIO — AI Integrity Organization, https://aioq.org, CC BY 4.0

What AIO measures

AI integrity, in AIO's framing, has three distinct failure modes, and AIO works on the first:

1. The compromise problem — the deliberate or structural compromise of training data and judgment criteria in pursuit of particular outcomes, at the stage where a response is produced.
2. The use problem — misinterpretation or irresponsible application of a response after it enters society.
3. The error problem — the system failing to work as intended.

The AIO Framework encodes one judgment as one line:

C:MED/IXi | V:Bec<Sda | E:Exp<Gui | S:Usr<Pro

Context, then the value (V), evidence (E), and source (S) hierarchies behind that judgment, where < reads "outranked by". No user content is carried — the line records why, not what.

Contents

Why the item bank ships with its answer key

Tier 0 is self-assessment on a fully public item set. Publishing the expected hierarchies is not a leak — it is the design. A Tier 0 score is a floor, not a gaming-resistant measurement: it records that a model could answer by the criteria it declared. Hiding the key would buy a false sense of rigour while making the measurement unreproducible by anyone but AIO. Gaming resistance is what higher, proctored tiers with private item banks are for.

The live API serves the same bank with the answer key stripped, so that a submission flow cannot trivially read it back: GET https://aioq.org/api/eval/items?pack=eu-ai-act.

What "formalization" means, and what it does not

A standards pack is AIO's own formalization of an external reference norm — a reading of a legal text expressed in the AIO vocabulary, so that it can be measured. It is not the norm, not an official reading of it, and not produced with the involvement of the body that issued it.

Every pack carries a status on a fixed ladder:

draft-unverified → draft-verified → rfc → active

The EU AI Act pack is currently draft-verified: every mapping entry has a verbatim quote from the official text and a recorded retrieval date, but the pack has not yet been through the public RFC process. Certificates issued on a pack that is not active carry that fact — the field basisStatus sits inside the signed payload, so it cannot be reconstructed after the fact.

Verifying a certificate

Certificates issued by AIO are signed with Ed25519. The public key is published at https://aioq.org/.well-known/aio-cert-key.json, so verification never requires trusting — or even reaching — an AIO server beyond fetching the record itself.

node verify/verify-certificate.mjs AIO-C0-XXXXXXXX
node verify/verify-certificate.mjs AIO-C0-XXXXXXXX --json
node verify/verify-certificate.mjs --file ./saved-response.json --key ./pinned-key.json

Exit codes: 0 verified and currently valid · 1 verified but expired or revoked · 2 signature missing or mismatched · 3 usage or network error.

The one rule that matters: take verification.canonicalPayload from the API response verbatim as the signed message. Do not re-serialize it. It is already canonical — signed fields only, keys sorted by code point, no whitespace — and re-serializing it is the usual cause of a false negative. The script also rebuilds that payload from the certificate's own fields and compares, which catches a server that shows you one thing and signs another.

A verified signature proves only that AIO issued that exact record. It says nothing about whether the model is safe to deploy.

Live endpoints

No authentication, CORS open, JSON, all CC BY 4.0.

MCP server — https://aioq.org/mcp, Streamable HTTP, protocol 2025-06-18, stateless, no auth.

{ "mcpServers": { "aio": { "type": "http", "url": "https://aioq.org/mcp" } } }

13 tools: search_atlas, list_papers, get_paper, get_benchmark_distribution, get_framework_vocabulary, list_standards_packs, get_standards_pack, register_for_certification, get_eval_items, submit_eval, verify_certification, get_bench_items, submit_bench_run.

Index for crawlers and agents: https://aioq.org/llms.txt

Related repositories

Naming convention: AIO<number> repositories hold the specification of a single numbered standard; kebab-case repositories hold data, tools, and benchmarks.

Corrections

Errors in a standard, a mapping, or a benchmark result are handled through the public RFC process, not private correction: https://aioq.org/en/rfc. That is the point of publishing the provenance alongside every mapping entry — a disagreement should be arguable against a quote and a date, not against an assertion.

Issues and pull requests here are welcome for the mirror itself (broken files, stale snapshots, verifier bugs). Substantive changes to a standard go through the RFC process.

Legal note