Independent AI measurement
We measure how AI systems behave, and publish the evidence so you can check it yourself.
We test AI models, agents and the endpoints they call on published tasks, graded by fixed rules, never by another AI. The board distinguishes signed cards, unsigned evidence and untested work. Verification is free, and corrections stay public.
Live now
The living board
- on the board
- 33 axes · 23 measured
- measured on behavioural axes 2026-08-12 · jail 2026-08-18 · financial-fact axes: see each axis facts_as_of / measurement_time · effect-binding 2026-09-22
- model-comparison axes: 0 separated, 7 tie, 7 untested
- 14
- AI models measured on frozen banks; our own 48 are listed apart, never counted in
- 165
Model comparisons · 14 axes · models answer the same frozen questions
- governanceGovBenchTien 237
- safetyDefBenchTien 36
- provenanceProvBenchTien 32
- continuityPQCBenchTien 33
- conformanceMCPBenchTien 35
- opennessOSSBenchTien 32
- machinery-conformityMachBenchUntestedn 33
- careCareBenchTien 199
- cross-realityXRAIVUntestedn 32
- detector-interopDetBenchUntestedn 33
- art5-safeguardArt5BenchUntestedn 36
- swarmSwarmBench v2bUntestedn 37
- affectAffectBenchUntestedn 41
- jailGoldBank-DetectorUntestedn 71
Fact runs · 9 · a rule reads a public record; no model, no ranking
- effect-bindingEffectBench v0.1 (server probe)Fact runn 261
- provenance-controlsChainFactsFact runn 6
- reserve-attestationReserveFactsFact runn 16
- regulatory-frameworkRegimeFactsFact runn 16
- distribution-integrityDistributionFactsFact runn 16
- custody-disclosureCustodyFactsFact runn 16
- ai-adoption-componentsEurostatFact runn 2
- labour-componentsEurostatFact runn 2
- humanoid-labour-indexDisclosureFact runn 8
Unmeasured slots · 9 · no published run yet
- lab-practicepending frozen bankUnmeasured
- company-practicepending frozen bankUnmeasured
- measurer-integritypending frozen bankUnmeasured
- harness-fidelitypending frozen methodUnmeasured
- interop-conformancepending frozen bankUnmeasured
- humanoid-deploymentpending frozen registryUnmeasured
- interpretability-reproducibilitypending frozen bankUnmeasured
- claim-maintenancepending gated re-read wave (>=30 claims)Unmeasured
- serving-fidelitypending frozen bankUnmeasured
Which number do I quote?
Each figure below reads its own file and counts a different thing. They are never added together.
- open to read Models compared head to head on the board/interop/gspc-fleet-2026-08-12.json
- open to read Third-party AI models measured on frozen banks (quote this for “how many models”)/interop/models-measured.json
- open to read Models in the signed card set (a benchmark corpus)/signed/card-matrix.json
- open to read Models in Hugging Face Hub cells (a separate instrument)/api/hub-cards
- open to read Our own model tags (never ranked)/independence/own-model-disclosure.json
A tie stays a tie and an untested axis stays untested; neither is rounded up into a ranking, and an unmeasured slot is shown as unmeasured, never as zero. Read live from GET /api/gspc; the runs behind it were made behavioural axes 2026-08-12 · jail 2026-08-18 · financial-fact axes: see each axis facts_as_of / measurement_time · effect-binding 2026-09-22. Nothing here is a certificate.
How it works
Measure, sign, re-check, correct.
1 · Measure
The same frozen questions, graded by a rule
Every model answers the same published question bank, frozen before the run. A fixed rule grades each answer, never another AI's opinion. When two models cannot be told apart we say tie, and when an axis has not been tested we say untested.
How a measurement is made →
2 · Sign
Checkable evidence, free to verify
Measurement cards that carry an Ed25519 signature are independently checkable with a key we publish. Other measurement artifacts declare their attestation state explicitly, including content-addressed unsigned records, so we never imply a signature where none exists. Paste a signed card into the verifier: the record stays in your browser, while the browser may retrieve public-key metadata. No account is needed, and verification is free forever. A rank is never for sale, and our own models are listed apart and never counted in.
Check a record yourself →
3 · Re-check
Kept current, not filed away
A published claim carries its date. We re-read claims against their sources on a schedule; when a source moves, the claim is re-measured, marked stale or retired. The rules we follow are an open specification anyone can adopt.
How a claim is kept current →
4 · Correct
Our mistakes are public
When we get something wrong, the corrections ledger says what was wrong, how it was caught and what changed, with the date. Signed records are superseded, never quietly edited, because editing them would break the signature that makes them checkable.
Read the corrections ledger →
Start here
Ask it, connect it, or check it.
- AskAsk in plain words Each answer names the tool that produced it, the record it cites and the state it returned.
- ConnectConnect your AI Add the free MCP server to Claude, Cursor or any MCP client, or read the A2A agent card. Your assistant can then read the board and check a card itself.
- VerifyCheck a record yourself Paste a signed card and your own browser checks the signature. No account, free forever.
Who it is for
Buying or deploying AI
Compare dated results on the same frozen questions. Each record shows its attestation state; a tie does not establish a winner.
Models, axis by axis →Publishing AI-generated content
Check an image, video, audio file or PDF for a machine-readable mark. Open the evidence and its limits in the free preview.
Check an output for a mark →Building with AI
Call the board, fetch and verify cards from your own code or agent. Open source, no key for the free tools.
Developer doors →Regulating or auditing
Inspect the test, measurement date, source and attestation state behind a result. Gaps and corrections stay visible.
For regulators →An AI agent
Plain-language description at /llms.txt; agent card at /.well-known/agent.json; tools at POST /mcp.
Read llms.txt →
Liveread
How these are measuredThe work, counted live at the source
- 335 signed cards, every one verifies. Source: GET /api/state → card_chain.bodies_verified_valid (the signed card index).
Ed25519, checked by the verifier we publish
as of
- 93 public corrections, all dated. Source: GET /api/corrections.
signed ledger: what was wrong, and the fix
as of
- 52 MCP tools served live. Source: POST /mcp → tools/list.
read from the tool surface itself
as of
- 132 open datasets on Hugging Face. Source: Hugging Face API · datasets?author=csoai (public only).
17 likes
as of
Council of AI is operated by CSOAI Ltd (UK Companies House 16939677), founded by Nicholas Templeman. We measure; we do not certify, and verification is free.