CSOAI - for regulators + policy bodies

An assurance baseline you can verify - and publish

You need something you can check without trusting the party who published it. The core here is MIT-licensed, the banks are public, and each measurement is Ed25519-signed against a key you fetch yourself - so anyone can recompute a number and challenge it.

Regulator atlas ->What a signed System Card is ->

Your path with CSOAI

Regulator atlas

The global regulation map

Open ->
The published crosswalk

Four regimes, mapped row by row

Open ->
Council Space

Talk to the governance AI

Open ->
The honesty record

Corrections, published not edited

Open ->

What you can check, right now

The row that matters most to a supervisor is the empty one. regulatory-framework is a published slot with no run behind it, and it sits here beside three measured axes rather than being left off the page. That is the whole proposition: you can see what we have not measured as plainly as what we have.

AxisBenchnLeader accuracy95% CISeparation
governanceGovBench23770.0%63.9–75.5%SEPARATED p=0.0086
art5-safeguardArt5Bench3697.2%85.8–99.5%TIE — indistinguishable p=1
conformanceMCPBench3574.3%57.9–85.8%TIE — indistinguishable p=1
regulatory-framework—nothing measuredunmeasuredunmeasuredUNMEASURED
Every figure above is read from GET /api/gspc when this page loads — none of it is written into the page. Recompute a signed card yourself at /gspc-verify, walk every chain position at /signed/chain.json, and read the rules the grades are computed under at /methodology. A leader is the highest point estimate on the board, not an approval; a TIE is not a win; and unmeasured means no run exists — never zero. We measure against these obligations; we do not enforce them and we certify nothing.

verifiable

Publish proof, not promises

Each measurement card is signed over its exact bytes, and the board snapshot carries its own site attestation naming the key that signed it. Re-verification needs no account and no CSOAI code.

What you can check on the board

The two axes a supervisory reader usually wants first: can a model recognise a prohibited practice, and can it place a system in the right risk tier.

Every figure below is read live from GET /api/gspc when this page loads — no count is typed into it. These axes measure how a fleet of models behaves on a frozen, published bank on a date. None of them is an assessment of your organisation, and none is a conformity opinion: determination stays with your regulator.

Board right now: 22 axes · 15 measured — a published slot is not a measurement, which is why both numbers travel together.

  • art5-safeguardMEASUREDArt5Bench

    EU AI Act Article 5 prohibited-practice trip: does the model recognise a prohibited practice when it sees one.

    What is graded:
    EU AI Act Article 5 prohibited-practice trip
    n:
    36 bank items
    Leader:
    97.2% (council-relationality-v3-light (council specialist))
    Separation:
    TIE — the leader's lead is not statistically separated, and a tie is never counted as a win.
    Frozen bank:
    csoai/gspc-art5
  • governanceMEASUREDGovBench

    EU AI Act risk-tier classification. The largest bank on the board, and the judgement most downstream duties hang off.

    What is graded:
    EU AI Act risk-tier classification
    n:
    237 bank items
    Leader:
    70.0% (council-embodiment-v3-light (council specialist))
    Separation:
    SEPARATED
    Frozen bank:
    csoai/gspc-gov
  • machinery-conformityMEASUREDMachBench

    Machinery Regulation self-evolving safety-function classification — a second statute, graded deterministically the same way.

    What is graded:
    Machinery Reg self-evolving safety-function classification (PART_A / OUT_OF_SCOPE / NOT_SAFETY_FUNCTION)
    n:
    33 bank items
    Leader:
    54.5% (llama3.2:3b (base model))
    Separation:
    TIE — the leader's lead is not statistically separated, and a tie is never counted as a win.
    Frozen bank:
    csoai/gspc-mach

Re-check it without us

Each published measurement card is Ed25519-signed over its exact bytes, and its id is the sha256 of those bytes. Pin the key from our DID document first — a card verified against the key it ships with proves only that the file is self-consistent.

Ed25519-signed
every published measurement card
Verify without an account
pin our key, recompute the bytes
Empty cells stay empty
unmeasured is published, not hidden
MIT-licensed core
no vendor lock-in

Questions, answered

How is CSOAI independent?

The core is MIT-licensed and every published measurement verifies against a key you fetch from our own DID document, so the assurance does not rest on trusting a single commercial vendor. We measure; determination stays with authorities.

Can these numbers be independently verified?

Yes - that is the design, and the method is published rather than described. Pin the key from /.well-known/did.json, fetch a card from the published index, recompute the sha256 of the exact signed bytes and check the Ed25519 signature. If the bytes do not verify, the card is not ours.

Council OS — regulator

Answers from published measurement, or it refuses. Your question is typed into the lobby — nothing sends until you press Ask.

Open Council OS

Deterministic pane commands · grounded /api/chat lane · consent checkpoint on consequential steps