Who runs us, who pays, and where we have an interest

The Council of AI is run by CSOAI Ltd, a private limited company, Companies House number 16939677, incorporated on 2 January 2026, with its registered office in London, England.

The register lists one officer and no resignations: Nicholas Brian George Templeman, director, appointed 2 January 2026. It lists one person with significant control, the same person, through ownership of 75% or more of the shares.

So one person directs and controls the company that writes our tests, runs them, signs the results and publishes them.

Source: Companies House, officers and company overview read 27 September 2026 at 01:31:59 UTC; persons with significant control read at 01:32:10 UTC.

Roles, shareholdings or paid work that anyone at CSOAI holds with an AI developer whose models we measure: not yet published.

Funding sources: not yet published; request via nicholas@csoai.org.

In-kind support (free compute, credits or programme membership): not yet published.

Reading what we have published, and checking any signature on it, needs no account and costs nothing.

This is our largest conflict of interest, so here it is with the numbers.

Of the 335 cards in our signed card index, 230 measure a model we built ourselves.

• 224 cards, across 46 model tags, name a model whose tag begins sov or clan. Our own fine-tunes are published under those two prefixes.
• 6 cards name a model whose tag begins with the word council (council-oowm:latest (3), council-safe:latest (3)). This is the same test the public board uses to recognise our models.
• 5 cards, on eat-unsloth-050b:2026-08-02 (2), qwen2.5-0.5b-mined:latest (2), muse-glimmer:latest (1), may also be ours. Their names suggest a model we derived, but nothing we hold confirms it, so they are counted on neither side.
• 100 cards match no rule. That describes the rule, not who built the model.

How each card was classed: only the model field of each signed card body is read, and it is tested against the name rules above, in that order. Ownership is decided by name, not by a registry of model digests, which is the weakness of this count.

Source: own-model-disclosure.json, derived on every build from /signed/card_index.json (SHA-256 4ff54c78dba0b15b…, index created 2026-08-19). The signed card index is one of three separate card collections we publish, and these counts describe that one only. They are not a count of everything we have measured.

How we keep it off the public board

The public board does not name our own model as the leader. Where one of our models held the point lead, the board shows no leader at all: it neither names our model nor promotes the runner-up without a fresh ranking. The measurements and their signed cards stay on the record.

Read live: the board withholds the leader on 8 of 14 model-comparison axes (governance, provenance, continuity, conformance, openness, care, art5-safeguard, affect). Of the 9 leaders it does name, 0 match any of our own-model name rules.

A gap we have not closed: the board’s test only recognises names that begin with council or are marked “council specialist”. It would not recognise a sov or clan tag. The check above runs every rule against the named leaders so you can see it holds today.

Source: /api/gspc → totals.own_leaders_excluded and axes[].public_leader_state = EXCLUDED_OWN_MODEL. The rule is isOwnCouncilModel in functions/api/gspc.ts.

Proposed, not built: each new card would carry a signed yes-or-no field saying whether it measures one of our models, decided from a published registry of model digests rather than by name.

Our DID document lists 6 Ed25519 keys, and names did:web:csoai.org as the controller of 6 of them. That is our own identifier: every key listed there is ours, and our cards and our board are signed with these keys, not by an outside party.

• did:web:csoai.org#site-release-1
• did:web:csoai.org#estate-chain-1
• did:web:csoai.org#board-attestation-1
• did:web:csoai.org#card-attestation-1
• did:web:csoai.org#gspc-board-22axis-2026
• did:web:csoai.org#card-attestation-2

The board is signed with #board-attestation-1. The cards in the signed card index are signed with #card-attestation-1: read live, 335 card bodies verify, under 1 signing key.

A valid signature shows that we signed the bytes. It does not show that anyone else has checked them.

Source: /.well-known/did.json (also served at csoai.org); /api/state → card_chain.

The same company writes the frozen tests, runs the models, grades the answers, signs the cards and publishes the board.

No outside party has yet re-run one of our measurements, as far as our records show. Outsiders have checked our published bytes: an outside audit on 26 August 2026 recomputed the jail axis from our published per-item rows and found one arithmetic error (correction C-2026-0826-09), and an outside implementer verified a card signature in Python (C-2026-0826-07). Recomputing from our rows and checking a signature are useful, but they are not the same as running the test again.

Read live: our corrections ledger holds 78 entries, of which 14 were reported from outside CSOAI.

Source: /api/corrections → corrections[].detected_by; readable at /corrections/.

1. Check it yourself first. Every signed card can be verified on your own machine with our public key, with no account and no permission from us: how to verify.
2. Ask for a re-check. Email nicholas@csoai.org with the card id and what you think is wrong. Disputes are answered by re-running the frozen test: appeals and disputes. Today the owner of CSOAI Ltd reviews every dispute; there is no independent arbiter yet.
3. Lodge a formal objection. /challenge/ gives you a receipt, but challenges are not stored on our side yet, so email us as well and keep the receipt.
4. See what we got wrong before. Every correction goes in the public corrections ledger, which says what was wrong, how it was caught and what changed.

A score describes one run of one model on one frozen test on one date. We do not certify, approve or endorse anything, we issue no conformity marks, and we are not a regulator. A good result from us is not a compliance verdict, and a bad one is not a ban.