We measure AI systems, sign the result, and publish what we could not measure
In 2024, as artificial intelligence began transforming every industry, a critical question emerged: Who watches the watchmen? Governments scrambled to regulate. Companies rushed to comply. But one thing was missing: trained professionals who could actually monitor AI systems for safety.
That's when CSOAI was born—not as another AI company, but as the solution to two problems at once: making AI behaviour checkable by people who are not the vendor, and training the people who will have to do the checking.
Living GSPC board — GET /api/gspc
Signed cards published and verifying — GET /api/state
Frozen provision bank — anchored by a published corpus hash
Published crosswalk — EU · UK · US-IL · CN
Making AI behaviour checkable
AI displacement forecasts vary enormously and we have measured none of them, so we will not put a number here. What is not a forecast: the EU AI Act already requires human oversight of high-risk systems under Article 14, and someone has to do it.
Without CSOAI
- ✗AI systems deployed without proper safety review
- ✗Companies struggle to find qualified compliance staff
- ✗Nobody outside the vendor can re-run the test that produced the claim
- ✗Governments lack trained personnel for AI oversight
With CSOAI
- Every published finding links to a signed, recomputable record
- Free training, and completion records that say what they are
- A published corrections ledger — appended, never edited, including our own failures
- An open, recomputable method — the banks, the grader and the rows are published, so anyone can disagree with us in our own units
Protecting Humanity While Creating Careers
We are not a certification body and never will be — we measure, sign and publish evidence, and the competent authorities decide. What we are building is the infrastructure for a role we think the law is about to require: the AI Safety Analyst. That this becomes a large profession is our bet, not a measurement — we have no forecast to cite and we are not going to invent one.
Train
Comprehensive training on EU AI Act, NIST AI RMF, and ISO 42001 frameworks. No coding required.
Complete
Finish the assessment and the Academy issues a completion record: a signed statement that you demonstrated the material on a date. It is not a certification, not an accreditation, and confers no regulatory status of any kind — we issue no conformity marks to anyone, ourselves included.
Contribute
Challenge a published measurement, re-run a frozen bank, or file a correction — the harness and the banks are public and a correction that lands is published under your name. A paid analyst marketplace is not yet available: we have no roster, no matching and no engagements to offer, and we will not describe one before it exists.
What is crosswalked, and what is only tracked
Our published, signed crosswalk maps one measurement across four regimes: the EU AI Act, the UK DRCF alignment, Illinois SB 315, and the Chinese TC260 alignment. That file is at /crosswalk/east-west-v1.json and it is the whole of what is crosswalked today. We track and write about many more jurisdictions than four; those are reference pages, not crosswalked measurement, and this page will not blur the two.
North America
Crosswalked: Illinois SB 315. Tracked, not crosswalked: NIST AI RMF, federal guidance and other state AI laws — reference coverage we write about, with no signed crosswalk behind it yet.
Europe
Crosswalked: the EU AI Act, and the UK DRCF alignment beside it. This is the deepest coverage we have — the frozen provision bank behind it is anchored by a published corpus hash. Measurement, never certification.
Asia-Pacific
Crosswalked: the Chinese TC260 alignment. Tracked, not crosswalked: Japan, India, Australia and Southeast Asia.
Middle East & Africa
Tracked, not crosswalked: emerging AI-governance regimes across the Middle East and Africa. Nothing here is measured yet, and saying so is the point of the column.
We're Not Just Talking. We're Building.
Multi-provider oversight — designed; latest test fully correlated
The intended architecture spreads review across providers so no single vendor decides alone. The council-seat figure is a design, not a live system. The latest point experiment measured rho=1 and n_eff=1 across three nominal legs, so it does not establish independent or unbiased review.
Why it matters: When a company's own AI reviews their AI, there's a conflict of interest. Our multi-vendor approach ensures unbiased safety assessments.
👁️ Watchdog: a public intake
There is a public intake for behaviour that looks wrong, at /public-watchdog. What it is: somewhere to report, that anyone can use. What it is not: a published incident register — we do not yet operate one, and a triaged public register of reports is not yet available. We also hold nobody accountable: we measure, and only a regulator can approve, ban, fine or clear anything.
Why it matters: an intake with no register behind it is worth less than one with, and pretending otherwise is the kind of claim this page exists to retire.
🔄 Re-attest, never edit
A model measured in August is not a model measured next year, so a measurement is a dated record and never a standing verdict. When we re-measure, we issue a NEW signed card and the old one stands — history here is append-only, and drift is visible by comparing dated cards. Scheduled automatic re-attestation is not yet available: re-measurement is arranged run by run today, so do not read this as a monitoring subscription.
Why it matters: a PDF from six months ago describes a model that no longer exists. A dated, signed card at least tells you which model it describes.
💼 Free training, and an honest account of where it leads
Council Academy training is free and its completion records are free. What we have not measured, and will not assert: how many people this puts into work, what they earn, or how our focus compares with other organisations' — we have measured one rating organisation on one criterion and that is the whole of our comparative evidence. The training exists and is free; the career outcome is unmeasured.
Why it matters: a training pipeline that promises income it has not measured is selling something. Free training with an unmeasured outcome, stated as such, is the honest version of the same offer.
By the Numbers
Every figure below is read from a live endpoint at page load, with the artifact and the date it came from. None is typed into this page — that rule exists because typing them once produced five different axis counts across this site simultaneously.
Living GSPC board — counts from GET /api/gspc. The larger number counts slots; the smaller counts measurements. A published slot exists so a gap is visible, and is not evidence that anything was measured.
Published cards that verify under our single public key — counts from GET /api/state
Platform Uptime — see /status for probed services
Withheld cards a signature actually attests. The rest are disclosed in an unsigned manifest — see the honesty gate
Our Commitment to You
We build in public. The banks, the grader, the rows, the signing keys and the corrections are all published, so the way to disagree with us is to recompute a number and show us a different one — and if you do, it goes in the ledger under your name. CSOAI LTD is a UK company (Companies House 16939677) with shareholders like any other; what makes us checkable is not our ownership but that nobody we measure pays for their place on the board, their score, or their removal from either, and verification is free forever.
Aligned with Global Standards
Our training and our instruments are written against three reference frameworks — the EU AI Act (Europe), NIST AI RMF (United States) and ISO/IEC 42001 (International). Written against is not measured against, and the difference is the whole point: the EU AI Act is the only one of the three with a frozen, corpus-anchored provision bank and a published crosswalk behind it. NIST AI RMF and ISO/IEC 42001 inform how the instruments were designed; no signed crosswalk to either is published, so treat them as reference rather than as coverage. We hold no accreditation under any of the three, issue no conformity assessment, and a course completion attests training and never conformity.
Insurance evidence
CSOAI LTD does not state professional indemnity cover until the policy document is on file.
No policy document on file
No coverage, policy number, insurer, broker, or term is claimed on this page. If evidence is filed, the public statement and corrections record will be updated together.
Corrections ledger C-2026-0902-06Evidence status
No professional indemnity policy document is currently on file
UK Registered
Companies House No: 16939677
Registered in England & Wales
Public correction
The withdrawn claim remains recorded in the corrections ledger
Frequently Asked Questions
Everything you need to know about CSOAI
Who can become an AI Safety Analyst?
Anyone with critical thinking and attention to detail. You do not need a computer science degree or coding experience. The training covers AI governance frameworks, risk classification and how our measurement method works — enough to read a system and a card critically. It does not make you a qualified auditor, it confers no regulatory standing, and it is not a route to accreditation, because we hold none to pass on.
How long does the attestation exam take?
The exam is 50 questions in 90 minutes with a 70% passing threshold, and you can retake it as often as you need. We have not published completion times or a first-attempt pass rate, so we are not going to quote either — there is no cohort record behind those figures. What you get on passing is a completion record: a statement that you demonstrated the material on a date. Not a certification, not an accreditation, and no regulatory status.
What do AI Safety Analysts actually do?
The honest answer is that this role is a bet on where the law is going, not a job we are currently staffing — see the earnings question below. The training covers reading a system against the EU AI Act and against the NIST AI RMF and ISO/IEC 42001 as reference frameworks: classifying risk, checking documentation, and writing up what is evidenced and what is not. What it does not cover, because we do not do it: making a final safety determination. Nobody here makes one. Determination stays with the competent authorities, and our part is the measurement they can recompute.
How much can I earn?
There are no analyst engagements to offer today. A paid analyst programme is not yet available — no roster, no matching, no published terms and no cohort — and we are not going to describe rates, hours or a body of working analysts before any of that exists. What is available now is free: the training, the completion records, the banks, the grading code and the right to challenge any published measurement.
Why should companies trust CSOAI?
Not on our architecture, and we will say why. We once claimed that a multi-provider council delivers decorrelated, unbiased review. The earlier DR-0007 experiment caused us to withdraw that guarantee; the latest point experiment measured rho=1 and n_eff=1 across three nominal legs. Trust us on the things you can check instead: no model judges a model — every verdict is deterministic code against pre-written gold labels; every published card verifies offline against a key in our DID document, with no account and no permission; nobody we measure pays for their place on the board, their score, or their removal from either; and what we got wrong is appended in public at /api/corrections and never edited.
How is CSOAI different from other AI safety organizations?
We have not surveyed the field, so we will not tell you what most AI safety organisations do — our own claims register records comparative coverage of the evaluation landscape as UNMEASURED (CR-020). What we can describe is what we publish: a measurement instrument — 23 axis · 23 measured, live from GET /api/gspc, where a slot with no run behind it is published UNMEASURED and never counted as a measurement — over frozen benchmarks, open on Hugging Face and Kaggle with the scoring code, so anyone can recompute what we claim or dispute an answer key. And we publish the results that go against us: our own fine-tunes sit below the base models they were built on in our own signed arena reference. That is on the honesty gate, read live from the artifact rather than typed.
