Third-party review

What a review of someone else’s system gives you

You are buying, funding, insuring or relying on an AI system you did not build. We read the documents its makers hand over — policies, model and system cards, evaluation reports, contracts — against the clauses you need checked, and give you one finding per clause: the evidence is present, absent or contradicted, or the documents do not answer it and we say so.

Every finding points at the passage it rests on, so you can open the document and check it yourself. A clause the documents do not answer is refused by name, never filled in. Findings are ranked on three axes: how serious, how thinly evidenced, and how cheap to fix. The Pareto frontier of those three, the findings no other finding beats on all three at once, is what to fix first, and the report opens on it.

Each report names every check it ran, the version of our engine that ran it and the date, and it carries a signed statement you can verify without contacting us.

What we check a system against

Beyond the clauses you choose, every review checks six properties of a system that keeps an honest account of itself:

• The system keeps an account in which every act is appended and none is edited, chained so that an edit anywhere shows.
• What the system is trying to achieve is written down, where a reviewer can read it.
• When the system declines to act, the refusal is recorded as an act, not left as a silence.
• Every decision carries a certificate a reviewer can re-open and re-check.
• What the system puts into its surroundings is counted, and each count says what it was measured on.
• The walls between customers are shown by trying to cross them, not by a policy that says they exist.

And the clauses you send are read against the frameworks buyers and regulators are asking about this year:

• The EU AI Act, as amended in 2026, including its transparency duties and the code of practice for general-purpose models.
• The NIST AI Risk Management Framework and its generative AI profile.
• The ISO/IEC standard for AI management systems, with its companion standards on impact and risk.
• AIUC-1 for AI agents.
• The AICPA trust services criteria, ISO 27001 and the NIST Cybersecurity Framework for the security underneath.
• SR 11-7 where the system is a bank’s model.
• California’s frontier AI transparency law, the Colorado AI Act and New York City’s automated hiring audit law, where they apply.

How a review engagement runs

The same three shapes as every engagement. A diagnostic of about a week reads one system’s documents against the clauses you name and ends in the ranked report. A pilot repeats it on the systems you depend on most, with the clause list agreed in writing first. Build and hand over puts the review in your own hands: your team runs it on every new supplier through the professional API, as the assurance_review call.

The honest limit: what we are not, and what we have not measured

We are not a certification body, and this report does not say "compliant" or "certified". It says evidence present, absent or contradicted, at commit X on date Y, by instrument Z. Certification is the job of the bodies authorised to certify; a review from us is the evidence you take to one, or use to decide whether you need one.

All six properties can only be re-checked when the system runs on our platform. When it does not, those findings read “not measurable”, by name, and the clause findings carry what its own documents say. We hold no licensed copy of any regulation or standard: you send the clause wording you want checked.

The review’s accuracy on third-party documents is not yet measured. These are the bars it has to meet before we publish a figure, and each cell below changes only when a measured run fills it.