Registry · Benchmark · Policy router · Marketplace

Nobody will tell you how to defend the AI you already run.

Every AI vendor publishes something about how they govern their system. Almost none of it says what you would need if someone asked you to defend the decision. This site reads what is actually published, scores what it finds and what it is silent on, and gives you two ways to check your own material below — no account, nothing uploaded, right now.

Request a demoOr walk through it yourself — four steps, no account →

Assessing this for a client? What tenant access costs, and what it does not include yet, is on one page →

Two things you can run yourself, now

Check what a vendor’s AI documentation is missing →

Paste any AI system’s published documentation — a model card, a usage policy, a safety page — and see which of seven governance questions it answers and which it is silent on. The same automated scan behind every card figure on the leaderboard.

See what is in your message before you paste it into a chatbot →

Account numbers, API keys, a colleague’s name buried in a thread you forwarded — check before you send, not after. The same patterns an organisation’s policy router strips on the way out, running here on your own text, in your own browser.

Both run entirely in your browser. Nothing you paste is uploaded, because neither page has anywhere to upload it to.

What a score here is, and what it is not

What this measures is the documentation, not the system. A high score means a vendor published material that answers the questions and can be checked. A low one means what is published does not answer them — not that the system is unsafe, and not that the vendor is non-compliant with anything. Every figure on this site carries the source it was read from and the date it was read, so you can disagree with it specifically.

AI systems scored from documents their vendors published

5/5

Frontier providers with an open evidence request, none answered

Published on this site. None has been served on a provider by any other route, so this records what the registry has received, not what anyone refused.

The provider table name specific companies, and are not published here. What is above this line is aggregate — true of the whole set and identifying nobody — and is published freely. The per-company half is withheld; today only the registry's operator can see it.

This is not coyness about the findings; every one of them is sourced, dated and checkable. It is that a statement about a named company needs someone identifiable standing behind it, and an informational-purposes notice is not that. The terms explain the difference.

Ask about an engagementOperator account? Sign in.

What reading everything they publish actually found

8 systems assessed by hand across all seven dimensions — safety frameworks, system cards, licences, deployer guidance and evaluations published by parties with no stake in the answer. The pattern is the finding, and it is the same one every time.

	What they are asked to document	Silent	When they do speak
D4	Fairness & Bias	6 of 8	they carry 41% of full marks
D5	Privacy & Data Handling	6 of 8	they carry 28% of full marks
D6	Accountability & Change Control	6 of 8	they carry 27% of full marks
D7	Regulatory Mappability	4 of 8	they carry 12% of full marks
D1	Documentation & Transparency	0 of 8	they carry 51% of full marks
D2	Evaluation Evidence	0 of 8	they carry 68% of full marks
D3	Safety & Robustness	0 of 8	they carry 41% of full marks

Every one of the 8 documents what the model is, how it was evaluated, and how it behaves under adversarial testing. Six of eight document nothing about fairness, nothing about data handling, and nothing about change control. The industry documents thoroughly for researchers and barely at all for the person who has to defend a decision to an examiner — and that gap is precisely what an examiner asks about. The full board →

Every card says what the model is. Almost none says what it does under stress.

58 of the most-downloaded open-weight model cards, read automatically on their published pages. Each row is a fact about documents, not a judgement about vendors — documentation held elsewhere is not counted, and every claim is capped at T2 because a card cannot verify itself.

	What the card is asked	Say nothing
D3	Safety & Robustness	54 of 58
D4	Fairness & Bias	40 of 58
D5	Privacy & Data Handling	40 of 58
D6	Accountability & Change Control	18 of 58
D2	Evaluation Evidence	15 of 58
D7	Regulatory Mappability	8 of 58
D1	Documentation & Transparency	0 of 58

Identity, licence and provenance are documented almost universally. Safety, privacy and fairness are documented almost never. That asymmetry is the finding, and no ranking shows it — a ranking sorts these documents by how much of the same small set of things they happen to mention. The full breakdown →

What this number is. It measures the documentation located about this system at a stated source on a stated date, against a published method — not the system itself. A low score means the located documentation is thin; it is not a finding that anyone governs their models poorly, that a system is unsafe, or that any law or rule has been broken. Documentation held privately or published elsewhere is not counted, because this registry cannot see it. It is published for information only, describes documentation as it stood on a stated date, and is not advice — verify anything you intend to act on. Terms, corrections and right of reply →

How a score is built

A dimension score is the strength of the evidence behind each claim — never an opinion, and never something a vendor can buy.

Tier	Value	Means	Who could have produced it
T0	0	No evidence located	Nobody
T1	25	Vendor assertion only	Marketing
T2	50	Vendor-published artifact	Vendor engineering
T3	75	Independently reproducible	Anyone, from published method
T4	100	Independently verified	A third party who reran it

Consequence: the only way to raise a score is to produce verifiable evidence. There is no lever money touches. Read the full method →

1 of 5

The policy router

One endpoint in front of every model your organisation calls. It applies your policy first — redacts before anything leaves, blocks models below your evidence floor, reroutes to an approved fallback, holds what needs a human — and hash-chains every decision. 6 policies configured.

Five things can happen, and all five are recorded the same way. A refusal is evidence in exactly the way an allowed call is — same chain, same hash, same export. That is the point: you can show what the policy stopped, not only what it let through. This router has not yet carried live traffic. Every request recorded so far is simulated, a test, or a refusal, and any claim made about it should say so.

Open the policy router →Next: Frameworks ↓

2 of 5

Frameworks

Control sets for NYC Local Law 144, ISO/IEC 42001, the EU AI Act, NIST AI RMF, SR 26-2 and more, each control cited to its source text, and each law marked when it is not yet in force.

What you say and what we can see are kept apart, permanently. A claim that a clause is covered is yours: recorded with your reason, never verified here, and never quietly upgraded. A signal is something this platform can actually observe — that your audit chain verifies, that a policy has human-review triggers, that a system is on the inventory. A signal is evidence a clause could be satisfied. It is never a claim that it is, and it is never written into your coverage.

See the frameworks →Next: Scan a card ↓

3 of 5

Scan a card

Paste any AI system’s published documentation and see which of seven governance questions it answers and which it is silent on. It runs in your browser; nothing is uploaded.

50 out of 100 is the top of this scale, not a bad score. A scan reads one document, and a document the vendor wrote can show they asserted something and published an artifact saying so. It cannot show that anyone independent reproduced or verified it — that needs somebody other than the vendor, which is what the upper half of the scale is for. So read a card result as how completely the document answers what a document can answer. It is not a measure of the system, and a low one is a statement about the page you pasted.

Scan a card →Next: Before you paste ↓

4 of 5

Before you paste

Account numbers, API keys, a colleague’s name buried in a forwarded thread: see what is in a message before it goes into a chatbot. The same patterns the router strips on the way out.

Nothing you paste leaves this page, and not because we promise. The detectors run in your browser, and this page has nowhere to send anything — no endpoint, no upload, no logging of what you typed. Close the tab and it is gone. You can check that claim rather than take it: open your browser’s network panel and paste something.

Check a message →Next: The method ↓

5 of 5

The method

Seven dimensions, five evidence tiers, every claim tied to a source and a date. The method is public so any score can be re-derived, and there is no lever money touches.

Seven questions, each asking for something you could go and find. Nothing here is a judgement about quality, a rating of a company, or an opinion about whether a system is safe. Each dimension names an artifact — a card, a red-team result, a retention statement — and asks whether it exists and what backs it. Two people scoring the same document should get the same answer, and where they would not, the question is wrong and we would rather hear about it.

Read the method →Next: everything else ↓

Nobody will tell you how to defend the AI you already run.

Two things you can run yourself, now

Check what a vendor’s AI documentation is missing →

See what is in your message before you paste it into a chatbot →

What a score here is, and what it is not

What reading everything they publish actually found

Every card says what the model is. Almost none says what it does under stress.

How a score is built

The policy router

Frameworks

Scan a card

Before you paste

The method

Everything else

Registry →

Leaderboard →

Marketplace →

Console →