Resources
Everything this registry produces, and what each of it costs
One page, so nothing has to be hunted for. The line between what is free and what is not is drawn on purpose and is stated below rather than left to be discovered.
Findings are free. Derived work is not. Every assessment, every score, every source it rests on, the method and the limitations stay open — the terms promise that a reader can re-derive any tier without trusting the assessor, and a paywall over the evidence would make that a lie. What is chargeable is the work that costs time to produce and cannot be reconstructed from the free half: the corpus as a time series, comparative figures across the whole set, and the documents an engagement delivers.
Open — 7 documents
Readable by anyone, no account. These are the method and the deliverable templates that make the method usable; keeping them open is what lets someone check the work.
| Document | Kind | Posture | Licence | |
|---|---|---|---|---|
| AI Vendor Due-Diligence Question Set Forty questions to send a vendor, ordered so that the answers map directly onto the seven GRI dimensions. Includes the follow-up that catches an evasive answer. | template | Either door | CC BY-ND 4.0 | Open → |
| EU AI Act Annex IV Documentation Checklist Article-level checklist for the technical documentation a high-risk system provider must hold, with the evidence tier each item realistically supports. | control library | Either door | CC BY-ND 4.0 | Open → |
| GRI Method Specification v0.1 The full scoring method: seven dimensions, five evidence tiers, five weighting profiles, and the stated limitations. Published deliberately — the method is not the moat. | template | Either door | CC BY-ND 4.0 | Open → |
| Governance Probe Battery — published subset The eleven publishable probes. The prompt-injection and name-origin disparity corpora are withheld: publishing them would let vendors tune to pass the test and destroy the instrument. | eval suite | Door B — Independent testing | CC BY-ND 4.0 | Open → |
| MSA Derived-Data Clause Set Contract language securing the right to retain de-identified, aggregated benchmark data from an engagement. Without this in engagement one, the comparative corpus never accumulates. | clause set | Either door | CC BY-ND 4.0 | Open → |
| Quarterly Governance Retainer — SOW skeleton Scope, deliverables and exclusions for a recurring quarterly governance review. The only engagement shape that is not priced by the calendar day. | clause set | Door A — Advisory | CC BY-ND 4.0 | Open → |
| SR 26-2 Generative-AI Gap Workbook Structured worksheet for documenting how an institution governs the generative and agentic AI that SR 26-2 explicitly excludes. Ten sections, each with the examiner question it answers. | workbook | Door A — Advisory | CC BY-ND 4.0 | Open → |
Gated — 1 document
Listed in full, including what is inside, so the decision to buy is informed rather than blind. Access is granted to an organisation, not a person, and payment is settled outside this system — there is no card processing here and none is planned.
| Document | What is in it | Price | Your access |
|---|---|---|---|
| Comparative Benchmark Medians — subscriber report Aggregate medians across assessed systems: where evidence typically fails, by dimension and by product category. Requires five engagements before any statistic is published. | Median dimension scores by product category · Distribution of evidence tiers by dimension — where the ceiling actually sits · The three claims most often made at T1 and never raised above it | not yet available | not granted |
The corpus
76 documents watched daily, 93 versions on file since 2026-09-08, 4 changed so far. 51 model card · 9 regulation · 7 independent evaluation · 4 usage policy · 4 governance page · 1 safety framework.
- What is watched, and why — open. Every document with the reason it is on the list.
- What changed, and when — open. A document that stops saying what it used to say, shown saying both.
- The series in bulk — gated. Every version of every document with its hash and dates, as a dataset rather than a page. This is the part that cannot be reconstructed by anyone who did not keep the earlier versions, and it is the reason there is a paywall at all.
Findings
8 assessed systems and the 9 independent sources their T3 and T4 claims rest on. Never behind the paywall — a reader must be able to check a finding without paying for the privilege. They are, separately, withheld from anonymous readers while no entity stands behind them; that is a different restriction with a different reason, set out in the terms.
The list of assessed systems and the sources behind them names specific companies, so it stays withheld here and is delivered to clients under engagement instead. The aggregate figures over the same set are public and name nobody.
Connecting your systems
For organisations routing their own AI traffic through the policy router.
Integration guide
Onboarding, the first call in curl, Node and Python, what to do with each decision and failure code, collecting the evidence, and installing the console on a phone.