Methodology
This registry is an evidence index, not a legal determination. It records what public documents say and don’t say about each offering, with a primary-source link, a verbatim quote, and the date the value was last confirmed on every cell. It never answers “can I use this legally”, that is a question for your counsel, informed by evidence like this.
How cells are researched and graded
Cells are researched and graded by autonomous LLM agents working against a published rubric, not filled in by hand. A daily process rolls through the registry, oldest cells first. For each cell, a verifier agent searches for the provider’s own authoritative page, fetches it (rendering JavaScript trust portals and reading PDFs), and grades the cell against the rubric, which fixes the meaning of every value, the rules for which sources count as authoritative, and the per-dimension precedents. It records a value only on evidence it actually fetched: a primary-source link, a verbatim quote from that page, a confidence, and the date. It is instructed never to assert a fact from memory.
Every grade is then audited by a second, independent agent (a different model) that re-reads the cited page and checks three things, that the source is the provider’s own authoritative domain, that the quote is really on it, and that the value fits the evidence, and corrects clear mistakes. Anything the agents genuinely cannot verify, because no authoritative page exists or the page will not load, is flagged for human review rather than guessed. Corrections sent in by the public are applied by a human on top of this.
What a cell means
Every cell grades the public evidence for the compliance-favorable answer to a fixed question, for example, “will they sign a HIPAA BAA covering this offering?” The value is a coarse index for scanning; the note and source on each cell are the actual evidence, read them. The grades:
| Status | Meaning |
|---|---|
| yes_public | Publicly verifiable from the provider’s own documents. |
| yes_sales_gated | The provider states it exists, but the artifact (report, agreement, feature) requires a sales conversation, NDA, or enterprise tier. |
| yes_platform_only | Available, but only through a particular deployment or platform path. |
| partial | Some of the question is publicly satisfied; a material part is not. |
| no_public_evidence | We could not verify the favorable answer from public sources. This is not the same as “no”: the artifact may exist behind an NDA or an enterprise agreement. The note says what we checked. |
| no_verified | We checked an authoritative, complete source and the answer is a confirmed no, not merely unverified. Used where such a source exists, for example the European Commission’s own GPAI Code of Practice signatory list: a developer absent from it has not signed, which is a stronger and more useful statement than “no public evidence.” |
| not_applicable | The question doesn’t apply to this row (for example, GPAI Code status for a serving platform that is not the model’s provider). |
| unclear | Sources exist but don’t settle the question. |
| conflicting_sources | Public sources disagree. The cell’s notes say how. |
Verification discipline
- Only primary sources are cited: the provider’s own trust center, policies, terms, and official EU pages. A third-party page that merely mentions the provider does not count, and press coverage and law-firm summaries are never cited as sources. Every grade is anchored to a verbatim quote from the page cited.
- Every cell carries the date its value was last confirmed, the last time a run found the evidence still standing, not merely the date a page was fetched.
- Each cell carries a confidence grade. high means the source states it directly; medium means it required minor inference or the source is partly gated; low means treat it as a lead, not a fact.
- Offerings are (developer, platform) pairs. “Claude via AWS Bedrock” is a different row from “Anthropic API” because the certifications, the data path, and the contract you sign are different. For distribution rows, vendor-trust and data-handling dimensions describe the serving platform’s terms for that offering; EU AI Act dimensions describe the model developer, whose obligations they are.
On audit reports (SOC 2, ISO)
Essentially every SOC 2 Type II report is confidential and released under NDA, so “can you download the report” is true of almost nobody and tells you nothing. What actually differs between providers is whether the provider publicly documents that this specific offering is inside the audit scope, and that is what these cells record.
- yes, public: a public page from the provider names this offering as in scope. AWS, for instance, lists “Amazon Bedrock” by name on its public SOC services-in-scope page.
- yes, sales-gated: the provider states it holds the certification, but the scope, the report, or both sit behind a sign-in, an NDA, or a sales conversation. Microsoft names no AI service on its public Azure SOC 2 page; the per-service scope is only in the report, behind the Service Trust Portal sign-in.
So an offering being sales-gated here is not a criticism of the provider’s security. It means you cannot confirm the coverage of the exact thing you are buying without asking them for it.
On the EU AI Act columns
The GPAI Code of Practice is voluntary. A provider absent from the signatory list may demonstrate compliance by alternative means. The registry records signatory status (full, partial by chapter, or absent from the list) and whether an Article 53 training-data summary is published, it never claims a provider “has no EU compliance path” or is “non-compliant.”
Corrections policy
If you work for a provider listed here and a cell is wrong or stale: send the primary source that corrects it. Corrections with evidence are folded in within days, credited in the changelog, and the cell’s verification date is refreshed. Email [email protected] with the source linked. Disputes about grading (rather than facts) get a note in the cell rather than a silent change.
Use the data
data.json is the whole registry, free, machine-readable. Attribution with a link is required. If you need dated, procurement-attachable evidence snapshots or a licensed feed, subscribe in the footer and say so - that demand signal is what decides whether those get built.