AI DevelopmentMethodology6 min readPublished September 20, 2026

12 arrangements · 3 labs · 2 governments · primary documents only · who pays, who publishes

Who Checks a Frontier AI Lab's Work? 12 Arrangements

A census of 12 external evaluation arrangements at Anthropic, OpenAI and Google DeepMind: who evaluates, who pays, what access they get, what gets published.

DA
Digital Applied Team
Research and practical guidance
Editorial dateSeptember 20, 2026
Sources readSeptember 22, 2026

"Independently evaluated" appears on model pages, in procurement answers and in policy submissions, and it can mean a dozen different things. On September 18, 2026 Anthropic announced that Faculty, Accenture's AI unit, will place evaluators inside the company with access comparable to an employee's, and that Anthropic will fund the work directly. In the same post Anthropic said there are as yet no standards for what such evaluators should see or how they should report.

That is the right moment to lay the arrangements side by side. This census reads the primary documents for 12 of them, across Anthropic, OpenAI and Google DeepMind, the nonprofits METR and Apollo Research, and the UK and US government institutes, and answers five questions for each: who evaluates, who pays, what access they get, whether results are published, and whether the evaluator can delay a release. Every cell is what the document states; where it is silent the cell says unknown. This blog runs on Anthropic's models, and Anthropic's rows are held to the same rule as the rest.

Key takeaways
  1. 01
    No document in the census gives an outside evaluator the power to delay a release.Where a document addresses the question, the mechanism either reviews a report, feeds an internal decision body, or is periodic and decoupled from launches; two rows leave it unaddressed and are marked unknown rather than no. That is an observed absence across these 12 documents, not a claim about arrangements we did not read.
  2. 02
    Where the payer is stated, it splits three ways: the lab, the government, or nobody.Anthropic funds Accenture directly and OpenAI offers compensation to all its assessors. UK AISI and US CAISI are government-funded. METR states it has taken no payment for company-identifying assessments and declines lab donations.
  3. 03
    Publication rights range from unrestricted to approval-gated to unstated.Anthropic's Risk Report reviewers may publish anything but confidential information. OpenAI's contract excerpt requires written approval before publication. The Accenture post states no publication mechanism at all.
  4. 04
    Access is described in words, rarely in items.Only two documents list access as items: OpenAI's account of the UK AISI engagement and METR's Frontier Risk Report. 'Comparable to an employee's' and 'shared access to data and ideas' are the norm; weights are never explicitly granted in any document here.

01The hookThe September 18 announcement

Anthropic's post frames the partnership as a step toward a commitment its CEO made in an essay to embed evaluators within the company. The scope is evaluating and red-teaming models, alignment assessments and safeguard testing. Three terms are stated plainly and matter for a buyer. Anthropic funds Accenture's work directly. Each company expects to invest at least $1 billion over five years in building capacity in this area, which is an investment horizon and not a contract term; no term is stated. And the partnership is non-exclusive, with other evaluators to be announced and Accenture free to work with other developers.

Two things the post does not say are as important. It does not say whether Faculty's findings will be published or on what approval. And it does not give the evaluators any power over releases; it says the opposite, that the safety of the models remains Anthropic's responsibility. The post also says Anthropic is in dialogue with METR and other nonprofits to pilot elements of embedded evaluation using their own funding, and that long-term funding should come from pooled or government sources. Our post on what the CEO's essay promised covers the origin of the commitment.

There are, as yet, no standards for what information embedded evaluators should have access to, or how they should report what they find. There is also no settled system for funding independent evaluation.Anthropic, embedded-evaluation partnership announcement, September 18, 2026

02The censusTable 1: who evaluates, who pays, what access

Twelve arrangements. An arrangement is one documented relationship between an evaluator and a lab, or one policy mechanism that defines such a relationship. "Unknown" means the primary documents consulted do not say. Where the payer is a government, the row states that and notes whether any lab payment appears.

Primary documents of the labs, evaluators and institutes named, read September 22, 2026. No document dated after September 20, 2026 is used; undated pages are cited as published at the time of writing.
ArrangementWho evaluatesWho paysAccess granted
Anthropic × Accenture embedded evaluationFaculty, Accenture's specialist AI businessAnthropic, directly, per Anthropic's postDescribed as comparable to an employee's: watching training, following build and deploy decisions, speaking to staff. No item-level list of weights, checkpoints or logs.
Anthropic Risk Report external review (RSP v3.4 §3.6)One or more third parties, unnamed in the policy; Board consulted, Long-Term Benefit Trust approvesNot stated in the policyThe private Risk Report, unredacted or minimally redacted, within one week of Board submission; follow-up questions answered.
Anthropic annual procedural compliance reviewAn unnamed third party commissioned by AnthropicAnthropic (it commissions the review)Not stated
METR review of Anthropic's Opus 4.6 Sabotage Risk ReportMETR, a nonprofitNot Anthropic: METR's conflict-of-interest policy says it has received no payment for company-identifying risk assessments to dateAn unredacted version of the Sabotage Risk Report and other materials, per METR.
METR Frontier Risk Report (Feb–Mar 2026)METR, covering Anthropic, Google, Meta and OpenAINot the participants; METR cites broad independent fundersEach participant's most capable internal models at the time, including raw chains of thought, plus non-public capability and monitoring information.
OpenAI third-party assessorsNamed examples in OpenAI's post: METR, SecureBio, Apollo Research, IrregularOpenAI offers compensation to all assessors; some decline; no payment contingent on resultsEarly checkpoints, select evaluation results, models with fewer mitigations, direct chain-of-thought access for some. Weights not stated.
UK AI Security Institute × OpenAI biosecurity red-teamingUK AISI, a government research organisationUK government; no OpenAI payment stated on either sideFive stated forms including helpful-only variants, safeguard prototypes, monitor chain of thought and selective mitigation disabling.
US CAISI agreements with OpenAI and AnthropicUS Center for AI Standards and Innovation, at NIST; the US AI Safety Institute when the 2024 agreements were signedUS government; no lab payment statedMajor new models before and after public release, per the 2024 NIST announcement. Weights, logs, variants: unknown.
UK AISI × Google DeepMind research MOUUK AISI jointly with Google DeepMindUK government funds AISI; cost-sharing unknownShared access to data and ideas and joint publications, per AISI. Model access terms: unknown.
Google DeepMind Frontier Safety Framework v3.1No third-party evaluator is named; external parties involved where required or appropriateUnknownUnknown; the framework states no access terms for external parties
Apollo Research × OpenAIApollo Research, a public benefit corporation as of 2026Unknown individually; covered by OpenAI's blanket compensation statementPre-deployment testing of o1 in 2024 and anti-scheming work in 2025, per Apollo; specific access forms unknown
UK AISI × Anthropic cyber capability testingUK AISIUK government; no Anthropic payment statedA pre-release Claude Mythos 5 running without cyber safeguards and with internet access, per Anthropic's incident account. Precise terms unknown.

03The censusTable 2: publication and release power

The same 12 rows on the two questions a buyer or policy reader most needs answered: does the public ever see the evaluator's findings, and can the evaluator stop or delay a model shipping.

Same sources as Table 1, read September 22, 2026. The source column gives the document's own date.
ArrangementResults published?Can delay a release?Source
Anthropic × AccentureUnknown; evaluators 'can also report incidents', no publication right or cadence statedNo such power stated; Anthropic says safety remains its responsibilityAnthropic, Sep 18, 2026
Anthropic Risk Report reviewYes: a written report made public, commentary within 30 days, no restriction beyond confidentialityNo; the mechanism reviews reports, not releasesRSP v3.4, Jul 8, 2026
Anthropic compliance reviewUnknown; the policy commits to commissioning, not publishingNo; scope is procedural compliance, not outcomesRSP v3.4, Jul 8, 2026
METR × Anthropic Opus 4.6Yes: two review documents with METR's disagreementsNo; the model had been deployed for weeks at review timeMETR, Mar 12, 2026
METR Frontier Risk ReportYes, with participants approving what non-public information could be disclosedNo; entity-based and decoupled from releases by designMETR, May 19, 2026
OpenAI third-party assessorsConditionally: summaries in system cards; assessor publications need OpenAI's written approvalNot stated; deployment decisions sit with OpenAI's Safety Advisory GroupOpenAI, Nov 19, 2025; Preparedness Framework v2, Apr 15, 2025
UK AISI × OpenAI biosecurityPartially, by OpenAI; no standalone AISI report citedNo; described as ongoing rather than tied to a launchOpenAI, Sep 12, 2025
US CAISI agreementsPartially: CAISI publishes evaluation news items; agreement text not publicUnknown; not addressedNIST, Aug 29, 2024; CAISI site
UK AISI × Google DeepMind MOUJoint publications intended; MOU text not publicNo such power stated; a research MOUUK AISI, Dec 11, 2025
Google DeepMind FSF v3.1Not committed; information shared with governments if a critical capability level is reachedInternal governance bodies onlyFSF v3.1, effective Apr 17, 2026
Apollo × OpenAIYes in named cases, after OpenAI's confidentiality and accuracy reviewNo such power statedApollo site; OpenAI, Nov 19, 2025
UK AISI × Anthropic cyberAISI reported an incident Aug 4, 2026; Anthropic published its own account Aug 31Unknown; not statedAnthropic, Aug 31, 2026
The two most specific documents

Anthropic's Responsible Scaling Policy version 3.4 is the only document here that excludes a class of evaluator by financial dependence: reviewers must not be teams whose revenue, reputation and success depend entirely on Anthropic and similar companies. METR's conflict-of-interest policy, last updated August 28, 2026, is the only evaluator document that states its payment position outright.

04The typesFour kinds of arrangement

The 12 rows sort into four types, and each type can tell a buyer something different. None is better in the abstract; the question is what evidence each can produce.

Type 1
Commercial embedded evaluator
Accenture/Faculty at Anthropic

Deepest access on paper and the lab pays. Can tell you how decisions were made inside; cannot yet tell you anything in public, because no publication mechanism is stated and no access or reporting standard exists.

Lab-funded, inside
Type 2
Contracted independent lab
METR, Apollo, SecureBio, Irregular

Pre-deployment or report-level testing under contract or on the evaluator's own funds. Can produce a published assessment; what appears in public depends on the contract's approval clause, which ranges from none to written approval.

Paid or self-funded
Type 3
Government institute
UK AISI, US CAISI

Publicly funded, with access described by the lab rather than by a published agreement. Can tell you a state body tested the model; cannot show you the agreement, and neither institute's page claims power over releases.

State-funded
Type 4
Published external review
RSP §3.6 reviewers; METR's Risk Report reviews

A review of the lab's own risk documentation, published with the reviewer's disagreements. Can tell you whether the lab's reasoning holds up; cannot test the model itself, and runs after or alongside deployment.

Report-level

05The questionsThree questions for any vendor

Put these to any vendor whose page says independently evaluated, and match the answers to a row above.

Who paid the evaluator, and was any payment contingent on results?OpenAI answers the second half in writing; METR answers the first; most documents answer neither
Payer
What exactly did the evaluator get: weights, checkpoints, helpful-only variants, logs?Only the UK AISI × OpenAI account and METR's report list items; ask for the list, not the adjective
Access
Where can I read the evaluator's own report, and who approved it before publication?Unrestricted, approval-gated, or nowhere: all three appear in the census
Publication

How a lab measures oversight once a model is deployed is a separate question, covered in our post on agent oversight metrics; what gets disclosed when something goes wrong is in our incident-disclosure post. If your organisation is writing an AI procurement policy, our AI transformation service uses this census as the starting checklist.

06The gapWhat no arrangement here does

None of the 12 documents gives an outside party the power to stop a release. Anthropic's external review covers Risk Reports, and the policy only says the company will try to complete it before Board review. OpenAI's Preparedness Framework gives the deployment decision to its Safety Advisory Group, with external input as one part of the evidence. Google DeepMind's framework names internal governance bodies for pre-deployment review and no external evaluator at all. The government institutes describe testing, not gating; UK AISI's own about page describes testing systems before public release and collaborating with companies to improve safety. And none publishes a full agreement text; every government arrangement here is announced but not published. A buyer who wants a veto in the chain will not find one in these documents.

07How we built itMethodology

Methodology

A census of stated terms, not an assessment of any evaluator's independence or rigour, and not a ranking.

Population
Twelve arrangements involving Anthropic, OpenAI or Google DeepMind and an outside evaluator, where a primary document from the lab, the evaluator or the government body describes the terms. Meta appears only as a METR participant, and xAI only as a company whose public models were available to that exercise; no Meta or xAI safety framework was consulted. Academic evaluators are absent because no primary document establishing one's access, funding or publication rights was located.
Inclusion rule
A row needs a first-party document: a lab policy or announcement, an evaluator's own report or policy, or a government body's own page. Trade-press figures were not used for any cell; the $1 billion figures are Anthropic's own.
Classification rule
"Unknown" means the documents consulted do not address the question. "No such power stated" means the document describes the arrangement without granting it. Neither is a finding that the power or the fact does not exist elsewhere.
Dates
Each row cites its document's own date. The Anthropic policy is version 3.4, posted July 8, 2026. The OpenAI Preparedness Framework is version 2 of April 15, 2025, which OpenAI's May 28, 2026 governance framework states it does not supersede. The Google framework is version 3.1, effective April 17, 2026.
What was excluded
An automated extraction that returned a newer Preparedness Framework version number, which did not survive verification. Gemini model cards, which were not read. Any characterisation of an arrangement as independent or not independent beyond the document's own words.
As-of date
This page belongs to the September 20, 2026 batch; its sources were collected on September 22, 2026. Undated pages, including the UK AISI and CAISI sites and the METR and Apollo about pages, are cited as published at the time of writing.
Known limitations
Twelve rows from three labs is a census of the documented, not of the field. Access described in words cannot be verified against what was actually shared. This blog runs on Anthropic's models; Anthropic's rows are classified by the same rule as the rest.
Refresh
Extended when a lab or evaluator publishes an arrangement with stated terms, when Anthropic names the further evaluators its September 18 post promised, or when any government agreement text is published.

08Next stepIndependent describes the evaluator, not the arrangement

Put it into practice

Ask for the row, not the adjective, before you sign

When a vendor says independently evaluated, ask which of the four types it was, who paid, what the evaluator saw and where the report is. Write the answers into the procurement record beside the vendor's claim. If the answer to any of the three questions is the same word the marketing page used, you do not yet have an answer.

Digital Applied

Buy AI on stated terms, not on adjectives.

We turn vendor claims about evaluation, safety and testing into a checklist of documented terms, and help you write the questions into your procurement process.

Vendor evaluation checklistProcurement questionsGovernance mapping
Your next project

Start with three questions

  • Who paid the evaluator
  • What the evaluator actually saw
  • Where the report can be read
Questions and answers

Applying this post

This post reports the stated terms and does not rule on the word. Anthropic describes it as independent embedded evaluation, states that it funds Accenture's work directly, and states that no access or reporting standards yet exist. A reader can weigh those three facts; the census does not weigh them for you.
Digital Applied newsletter

Deep dives on AI, marketing and development.

Practical guides and fresh insights by email. No recycled takes.

Related dispatches

Continue reading

AI Development

What Evidence Should an AI Agent Leave Behind? A New Draft

Proof-of-Control, a 127-requirement draft standard open for comment, grades agent evidence by who you must trust. The tiers, the six domains, what to do now.

September 19, 2026 · 8 minRead
AI Development

Each AI Agent Step Passed the Rules. The Workflow Didn't.

A September 2026 paper names four ways an agent workflow breaks a policy while every step passes its own check. The types, worked examples and the fix for each.

September 17, 2026 · 7 minRead
AI Development

Is Anyone Watching Your AI Agents? Anthropic's Three Numbers

Anthropic proposes three oversight metrics for AI agents and reports its own: 30,000 agents, 100% monitored, 1 in 47,000 blocked. How to measure yours.

September 17, 2026 · 9 minRead
AI Development

Eight Worlds of AI Agents Faced Three Attacks: None Passed

Emergence AI ran eight worlds of ten agents for up to 21 days, then staged three attacks. No world passed all three. The scores, and three fixes for builders.

September 16, 2026 · 8 minRead
AI Development

AI Agent or Fixed Workflow? A Business Decision Table

Choose rules, an AI-assisted workflow or an autonomous agent by checking task uncertainty, verification and consequences with a business decision table.

September 13, 2026 · 5 minRead
AI Development

Which AI Agent Features Fall Outside Zero Data Retention

Compare 31 agent features across Anthropic, OpenAI, Google and AWS, with dated sources for retention, ZDR eligibility, contract scope and deletion.

September 8, 2026 · 9 minRead