AI EVALUATOR FORUM

OLYMPUS opened an institutional file on the AI Evaluator Forum because it is the evaluators organizing as a bloc. The outside testers of frontier models, who sell their findings to no one in particular and are paid by a short list of philanthropic funders, governments and, in some cases, the personal holdings of lab employees, have written down the conditions under which they say an evaluation counts as independent. This is not a psychometric profile. It is a member list, a standard and a letter. The finding is the shape of the arrangement: the people who do the grading wrote the rule for how grading should be disclosed, a regulator is reported to have taken key provisions of it, and the funding that rule asks evaluators to disclose is the funding the members themselves receive. Not a cartel. A trade table, with no money of its own and no legal form.

Institutional Archetype

THE TRADE TABLE OF THE INDEPENDENT GRADERS — Subject describes itself as “a network of organizations conducting independent, third-party evaluations of AI systems” that “brings together practitioners to share methods, align on norms of professional practice, and inform the broader standards and policy ecosystem” (AI Evaluator Forum). It tests no model, licenses no one and holds no money. What it has is a standard, AEF-1, a checklist evaluators can publish beside their results, and a letter asking every frontier developer to embed outside evaluators on stated terms. The throughline is that the definition of an independent evaluation is now being written by the evaluators, in a document addressed to the labs, the regulators and the funders who pay them. This profile scores the table, not any member.

Mandate & Origin

• Announced December 2025, with the standard. The AEF-1 page is headed “AEF-1: Minimum Operating Conditions for Independent Third Party AI Evaluations” and dated “Version 1, updated December 4, 2025” (AI Evaluator Forum, AEF-1). A listing for “the public launch of the AI Evaluator Forum, a collaborative network of leading independent AI evaluation organizations working in the public interest” says that at the event “we will announce the forum” and “release a new standard for independent evaluations”; it is presented by Transluce, names Sarah Schwettmann, Conrad Stosz and Jacob Steinhardt as hosts, places the event in San Diego, California, and carries event data starting at 10:30 Pacific time on 4 December 2025 (Luma, AI Evaluator Forum). Princeton’s Center for Information Technology Policy posted on 12 December 2025: “A group of leading AI researchers announced the creation of the AI Evaluator Forum (AEF), a network of independent organizations assessing AI capabilities and risks in the public interest” (Princeton CITP). AVERI says of itself: “AVERI is a founding member of the AI Evaluator Forum and some AVERI staff helped write their first standard, AEF-1” (Miles Brundage, 15 January 2026).

• No legal form, no host. “The Forum is not a legal entity in its own right and relies on voluntary participation from its members” (AI Evaluator Forum). No fiscal sponsor, secretariat or host organization is named on the home, members, mission, path-ahead, support, contact or membership pages. The standard’s PDF names a Transluce address for correspondence: “Direct correspondence to Conrad Stosz: [email protected]” (AEF-1 PDF). The Forum’s own contact address is on its own domain.
• The mission, in its own words. The home page tagline: “Advancing the rigor, credibility, and impact of independent AI evaluations that serve the public interest” (AI Evaluator Forum). Its goals include to “Establish and promote best practices for credible independent AI evaluations and contribute to standards development” and to “Broaden the ecosystem of independent evaluators, so that credible oversight does not depend on any single organization” (AI Evaluator Forum, Mission). Its home page lists what it develops: “shared requirements for credible third-party evaluations: conflict-of-interest disclosure and management, funding relationships, recusal, depth of system access, and transparency about the terms of an evaluation” (AI Evaluator Forum).
• The members. Eight: the AI Verification and Evaluation Research Institute (AVERI), the Collective Intelligence Project, Meridian Labs, METR, the Princeton Holistic Agent Leaderboard (HAL), RAND, SecureBio and Transluce. “Membership is limited to organizations that conduct and publish rigorous independent technical evaluations of general-purpose AI systems in the public interest,” and the page adds “More members coming soon” (AI Evaluator Forum, Members). A membership form asks applicants for their legal form, from “Nonprofit” to “FFRDC / national lab” to “Government,” and for-profit evaluators are among the options (AI Evaluator Forum, Membership).
• Leadership. The members page lists “Conrad Stosz” as Chair and “Patricia Paskov” as Senior Advisor (AI Evaluator Forum, Members). Stosz is Head of Governance at Transluce; Paskov is Standards Director at AVERI, “also an Adjunct Researcher at RAND, a Research Affiliate at the Oxford Martin AI Governance Initiative” (AVERI, Team). No other officer, board or staff is listed.
• The standard. AEF-1 is “a standard and checklist that third-party evaluators can use to demonstrate how they achieved a set of operating conditions that support a baseline level of independence, access, and transparency.” It has five principles: Sufficient Access and Resources, Minimized Conflicts of Interest, Analytic Autonomy, Transparent Methods and Results, and Protection of Sensitive Information. The PDF describes itself as “an initial voluntary standard” and lists 26 authors from Transluce, METR, AVERI, SecureBio, the Oxford Martin AI Governance Initiative, Princeton, MIT, Stanford, GovAI, the Collective Intelligence Project, Meridian Labs, Epoch AI and UC Berkeley; no RAND affiliation appears in the author list (AEF-1 PDF; AI Evaluator Forum, AEF-1). “Evaluators demonstrate adherence by completing the checklist and publishing it alongside their results. Where a requirement cannot be met literally, the checklist is used to document which conditions were not met, and why.”

• The European endorsement, on the Forum’s own account. “The EU AI Office has endorsed key provisions of the AEF-1 standard as a means for providers to comply with the independence provisions of the General-Purpose AI Code of Practice” (AI Evaluator Forum, AEF-1). The Commission’s own call for a workshop on external evaluators, published 12 June 2026, concerns “independence and qualification requirements that organisations external to GPAI model providers should meet” and does not name AEF-1 or the Forum (European Commission, call for participants). The endorsement stands here as the Forum’s statement about the Commission’s act.

• The letter, 18 September 2026. “Published September 18, 2026 · 200+ Signatories,” titled “Minimum Conditions for Embedding Evaluators.” It opens: “We, the undersigned, are encouraged to see frontier AI companies call for embedding third-party organizations to evaluate rapidly escalating AI capabilities and risks,” and says that “all signatories sign in their personal capacity. Affiliations are listed for identification only” (AI Evaluator Forum, letter). The first notes link posts on X by the accounts of Dario Amodei, Demis Hassabis, Elon Musk, Sam Altman and Satya Nadella. IBTimes reported: “The coalition, organized by the AI Evaluator Forum, published a public letter Friday” (IBTimes, as reported); at publication the count it reported was “more than 100.”

Funding & Backers

• The Forum says it has none. “Since its launch, the AI Evaluator Forum has operated without funding, sustained entirely by voluntary participation of its members. We are now launching an open call for independent funding from a broad range of supporters, both to meet the Forum’s near-term needs and to explore sustainable mechanisms for funding independent evaluators more broadly.” The page adds: “A diversified funding base protects evaluators’ independence as we grow to meet the demands of the third-party evaluation ecosystem” (AI Evaluator Forum, Support). No funder, grant, fee or in-kind donor is named on any Forum page. The cost the Forum carries, a website, a standard, a letter and a launch event, is paid by members and by their own funders.
• The pooled-funding plan. “The Forum will advance mechanisms to fund public-interest evaluators, such as pooled funding, that let them conduct highly rigorous evaluations at the necessary scale and speed, without compromising their impartiality. AEF members will collaborate on these approaches to help grow a pluralistic ecosystem of independent evaluators that keeps pace with multi-trillion-dollar companies” (AI Evaluator Forum, Path Ahead). The page gives no amount, funder, timeline or vehicle.
• The members’ money is the Forum’s money by proxy. What each member publishes about its own funding, from its own pages:
  • Transluce: “Transluce received 6% of its revenue from the personal holdings of OpenAI employees” and “32% of its revenue from the personal holdings of Anthropic employees,” with no revenue directly from developers as organizations; it says its policy is consistent with requirement 2.3 of AEF-1 (Transluce, policy). Its fundraiser says earned revenue “constitutes approximately 20% of our funding this year, and will likely be higher in future years,” and that it serves “as a contractor to the EU AI Office” (Transluce, fundraiser).
  • AVERI: “AVERI’s funding comes from Halcyon Futures, Fathom, Coefficient Giving, Geoff Ralston,” “Craig Falls, Good Forever Foundation, AI Underwriting Company, Sympatico Ventures,” and “several non-executive employees and alums of frontier AI companies”; “No donor makes up a majority of our funding” (Miles Brundage, 15 January 2026). Its about page says: “we do not accept cash donations or compensation from frontier AI companies or their executives,” and also: “We’ve received API credits as donations or compensation from Amazon, Anthropic, Google DeepMind, Microsoft, OpenAI, and Thinking Machines Lab” (AVERI, About).
  • METR: “METR has not accepted funding from AI companies, though we make use of significant free tokens”; its about page lists Schmidt Sciences and the Survival and Flourishing Fund among its funders and says “METR cannot accept donations made by or at the direction of frontier AI company employees”; “a small part of our income is from a technical assistance contract with the European AI Office” (METR, About). The Audacious Project “catalyzed approximately $38 million of funding for Canary,” a collaboration with RAND (METR, 9 Oct 2024).
  • SecureBio: “We are primarily funded by Coefficient Giving, the Survival and Flourishing Fund, and Sentinel Bio,” with “SecureBio received compensation from OpenAI limited to covering the costs of this assessment” (SecureBio, GPT-6 Astra assessment); its policy caps AI-company revenue: “no more than 25% of its annual revenue is derived from services supplied to AI companies” (SecureBio, conflicts of interest).
  • Princeton HAL: “HAL is funded by Coefficient Giving, Schmidt Sciences,” and “We are grateful to OpenAI and Google for providing API credits to evaluate their models” (HAL).
  • Collective Intelligence Project: “Our work is made possible through the support of committed philanthropic partners”; the funders page lists the Robert Wood Johnson Foundation, Google.org, the Ford Foundation, Omidyar Network, OneProject, the Future of Life Institute, the Amaranth Foundation and the Survival and Flourishing Fund, and names Anthropic and Mistral AI among partners (CIP, Funders and Partners).
  • RAND: Open Philanthropy (now Coefficient Giving) “gave a gift of $10 million over three years to RAND in support of its work on Canary, an effort to develop advanced methods for evaluating AI systems, test systems for dangerous capabilities, and share the findings to help inform evidence-based policies” (Coefficient Giving, grants search).
  • Meridian Labs: no funder is named on its pages; its Petri 3.0 post thanks Anthropic “for entrusting its development to us” (Meridian Labs, Petri 3.0).

Institutional Voice & Intent

The register is procedural and consensual, and it moved between two documents. The standard (December 2025) speaks as a trade body setting terms: “conditions,” “requirements,” a checklist. The letter (September 2026) speaks as a coalition addressing the labs: “all frontier AI companies should embed evaluators,” with access “equivalent to that of their own highly privileged employees” (AI Evaluator Forum, letter). Both give the limit of their own reach. AEF-1: “The success of AEF-1 requires adoption by a range of stakeholders, including frontier developers” (AI Evaluator Forum, AEF-1). The letter: “Embedded evaluations cannot address all oversight needs and should be treated as a complement to, rather than a replacement for, broader efforts by frontier AI companies to expand external oversight.” Stosz to the Associated Press, in the weeks before the letter, said “even the forum has questions about what Amodei and OpenAI CEO Sam Altman want” (AP via PBS NewsHour).

Stated intent: Advance “the rigor, credibility, and impact of independent AI evaluations that serve the public interest”; set shared requirements for conflicts of interest, funding relationships, recusal, access and transparency; broaden the field so that “credible oversight cannot rest on a handful of organizations” (AI Evaluator Forum).

Observed intent: A trade body that has produced three things in ten months: a 26-author standard, a checklist that three members have attached to their own published evaluations (Transluce’s mental health evaluation, August 2026; METR’s Frontier Risk Report, May 2026; SecureBio’s GPT-6 Astra assessment, September 2026), and a letter addressed to the labs (AI Evaluator Forum, AEF-1). The METR entry describes its report as “carried out with more direct access to non-public information and more editorial independence than previous external evaluation engagements.”

Gap: The standard asks for disclosure of the conditions under which an evaluator worked. It names, as a category to disclose, that “a meaningful fraction of the evaluator’s funding (either revenue or donations) comes from a system provider, its employees, or its direct competitors” (AEF-1 PDF), and it requires that “The system provider does not control any of the evaluator’s board seats.” The members’ own pages show that category in use: Transluce’s 38 percent from OpenAI and Anthropic employees’ holdings, AVERI’s non-executive lab staff and alumni, HAL’s API credits, SecureBio’s cost-covering compensation. The letter goes further than the standard on one point, saying embedded evaluators “should not accept any form of payment or other reward contingent on the evaluator’s findings” and should not “have other significant commercial business with” the companies. Whether the standard or the letter is the Forum’s position is not stated; the letter calls the standard an example, and says “far more work will be necessary.” What the record can settle is who disclosed what; it cannot settle whether disclosure is sufficient, and the finding does not need it.

Interest: What each position would pay the members who argue it, documented, with the Forum’s own best case. The positions are the Forum’s; no member’s file is the author of them.

1. Position: embedded third-party evaluators inside every frontier AI company, with employee-equivalent access. The letter: “all frontier AI companies should embed evaluators to independently assess AI risks” (AI Evaluator Forum, letter).
   • What adoption would pay the members. A rule that labs admit outside evaluators is demand for what the members sell or publish: Transluce says revenue from companies and governments “will likely be higher in future years”; SecureBio’s policy allows up to a quarter of revenue from AI companies; METR’s report cites the deeper access as the difference from earlier engagements. No fee or contract value is published for the embedded arrangement itself (Transluce, fundraiser; SecureBio, conflicts of interest).
   • The Forum’s best case, steelmanned. Embedded evaluation is credible only with independence, transparency, protection from retaliation and access, and “should be treated as a complement to, rather than a replacement for” broader external oversight.
   • Who else gains from the same rule. Anthropic, whose chief executive proposed it: “Each frontier AI company commits to giving ongoing, employee-like access to a team of embedded third-party evaluators (such as METR)” (Dario Amodei); the other labs whose chiefs the letter’s footnotes link; and evaluators outside the Forum, among them Apollo Research, FAR.AI and GovAI staff, listed as signatories.
2. Position: a published minimum standard for evaluator independence, adopted voluntarily and referenced by regulators. AEF-1, with the EU AI Office reported by the Forum to have endorsed key provisions.
   • What adoption would pay the members. The standard sets a bar their own reports already clear: three members’ evaluations carry the checklist, Transluce’s policy says it is consistent with requirement 2.3, and SecureBio “uses AEF-1 as a foundation of its organizational principles” (AI Evaluator Forum, AEF-1; Transluce, policy). A regulator’s reference to the standard turns members’ practice into the compliance route for the labs they test.
   • The Forum’s best case, steelmanned. “The operating conditions of third-party evaluations can be opaque, even though they can significantly impact whether an evaluation is trustworthy and impartial”; a checklist makes the conditions readable. The PDF adds that “the checklist could be used by developers to report details of third-party evaluations” under the EU Code of Practice and California’s Senate Bill 53.
   • Who else gains from the same rule. Developers that need a documentation format for third-party testing under those two regimes; regulators that want a reference text without commissioning one; the Frontier Model Forum, whose April 2025 “Technical Report on Third-Party Assessments” the PDF cites as reference 2.
3. Position: pooled funding for public-interest evaluators. “Mechanisms to fund public-interest evaluators, such as pooled funding” (AI Evaluator Forum, Path Ahead).
   • What adoption would pay the members. A pooled mechanism would pay the Forum’s own members; the page names no contributor. The Forum itself currently has no funding, so the first beneficiary of an open call is the Forum.
   • The Forum’s best case, steelmanned. “A diversified funding base protects evaluators’ independence”; the letter asks for “funding mechanisms that give them confidence they will remain funded even in these cases,” meaning when findings displease the company.
   • Who else gains from the same rule. The philanthropic funders already behind several members (Coefficient Giving, the Survival and Flourishing Fund, Schmidt Sciences), who would fund one table instead of eight grantees; and the labs, which would cease to be the default payer.
4. Position: a talent pipeline for evaluation organizations. The Forum is “committed to expanding the field of independent AI evaluation and ensuring its member organizations can compete with frontier companies for top talent” (AI Evaluator Forum, Path Ahead).
   • What adoption would pay the members. Staff for member organizations, from a pipeline the Forum would develop and members would draw on.
   • The Forum’s best case, steelmanned. The labs hire the people who could evaluate them; “multi-trillion-dollar companies” outbid nonprofits.
   • Who else gains from the same rule. Training programmes and funders already in the field; the labs, which would see their evaluators better staffed.

Position in the Apparatus

• The graders’ trade body, beside the labs’ and the governments’. Three tables now write rules about frontier evaluation. The Frontier Model Forum is the six labs’, funded by their fees; the International Network for Advanced AI Measurement, Evaluation and Science is ten governments’; the Forum is the evaluators’. They cite each other. AEF-1 cites the Frontier Model Forum’s April 2025 technical report on third-party assessments and its issue brief on frontier safety frameworks, and refers to “the International Network of AI Safety Institutes” among government testing bodies (AEF-1 PDF). The government network’s best-practice document is addressed to the same third-party evaluators (AISI, 23 Jul 2026). NIST says “The International Network was founded by the Center for AI Standards and Innovation (CAISI) in November 2024,” the institute Stosz previously led; NIST says CAISI hosted Network members “in San Diego on the sidelines of NeurIPS” in December 2025, the same month and city as the Forum’s launch (NIST, 13 Feb 2026).
• What it sits above. The evaluators mapped in The Evaluator Alliances and The AI Evaluators: the Forum is the part of that field that wrote down terms, and it names eight of them. The ones outside it include Epoch AI, whose chief executive Jaime Sevilla is a listed AEF-1 author, and GovAI staff, listed as authors and signatories, without either organization being on the members page.
• Who sits where, member by member, and which other alliances on the map each carries.
  • METR: supporters including Schmidt Sciences, the Survival and Flourishing Fund, Geoff Ralston (also an AVERI funder) and “the AI Security Institute”; “a technical assistance contract with the European AI Office”; membership of the NIST AI Safety Institute Consortium; the Canary collaboration with RAND; free tokens from the labs it tests. Its president, Chris Painter, and CEO, Beth Barnes, are AEF-1 authors (METR, About; AEF-1 PDF).
  • Transluce: the Chair’s employer, which presented the launch event; contractor to the EU AI Office; 38 percent of revenue from OpenAI and Anthropic employees’ holdings; Percy Liang among its advisers; board member Mike McCormick, CEO of Halcyon Futures, who is also on AVERI’s board (Transluce, About; AVERI, Team).
  • AVERI: funders Halcyon Futures, Fathom, Coefficient Giving and others; board members McCormick and Grace Werner, an AEF-1 author; executive director Miles Brundage, an AEF-1 author and letter signatory; Geoff Ralston is also a named METR donor; the Forum’s Senior Advisor is its Standards Director (AVERI, Team).
  • SecureBio: funders Coefficient Giving, the Survival and Flourishing Fund and Sentinel Bio; a grantee of the Frontier Model Forum’s AI Safety Fund, in the cohort announced 11 December 2025, a week after the Forum’s launch; “a partner to the European AI Office”; compensated by OpenAI for the cost of the GPT-6 Astra assessment (Frontier Model Forum, 11 Dec 2025; SecureBio).
  • RAND: Canary, funded by a $10 million Open Philanthropy gift and shared with METR; Paskov is an adjunct researcher; no RAND-affiliated author is on the AEF-1 PDF and no RAND-affiliated name appears among the letter’s listed signatories (Coefficient Giving; RAND, Paskov).
  • Meridian Labs: its flagship, “Inspect AI, an evaluation framework built by Meridian’s founding team in collaboration with the UK AISI,” is described on its own page as “now the standard framework for government organizations including UK AISI, US CAISI, EU AI Office”; co-founder J.J. Allaire is “a visiting researcher at the US CAISI”; CEO Charles Teague is an AEF-1 author and letter signatory; no funder is named (Meridian Labs, About; Meridian Labs, Team).
  • Princeton HAL: funded by Coefficient Giving and Schmidt Sciences, with OpenAI and Google API credits; Stanford’s Percy Liang is among its listed authors, as he is among Transluce’s advisers; Princeton’s Arvind Narayanan and Sayash Kapoor are AEF-1 authors (HAL).
  • Collective Intelligence Project: funders include Google.org and the Survival and Flourishing Fund; Anthropic and Mistral AI listed as partners; two staff, Divya Siddharth and Faisal Lalani, are AEF-1 authors and its executive director, Joal Stein, a letter signatory (CIP; AI Evaluator Forum, letter).

• The overlap, counted. Coefficient Giving (formerly Open Philanthropy) funds AVERI, HAL, SecureBio and RAND’s Canary: four of eight. The Survival and Flourishing Fund funds METR, SecureBio and the Collective Intelligence Project: three of eight. Schmidt Sciences funds METR and HAL. Halcyon Futures funds AVERI, and its CEO, McCormick, sits on the boards of AVERI and Transluce. Geoff Ralston is a named donor to both AVERI and METR. The labs, jointly, fund one member through the AI Safety Fund and cover costs or credits at three others. The same few funders, then, sit behind the tables that write the rules on both sides of the evaluator relationship, and the Forum’s standard asks evaluators to disclose exactly this kind of money.

Actions & Leadership Choices

Read by deeds, the Forum is a body that has published a standard, attached it to its members’ reports and written a letter. The conduct is the conduct of its members, in turn.

Actual founding purpose. To give the independent evaluators a common set of terms before a regulator or a lab supplied its own. The launch page promised to “release a new standard for independent evaluations”; the standard came the same day (Luma; AI Evaluator Forum, AEF-1). As coordination it produced output at once: a 26-author PDF, a checklist template, a three-entry list of evaluations carrying the checklist.

• AEF-1, Version 1 (4 December 2025). Five principles and their conditions, from “Technical access,” “Time” and “Safe harbor” to “Contingent compensation,” “Organizational control,” “CoI policy,” “CoI disclosure,” “Recusals” and “Separate agreements,” “Editorial control,” “No contingent release,” “Redactions,” and “Responsible disclosure policy” (AI Evaluator Forum, AEF-1). Requirements are marked as requirements or recommendations in the PDF; a condition that cannot be met is documented in the checklist, not excused.
• Evaluations carrying the checklist. Transluce’s mental health evaluation of 77 model variants (August 2026); METR’s Frontier Risk Report, an assessment of “misalignment risk from AI agents deployed internally at Anthropic, Google, Meta, and OpenAI” (May 2026, checklist at page 67); SecureBio’s pre-release assessment of OpenAI’s GPT-6 Astra (September 2026) (AI Evaluator Forum, AEF-1). All three members publishing a checklist are among the AEF-1 authors’ organizations.
• The letter (18 September 2026). Six conditions for embedding evaluators: independence and full editorial control, differing viewpoints and multiple evaluators, transparency, protection from retaliation, access “equivalent to that of their own highly privileged employees,” and a closing paragraph that the list “is not comprehensive.” Notable signatories include Geoffrey Hinton, Stuart Russell, Arvind Narayanan, Percy Liang, Steinhardt, Schwettmann, Stosz, Brundage, Teague, Joal Stein, Adam Gleave and, by listed affiliation, GovAI, Apollo Research and SecureBio staff (AI Evaluator Forum, letter).
• The path ahead. Three lines of effort in “the coming months”: shared norms for embedded evaluators with “strengthened provisions on conflicts of interest,” pooled funding, and a talent pipeline; it invites a “broader open consortium of practitioners, academics, and researchers” (AI Evaluator Forum, Path Ahead).

Leadership choices. The leadership is two names. The Chair is employed by a member that also hosted the launch and whose address appears on the standard. The Senior Advisor holds a standards seat at one member, an adjunct post at another and an affiliate post at a university group whose name sits on the standard. The Forum names no board, staff, treasurer or host. Whoever carries the work is carried by the members’ employers.

CONDUCT: STANDARD-SETTER — EIGHT GRADERS WRITING THEIR OWN CONDITIONS OF INDEPENDENCE, WITH NO BUDGET OF THEIR OWN. The Forum is not hostile to anyone, and its standard is technically specific. It is a table at which the outside evaluators agree on how an evaluation should be disclosed, which regulators are told is the route to meeting a Code of Practice, and which is addressed to the labs that pay or host some of the members’ work. The finding is that the reference text for “independent evaluation” is being written by the independent evaluators, who disclose their funders in the same document that asks the question. The reach is the finding; the malice is not asserted.

Alliance Map

These run at the same time. One table with no legal form, no stated funding and no named host writes the terms of independence for eight evaluators; four of them take money from the same philanthropic funder, three from a second, and several carry lab-employee money or lab credits that the standard asks them to disclose; one is paid by the labs’ safety fund; a regulator is said to have endorsed key provisions; and the Chair previously led the federal institute that founded the government network writing guidance for the same evaluators. Each tie is stated by the Forum, a member, a funder or an agency. The overlap is the finding.

Reach Assessment

Institutional: Moderate. Its reach is the reach of a reference: a standard with a checklist, used by three members, endorsed in part by an EU body on the Forum’s account, and addressed to developers who have not been shown adopting it. It cannot require a test, a release or a payment.

Memetic: High for its size. It supplies the vocabulary for the argument over embedded evaluators: “minimum operating conditions,” independence, access and the checklist. The letter put the standard’s terms in front of the public after the labs’ own proposals.

Civilizational: Moderate, with low durability. The body builds no model and tests none. It writes how independence is described, and the grade a widely deployed model receives is shaped upstream by who counts as an independent grader. The hazard is the reach of a rulebook written by the graders, held by a table with no legal form, no budget and shared funders. The reach is the finding; the malice is not asserted.

Score move: none (new profile). Calibrated against the International Network for Advanced AI Measurement, Evaluation and Science (5/7/5, hazard 57), the Frontier Model Forum (9/6/7, hazard 79), Transluce (6/7/6, hazard 52) and METR (9/8/7, hazard 84), with Apart Research (5/5/5, hazard 50) and Epoch AI (5/7/7, hazard 64) as the nearest reference-setters.

Score basis. Atk 5 is reach: on the Forum’s account the EU AI Office “has endorsed key provisions of the AEF-1 standard as a means for providers to comply with the independence provisions of the General-Purpose AI Code of Practice,” three members’ reports carry the checklist, and the letter lists “200+ Signatories” (AI Evaluator Forum, AEF-1; AI Evaluator Forum, letter). It is not higher because the Forum “is not a legal entity in its own right,” the PDF calls itself “an initial voluntary standard,” the standard’s own page says its success “requires adoption by a range of stakeholders, including frontier developers,” and the evaluations listed as using it are members’ own. It matches the Network at 5, whose best-practice document is also written for third-party evaluators, and sits far below the Frontier Model Forum’s 9, where six member firms signed an agreement covering vulnerabilities and capabilities of concern. Def 6 is funding insulation: the Forum “has operated without funding, sustained entirely by voluntary participation of its members,” so no funder can withdraw from it (AI Evaluator Forum, Support). It is not higher because the members’ money is concentrated: Transluce reports 6 percent and 32 percent of revenue from OpenAI and Anthropic employees’ personal holdings, AVERI lists non-executive employees and alumni of frontier AI companies among its donors, HAL thanks OpenAI and Google for API credits, and four of eight members have a shared philanthropic funder, with the standard asking for disclosure of this category and not for its absence (Transluce, policy; Brundage; HAL). It sits level with the Frontier Model Forum at 6 for opposite reasons, that table’s six payers are its members and the Forum has none, and one point under the Network’s 7, whose members are each paid by a state. Hp 4 is entrenchment: Version 1 of 4 December 2025 is a published, 26-author text with a regulator’s reported endorsement, and a letter and a checklist are in circulation (AI Evaluator Forum, AEF-1). It is not higher because there is no legal personality, no stated funding, no staff or host named, a Chair who is an employee of a member, and a membership of eight that “relies on voluntary participation”; removing it costs a website and a PDF, which is why it sits one under the Network’s 5, whose ten members are governments, and three under Epoch AI’s 7, whose benchmark is installed in the labs’ own reports. Peers: Apart Research at 5/5/5, hazard 50, matches the hazard with a different shape, more entrenchment and less insulation; Transluce at 6/7/6, hazard 52, is the member that chairs it and scores higher on all three because it has a lab-facing evaluation business and a funded staff. The values hold at 5/6/4, hazard 50.

Sources: AI Evaluator Forum; AI Evaluator Forum — Members; AI Evaluator Forum — Mission; AEF-1 Standard; AEF-1: Minimum Operating Conditions for Independent Third Party AI Evaluations (PDF); AI Evaluator Forum — Public Letter, Minimum Conditions for Embedding Evaluators; AI Evaluator Forum — The Path Ahead; AI Evaluator Forum — Support; AI Evaluator Forum — Membership; Luma — AI Evaluator Forum; Princeton CITP — HAL is part of the Forum; Miles Brundage — The Launch of AVERI; AVERI — About; AVERI — Team; Princeton HAL; Collective Intelligence Project — Funders and Partners; Meridian Labs — About; Meridian Labs — Team; Meridian Labs — Introducing Petri 3.0; RAND — Patricia Paskov; Coefficient Giving — grants search, RAND; METR — About; METR — New Support through the Audacious Project; Transluce — Independence and Transparency Policy; Transluce — Fundraiser; Transluce — About; SecureBio — AI; SecureBio — Conflicts of Interest; SecureBio — GPT-6 Astra assessment; Frontier Model Forum — New AI Safety Fund Grantees, 11 Dec 2025; Frontier Model Forum — AI Safety Fund; NIST — International Network for Advanced AI Measurement, Evaluation and Science, 13 Feb 2026; UK AISI — Inspect Evals; UK AISI — international evaluation best practice, 23 Jul 2026; European Commission — call for participants, external evaluators workshop; Dario Amodei — We Must Pace the Frontier; AP via PBS NewsHour, 28 September 2026; IBTimes — More Than 100 AI Experts Sign a Letter.