On September 18, 2026, the AI Evaluator Forum published a public letter titled “Minimum Conditions for Embedding Evaluators.” The letter’s page identifies it as having more than 100 signatories. Its publication is a documented advocacy event: the Forum proposed a framework for embedded third-party evaluations, but the letter does not establish that companies have adopted the framework or currently provide the access and protections it recommends.
What the letter proposes
The Forum’s letter recommends that embedded evaluators receive access equivalent to that of highly privileged employees for evaluation purposes. The proposed access includes relevant systems, data, tools and physical spaces, as well as candid communication with relevant staff, subject to protections for sensitive third-party data.
The recommendation addresses a basic institutional tension in embedded evaluation. Evaluators would need proximity to the systems and people they assess, while the Forum’s proposed conditions would require them to retain independence from the companies providing that access. The letter presents access not as a concession that companies may define unilaterally, but as one element of a wider arrangement intended to make evaluation possible without giving the evaluated company control over the evaluation’s conclusions.
The Forum also recommends that evaluators retain editorial control over their work. It calls for disclosure and mitigation of conflicts of interest, and recommends that evaluators not be owned or governed by the companies they assess, maintain significant commercial relationships with those companies, or receive payment or other rewards contingent on their findings. The letter further recommends incorporating multiple viewpoints and evaluation organizations.
These provisions are recommendations attributed to the Forum. They do not establish that any particular evaluator has those protections, that companies have agreed to them, or that the proposed safeguards would be sufficient in practice.
Transparency and oversight
For transparency, the letter recommends disclosure of evaluation methods, findings, access and contractual terms. It calls for prompt and unfiltered communication with company boards or other privileged oversight bodies. It also recommends public release of findings and supporting evidence, subject to limited, time-limited redactions.
The proposed transparency rules connect the evaluator’s relationship with an AI company to a broader oversight structure. In the Forum’s framework, an embedded evaluator would not report only through ordinary company channels: the letter recommends access to privileged oversight bodies and public disclosure of findings, while also preserving protections for sensitive third-party data. The arrangement described is therefore neither ordinary internal review nor unrestricted public access. It is a proposed combination of company access, evaluator independence and outside scrutiny.
The letter says embedded evaluations should complement, rather than replace, broader external oversight, public transparency and additional forms of access for independent researchers. It also recommends that such evaluations have scientific objectivity, transparency, independence and robust protection against interference by the companies being evaluated. Those statements describe the Forum’s proposed model; they are not independent findings that the model exists or that it has been implemented.
Protection, funding and institutional limits
The Forum’s proposed conditions include protection from retaliation, including retaliatory litigation. The letter also recommends funding arrangements that would allow evaluators to remain funded when their methods or findings are unfavorable to the companies being evaluated. In the proposed relationship, independence depends not only on formal editorial control but also on whether unfavorable work can continue without exposing evaluators to financial or legal pressure from the organizations under review.
The Forum describes itself as a network of organizations conducting independent, third-party AI evaluations. It says it is not a legal entity in its own right and relies on voluntary participation. The letter states that all signatories signed in their personal capacity and that listed affiliations are provided for identification only.
Those institutional details limit what the publication represents. The letter is a public proposal from a voluntary network, not a regulation, court order or binding standard issued by a legal entity. Its signatories’ affiliations also should not be read as evidence that their organizations formally adopted the letter’s conditions.
Why this enters the Record
The letter matters because it makes the governance problem of embedded evaluation explicit. Access to company systems can improve the evaluator’s ability to examine how an AI system is developed or operated, but proximity can also create questions about editorial control, confidentiality, funding, retaliation and the route by which findings reach oversight institutions and the public. The Forum’s intervention records one attempt to define the conditions under which those relationships could be treated as independent.
For a future account of AI governance, the publication provides a baseline of demands against which later evaluation arrangements could be described or compared: privileged access, multiple viewpoints, control over methods and findings, disclosure of terms, protection from interference, durable funding and continued external oversight. The verified record does not show that companies accepted those demands, that the proposed access exists, or that the conditions have been tested in practice. What it establishes is narrower but consequential: on September 18, a named evaluator network publicly set out the protections it regarded as minimum conditions for embedding evaluators inside AI companies.