The report
An independent evaluator can be allowed to test a frontier model while remaining contractually constrained in what it may tell the public. In an October 13 investigation, Reuters reported tensions over confidentiality, inspection access and provider approval of external evaluation findings. The accounts concern reported contract language and negotiations; a complete collection of signed evaluator agreements was not available for direct review.
The providers' account and the independent researchers' concern are not identical. OpenAI and Anthropic say sensitive intellectual property and security details need protection, and contest interpretations that would give them a blanket veto over all findings. Evaluators need enough publication independence to communicate adverse results and relevant methodological limits. Protecting a specific secret through a limited redaction is different from making the assessed company the final editor of every critical conclusion.
Two different tests of independence
First, could the evaluator obtain a sufficiently complete model, tools, logs and transcripts early enough to run the promised assessment? Second, could the evaluator publish its material results, including negative findings, on terms that do not depend on discretionary approval by the subject? The first is an access question and the second is a disclosure question. The October UK parliamentary session illustrates a related difficulty for a government lab that relies on companies and foreign-government permissions for access; it does not establish any particular Reuters-described contract clause.
The frontier evaluator Case preserves the competing positions and unresolved contract terms. The original Reuters page was unavailable to direct retrieval in this review, so additional contractual specifics have not been inferred from headlines. There is no supported finding that evaluations were falsified. The strongest next evidence would be dated versions of signed publication clauses, concrete redaction requests and records showing when the evaluator received model artifacts relative to release.