Frontier evaluation publication rights remain disputed despite commitments to independence

External evaluators need enough model access and tool traces to test consequences, but the independence of their final disclosure depends on the agreements that grant that access. Reuters examined contract language that restricted publication without company consent and reported negotiation over embedded evaluator supervision and cross-lab staffing. Anthropic and OpenAI publicly expressed support for independent results while describing protection for sensitive intellectual property.

This is a documented dispute about process, not proof that an evaluator fabricated a result or that either company holds a complete veto over every future report. The actual signed contracts, permitted redactions and transcript delivery timelines matter more than generalized assurances about independent safety reviews.

The October 13 UK hearing illustrates a distinct but adjacent constraint: the AI Security Institute needs permission to access model systems and cannot stop a release under its current remit. That says nothing by itself about whether METR or Redwood may publish unredacted transcripts under their own agreements. Access to the model, delivery of relevant tool traces, editorial control over findings and authority to compel a change must be separately established from the respective contracts or legal instruments. Combining them into a single binary claim of independent oversight would conceal the actual gaps.