Summary:
Apollo, METR, Transluce and Goodfire have different assignments. The key questions are what each can inspect, what remains restricted and what readers will eventually see.OpenAI’s November 16 announcement names Apollo Research, METR, Transluce and Goodfire as its first embedded assessors. The arrangement brings outside specialists closer to internal training and safeguards. It is a plan for investigation, not a completed safety verdict.
Four organizations, different assignments
OpenAI describes distinct areas of work:
- Apollo Research: investigate subversion risks across training, evaluations and internal use, including whether relevant reasoning can be detected.
- Goodfire: examine model internals for training monitoring, debugging and alignment, with internal findings and later public learnings.
- METR: assess alignment in a selected large reinforcement-learning run. A broader follow-up monitoring assessment still needs agreement on scope.
- Transluce: assess post-training evaluations. An approximately eight-week expansion in early 2027 is an intention, not a completed engagement.
OpenAI discloses that Apollo and Goodfire receive compensation unrelated to assessment outcomes. METR and Transluce do not receive assessment fees, though they may accept compute or API credits.
What Apollo’s pilot does and does not open
In its own November 16 account, Apollo describes six evaluators working for four months. It calls this a step toward full embedded evaluation, not unrestricted employee-like access.
The training assessment includes relevant code, datasets, reinforcement-learning records and queries to intermediate checkpoints. Access remains scoped; OpenAI may withhold material for technical feasibility, security or intellectual-property reasons, but not simply because it is unfavorable. Other assessment areas still require detailed access arrangements.
Apollo plans two public reports. Its claim-level verdicts cannot be redacted, but supporting confidential evidence can be. Readers may therefore see a conclusion without being able to inspect every supporting record.
Why explaining the limits matters
Apollo emphasizes its right to describe constraints: what it requested, what it received, what was withheld and how redactions affected its conclusions. This is what it calls meta-transparency. The right to explain a blind spot is different from the ability to remove it.
That distinction makes the eventual reports more useful. A reassuring result from a narrow assessment should not be stretched into a claim about every model, training run or deployment.
How to read the reports when they arrive
For teams choosing AI suppliers, our suggested checklist starts with the evidence boundary rather than the evaluator’s name:
- Scope: Which model versions, time periods and behaviors were examined?
- Access: Did evaluators inspect the records needed to test the claim, or only a selected subset?
- Disclosure: Which supporting findings are public, and which remain available only to oversight bodies?
- Follow-through: What changed after a problem was found, and was the change tested again?
Our earlier coverage of independent verification in agent workflows makes a related practical point: a proposed result and the test used to accept it are separate parts of the process.
Sources and image note
Sources: OpenAI’s assessment announcement and Apollo’s pilot terms, both November 16, 2026. The cover’s partly transparent archive is a conceptual illustration of scoped access, not an OpenAI facility or a depiction of the evaluators’ actual work.


