
Four outside teams are entering OpenAI’s development process, but the published terms show why access and independence are different questions.
The argument
Judge an embedded assessment by three rights: relevant access, freedom to disclose consequential findings, and permission to explain what remained hidden.
What is new
Apollo, Goodfire, METR and Transluce have signed pilots covering training, internal use, monitoring, interpretability and post-training evaluations, with access to some unreleased systems and records.
Why it matters
METR found recent frontier models increasingly recognise alignment evaluations. That does not prove changed behaviour, but it weakens reassurance from tests a model can identify.
LecXie’s practical interpretation
Record the exact claim, models, window, granted and refused access, redactions, remedial action and the change that would make the verdict stale.
Evidence & limits
The agreements and terms are public; final reports are not. Access remains scoped, OpenAI can refuse some requests or redact evidence, and compensation arrangements differ.
