Comment by enugu

> Either the eval maintainers need to be given the closed source models (which will likely never happen)

Given that the models are released to public, the test maintainers can just run the private tests after release, either via the prompts or via an api. Cheating won't be easy.