It's not totally a mystery

https://arxiv.org/html/2604.24827v1

The short of it is by using hard facts knowledge that is difficult to compress, and then quizzing models on these facts and calibrating against a bunch of open models, you can kind of feel out the size of closed models.