Personally, if I can’t go from human readable data to a complete model then I don’t consider it open source. I understand these companies want to keep the magic sauce thats printing them money but all the open source marketing is inherently dishonest. They should be clear that the architecture and the product they are selling is separate, much like proprietary software just has all the open source software they used as a footnote in their about screens.
It would be luck based for pure LLMs, but now I wonder if the models that can use Python notebooks might be able to code a script to count it. Like its actually possible for an AI to get this answer consistently correct these days.