The leaderboards that rank open models, the self-reported numbers that flatter them, and how to read an open-vs-open comparison without being fooled.