The existence of R1-zero is evidence against any sort of theft of OpenAI's internal COT data. The model sometimes outputs illegible text that's useful only to R1. You can't do distillation without a shared vocabulary. The only way R1 could exist is if they trained it with RL.