“Chinese AI labs commonly use distillation by "jailbreaking" the APIs of models like Claude and GPT to extract reasoning tokens for use as SFT or mid-training data.”