Further we confirm these empirically finding no performance difference between SFT -> RL and just doing RL. Overall we show you don’t need SFT to make KV-streams work. Thus any agentic post-training…