“The emergence of Multi-teacher On-Policy Distillation (MOPD) is partly because mixing math, code, and agentic RL in a single training run leads to capability trade-offs.”