IQP circuits used as RL policies exhibit a trainability-simulability tradeoff: action decoders that protect initialization variance via low-weight Pauli components can paradoxically make policies classically simulable. We characterize this via exact variance laws and majority-...