By carefully coordinating off-the-shelf models with inference only, we show the DSP framework can achieve surprisingly good results in theorem proving, comparable to the frontier models with RL-based large-scale training.