If reinforcement learning deployment continues to gain relevance, how do we evaluate whether it changes user behavior in unexpected ways — and which early signals matter most?