The output head byte budget is surprising. Did you try any tradeoff where the head is compressed more aggressively but experts stay mostly untouched?