I realize before you mix anything the result is 100% perfect, i mean the window/mask says it seen a 33 times come next, b 22, c 5, d 2....so during Arithmetic Encoding them you give the common ones smaller codes, almost optimal if the arithmetic coder was perfect, great! But we gotta normalize when mix and that has a solution i can probably cook up myself. The normalizing/rescaling/averaging is tedious since they differ! That causes issue I wager. Cus we make them into %s from counts and them mix them.