Compare

The same question through several architectures at once. This is where the story stops being a straight line: on plenty of questions a cheap early stage already answers correctly, and one of the elaborate later ones talks itself out of the right answer.