Self-consistency versus more thinking

Both strategies spend extra tokens on the same problem; they differ in shape. One high-effort call spends them in a single long chain where later steps can check earlier ones. Five low-effort calls spend them in independent short chains, run at the same time, and combine them only at the end by voting. The decision that matters most is what gets voted on: the normalised final answer, so "£2,500" and "2500" count as the same vote, and an unparsed reply doesn't vote at all.