ai's reasoning gap: why llms fail at simple math when details change