Agnostic proper learning of monotone functions: beyond the black-box correction barrier