What properties of reasoning supervision are associated with improved downstream model quality?