Highlights faithfully attributed to two fictional select-predict models on an elaborate Amazon Reviews sentiment classification example. Although highlight (a) is easier to understand, it is also far less useful, as the selector clearly made hidden decisions.