Separating Distribution-Free and Mistake-Bound Learning Models over the Boolean Domain