That is, in certain cases there is no way to post-process a multiaccurate predictor to get a weak learner, even assuming the best hypothesis in C has high correlation. However, when we also require p to be calibrated, we recover not just weak, but strong agnostic learning.