I'm not convinced that improving calibration will not improve accuracy because predictions are often nested within other predictions. For example, suppose we are trying to make a prediction about P, and the truth or falsity of Q, R and S are relevant to the truth of P in some respect. We might use as a basis for guessing P that we are ninety five percent confident in our guesses about Q, R & S, (suppose the truth of all three would guarantee P). Now suppose we become less confident through better calibration and decide there is only a 70% chance that Q, a 70% chance that R and a 70% chance that S, leading to a compound probability of less <50%. Thus overall accuracy can be improved by calibration.