More accurately speaking, reduce the total inaccuracy contributed by all operations. The inaccuracy will be (generally) proportional to the number of operations

More accurately speaking, reduce the total inaccuracy contributed by all operations. The inaccuracy will be (generally) proportional to the number of operations, but keep in mind that some operation would be exceptionally inaccurate due to, for example, catastrophic cancelation. So do not normalize percentage into [0,1] if the input was already in [0,100] and an additional operation is needed for normalization. (By the way, cospi etc counts as a single instruction here which is why it is so good to have.)