Thank you for pointing this out! You are right that the total of fitness adjustments is not constant. It is exactly what you say. I missed this because I was focused on the relative accuracies totaling 1.0, and mistakenly took that for the total of the adjustments.
However, you have kindly shown the way out, too! The beginning of the next paragraph should have said, "Because the fitnesses estimate the relative accuracies, and because the relative accuracies sum to 1.0, the effect is that the various action sets within a given match set [M] have approximately the same total fitness."
At the first sentence of page 11, you mention that the total of fitness adjustments to the members of the Previous Action Set is constant, but I don't understand why. I guess the total is Beta * (1.0 - sum of (F sub j) in the previous action set), which does not seem to be constant. Since (F sub j) is an estimate of the relative accuracy, whose sum is 1.0, I suppose the discussion about mapping equal numbers of classifiers to each condition-action map is still valid though.