I'm saying that the theoretical limit is 62%. There are two methods to see this (one a thesis paper using machine learning and one a blog post comparison of win distribution) and they came to the exact same results (the only one that it was exactly the same in all significant digits). Both are linked in the article. That's the theme for all of these estimates in the graphs; they are all verified from multiple people through multiple trials and experiments. There is some wiggle room, and the ratios will slightly change as the environment will change... but the basic relative relationship is what it is.