[mapt]

Color me skeptical. The value of things like Elo is they provide a ramping scale with some degree of statistical significance, because they cover play over many games. I don't think a computer is going to be able to extrapolate with a high degree of confidence, because human play is variable from game to game and long-run strategies are non-obvious constructs for the computer.

You're thinking in terms of 'The AI has solved Go mathematically', but that's not the case; Just because you can run a Monte Carlo best-choice-picker/guesser algorithm doesn't mean you can meaningfully rank how deliberate choices compare with each other more than a few plays away.

[eru]

Long-run strategies are a human crutch. It's easiest to see when you solved a game mathematically, that you can just value positions independently.

Go hasn't been solved to that level, but it's apparently been solved to higher level than humans ever reached.

I am just parroting http://www.uschess.org/content/view/12677/763 here, so I might as well quote:

"To catch an alleged cheater, Regan takes a set of chess positions played by a single player—ideally 200 or more but his analysis can work with as few as 20—and treats each position like a ques­tion on a multiple-choice exam. The score on this exam translates to an Elo rating, a score Regan calls an Intrinsic Perfor­mance Rating (IPR)."

This approach also allows to score historic players absolutely, instead of only relatively and trying to find sets of overlapping lifetimes until we reach the modern age.