Perfect play, by definition, evens out. If everyone plays perfectly, then the ev should be zero (modified by position around the table). So it should balance to zero if everyone plays an equal number of hands from every position. The payoff would be adjusted by the rake.

But that is a wildly inaccurate assumption for any real table.

I still wouldn't call it perfect play because as soon as someone adopts a fixed strategy, the meta game shifts and it is no longer optimal.

If anything, the assumption is equal play, not perfect play

To your question above, the answer is better than all levels.

The meta idea is what I first thought. It’s hard to evaluate though. The ai does not adapt, but continues to beat pros, who are welcome to try and adapt.

I’m curious if there are some zany cheese strats against this kind of AI that work reliably but the answer may be no. The only really practical tool you have that could work is a crazy bluff, but I doubt that’s sustainable.

It doesn’t really matter if perfect play is considered though. Games should be rated assuming equal competence. It’s not misleading if in fact your turn out to be a worse player in a game of skill.

Perfect play would take into account the meta game as well.

My point is that if the casinos have to post odds, the only odds on a poker table are flat (any deviations from 0 coming from skill differences) minus any rake (with maybe some slight variation if the button doesn't go down an even number of times).

The fact that we cannot define perfect play is irrelevant.