I am not sure the "get 5 duchies before first duke" idea is bad. Consider if you are playing against a Province player, you will be contested for the Duchies, but not for the Dukes. I am sure I could find decent players recommending it if I searched enough.
Of course, this does suffer from being unconditioned on opponent's strategies. There could be two very different paths depending on the opponent's strategy, and what you see is a probability weighted mix over all opponent's strategies. So really you might want to go in one direction vs one kind of opponent, and a totally different direction vs another, and you'll see some muddled incoherent mixture that does well in the this output.
The no second potion thing MUST be wrong! I usually like getting the second potion in Alchemist games.
Of course, I don't recommend just blindly trusting the data, but it could be useful to think about. Also, optimal play is sufficient but not necessary for the stats to tell the right thing. You need that the players play equally badly conditioned on their state. Then there is no bias since the skill problems cancel. I think this condition will be violated for things like Ambassador (I'd bet on a randomly picked player who went 2 Ambassadors vs a completely random player), but should usually help the data from not being too off.