Hmmm, interesting theory.
I ran it against Mountebank (that's an optimized bot that buys two Mountebanks, btw), and it shows a 78.46% win rate for NBx2 with an average game length of 30.1 turns.
So here are 10 Sample games in a row:
Game 1: NB LOSES, 28 turns.
Game 2: NB LOSES, 20 turns.
Game 3: NB LOSES, 19 turns.
Game 4: NB LOSES, 27 turns.
Game 5: NB LOSES, 23 turns.
Game 6: NB LOSES, 23 turns.
Game 7: NB LOSES, 25 turns.
Game 8: TIE, 21 turns.
Game 9: NB WINS, 26 turns.
Game 10: NB LOSES 20 turns.
That certainly decent evidence for DG's hypothesis that there's something different about the sample game play and the multi-game play. And, on the face of it, I'd tend to agree with DG that it's likely that the sample game play is correct and the multi-game play is incorrect, because... NB? Beats Mountebank? It doesn't seem likely.
Geronimoo, can you investigate?