The AI sometimes selects surprisingly bad moves. Then, when I switch on the practice mode to review the game, it turns out that it knew, at the same search depth, that the move was bad, often by a margin of 10 to 20 disks compared to the move it considered best. I ran into this with randomness at the lowest nonzero setting (1–2 disks) and with randomness turned off.

Unless it still drew on some sort of cached subtree that allowed it to be smarter during the review than during the actual game, this seems to me like there’s a bug in how the assessments of the engine are translated into moves.