Given some of the things @Sirrian has told us about the AI, and the game’s code in general, it could be difficult to predict its behavior under all circumstances. The game is event-driven, and events may not complete in a consistent order, particularly under load. He presented the basic algorithm as a series of linear steps, but they may not actually occur that way in practice. The AI also has trouble making good choices about newer spells, because of limitations in the scripting system.
So I wouldn’t be surprised if changes that aren’t related to the AI could influence the outcome. Dead troops or status effects that haven’t been properly cleaned up may be inappropriately taken into account. Some part of the logic may be overriding the decision to take skulls, but failing its own check, allow the logic to drop down to taking a match-3. There have been multiple threads suggesting that the game doesn’t correctly identify match-4s under load, with the gems not matching, or them not granting an extra turn. How the game counts “blobs” of gems has been an issue on all platforms for a while now, so perhaps there are instances in which the game incorrectly identifies a potential match-4 as a match-3 instead, and thus passes over it.
As in all cases, the more pieces of the puzzle that are available, the more likely they can track it down. Video recordings of the whole battle, screenshots, or at least the team composition for each side.