I wonder if maybe each checkpoint is treated as the entrance to a new hedge maze in the code, so what's actually happening is that the lower-intensity AIs have more attainable path-finding goals because they're searching for the next "ride" (checkpoint), whereas the other AIs are only trying to navigate to the exit. From a path-finding perspective, the next checkpoint is always closer than the exit, so the low-intensity AIs end up having a better heuristic for proceeding through the maze.