"Novel" here simply means that the exact sequence of moves that is the solution cannot possibly be in the training set (mutatis mutandis). You can easily write a program that generates these kinds of puzzles at random, and feed them to the model.