You could train a classifier on all neat looking fortresses, then run outputs through the classifier. But you'd measure the cosine similarity to all previously-seen activations, and discard any that are too similar -- thus making sure that people need to come up with unique neat-looking fortresses.