How do you validate whether the fortress was “neat-looking”?

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.

Today's stupid ML idea has been brought to you by: Gabe Newell. Inspiring future ML devs since 1996. A word from our sponsor: https://www.youtube.com/watch?v=jpw2ebhTSKs&ab_channel=TheCh...

But then anyone else can do that and write a neat fortress generator trained to produce outputs that pass the neat fortress classifier.