> when it can't be derived from the training data

This sounds like a goalpost on wheels. Can you define clearly where your stake in the ground is?