I started reading this literature while building an AI scientist (Can an AI Scientist Start With a Dataset Instead of a Goal) that scores hypotheses by how much a test result moves the model’s belief.