A constraint that limits how far an edited prediction can drift from the original model's prediction, measured by KL divergence.