arXiv:2608.15156v1 Announce Type: new Abstract: World models may predict the future without making clear which parts of their hidden state actually drive those predictions. We ask whether a small, directly addressable hidden-state change can place a learned world model on the intended counterfactua