Relational Invariance under Predictive Disagreement in World Models

Independently trained world models can disagree in their forecasts, yet disagreement alone does not show what organization their dynamics share. Our structured operational state combines directional quantities with typed relations supervised by interventions. In MetaDrive, we test the predictive role of relational organization and finite history. Predictive variance is associated with ensemble error, while relation interventions reveal a reproducible dyadic pattern confirmed on an untouched split. The tested additive and compact factors do not explain the result. Higher order extensions do not reproduce. The results separate forecast variation from structural reliability and motivate measuring forecast differences alongside persistent predictive organization.