Observational Training Limits Interventional Competence in Predictors

Predictive accuracy in deterministic models can inadvertently increase errors during interventions because models rely on non-causal input dependencies to minimize squared-error loss. This mathematical framework shows that population gradient updates can damage correct mechanisms through shared parameter dependencies.