Preprint. The proposed Glassbox Framework places Bayesian networks between domain knowledge and generative-model outputs, rather than explaining decisions only after they have been produced. These probabilistic mediation layers are intended to make assumptions, uncertainty and reasoning traces open to inspection and challenge. A benefit-eligibility example illustrates the architecture. The paper identifies unresolved problems in semantic alignment, dynamic construction, probabilistic grounding and human governance that must be addressed before such systems can operate at scale.