Abstract: Previous work reporting Beam Enumeration showed that \textit{probable} substructures extracted from a generative model contains chemically meaningful information and can act as a source of explainability. In this work, we propose Dynamic Beam Enumeration as an extension to extract larger substructures. We show that this extracted insight can be made actionable and used to filter compounds in ultra-large make-on-demand libraries ($10^{9-12}$). The resulting molecules possess properties more aligned with the target objective than random sampling. Importantly, the results suggest that Dynamic Beam Enumeration can act as a bridge between generative design and library screening, such that even if generated molecules cannot be easily synthesized, extracted knowledge from the model can be used to find promising molecules that are make-on-demand.