Searching for archival voids is a central concept in critical archival research, yet traditional approaches remain largely restricted by what is present and legible: Natural Language Processing and text-based methods tend to operate at the level of the document, struggling to engage with what is structurally absent. Operationalizing the archival principle of respect des fonds, instead, demands methods that preserve the structure and hierarchy of a collection as a whole. In archival collections of the Holocaust, silences take many forms: evidence destroyed by perpetrators, undocumented stories, records lost to dispersal or rendered invisible by digital archiving practices. Though computational approaches cannot tend to all voids, we aim to support the interpretive work of archival justice research by exploring ways for reading absence itself as a form of evidence and recovering fragmented information that might escape conventional approaches with the help of provenance and relational context. To this end, we introduce a Graph Neural Network pipeline for the critical engagement with collections from the European Holocaust Research Infrastructure. Our methodology treats documents not as isolated nodes but as entities embedded in relational structures of institutional provenance, hierarchy, and semantic content. We present a series of investigations demonstrating how our GNN approach can detect anomalies, discontinuities, and structural gaps that signal archival silences, offering a transferable framework for a critical interrogation of archival voids and the reconstruction of dispersed records in historical collections.