Granular Knowledge Provenance Bi-Directional Document-to-Node Mapping for High-Resolution Neural Network Interpretability

Current approaches to neural network interpretability often operate at high levels of abstraction, attributing model behavior to broad concepts or entire datasets. However, critical applications in healthcare, law, and copyright compliance require document-level granularity. We introduce a system for ”Granular Knowledge Provenance” that establishes a bi-directional link between individual training documents and specific network parameters. By tracking gradient contributions at the sample level during training, our system allows users to: (1) query a specific node to reveal the top-contributing source documents, and (2) filter the entire network to visualize the ”architectural footprint” of a single document. We address the significant computational challenges of this high-resolution tracking and demonstrate, through a medical case study, how this capability enables ”source-code level debugging” for data, allowing for precise verification of facts and identification of training data leakage.