Deep learning (DL) has excelled in tissue image classification, presenting opportunities to discover biological behaviors escaping visual inspection. However, the methods to produce fine-grained insights from spatially scattered information do not exist. Here, we introduce Delta-Marches, a framework for mechanistic interpretability that leverages generative models to produce high-fidelity images from semantic latent representations. By identifying directions in this space corresponding to class transitions, we simulate controlled morphological changes between classes. Comparing each image to its class-shifted counterpart enables a secondary model to nominate features most affected by the shift. This approach overcomes sample-to-sample variability and yields idealized, interpretable transformations at subcellular resolution. We prototype the approach in the context of histopathological grading of clear cell renal cell carcinoma. Delta-Marches generate synthetic grade transitions indistinguishable from real images and autonomously pinpoints nuclear enlargement and increased nucleolar count in tumor cells as key properties of higher grades - features identifiable only through the method's subcellular precision. In addition to these features mirroring clinical criteria, it also reveals reduced vasculature, a pattern reported in multiple studies but absent from standard grading rubrics. These results indicate Delta-March's potential to convert complex and spatially-distributed features into rules for image classification.