Presents a model-agnostic explainability pipeline that uses diagnostic classifiers to probe graph neural network embeddings for structural graph properties, clarifying what information their hidden states retain across architectures and datasets.