Train the Model, Not the Reader: Decodability Supervision for Verifiable Activation Explanations | ArxivCSExplorer