Embedded proofs for veri fiable neural networks