ExactK: Iso-K Hinge Loss for Robust K-Leakage Reduction in QEC Neural Decoders

ExactK is a production-ready ML decoder for quantum error correction on surface codes. It uses iso-K hinge loss (exact syndrome-count matching) to eliminate K-leakage confounding in factor-graph neural network decoders. Validated on d=5 (+31.2% G1 reduction) and d=7 (+23.4% OOD generalization), with a V1.0 holdout achieving +45.0% science delta, 80% safe yield, and 0% topology failure on unseen seeds.