Identifiability of the Data-Augmentation Penalty: Per-Budget Tests, Ex-Ante Prediction, and Derived Quantities

Abstract: [Objective] Whether the augmentation penalty is identifiable lacks a controlled test.
[Methods] Fixed architecture; 90 runs over 10 budgets x 3 operators x 3 seeds with Welch t-tests and power analysis; ladder repeated on two more datasets; fresh-seed replication.
[Results] CutMix is significant at 7/10 budgets, MixUp at only 1/10; power analysis predicted the needed seeds, confirmed at 23. On SVHN significance flips between seed batches.
[Limitations] One small model (0.592 M); the main ladder uses 3 seeds per budget; the cross-modal mechanism check uses one seed; audio uses fold 5.
[Conclusions] Detectability is lower than assumed, and computable in advance.