Abstract
We evaluated whether adding prediction consistency (PC) to hard augmentation yields repeatable robustness benefits in industrial surface-defect classification. Five training strategies were compared using ResNet18 on NEU-CLS, Magnetic Tile, and X-SDD across five repeated stratified splits and three seeds (225 runs). PC changed mean corrupted-image Macro-F1 by −0.0026, +0.0168, and +0.0051, respectively. All corrected dataset-level intervals included zero, and no condition-level effect survived Holm correction. Component ablations and a further 30-run NEU control with training-only coefficient calibration did not establish a PC gain or selection–PC interaction. Archived development-set endpoint diagnostics characterized prediction agreement and gradient magnitude but did not establish a causal mechanism. On a sealed 270-image NEU set, 75 frozen models reproduced the broad method-by-condition pattern (r = 0.9983), while small PC effects were inconsistent. Additional comparators placed IPMix highest in mean performance under the original evaluator, while evaluator controls changed the speckle margin between HardAug + PC and AugMix. Between-split variance remained imprecisely estimated. These findings distinguish repeatability of small incremental effects from broad performance agreement and demonstrate sensitivity to evaluation choices. Conclusions concern the tested complete training strategies under one backbone, three public datasets, and synthetic corruptions.