When Validation-Gated Selection of Diffusion Augmentation Fails to Generalize: A Retrospective Audit of Rare Seagate KQI Classification

Validation-based rules often decide when synthetic minority augmentation should replace conventional imbalance treatment but may not transfer to new targets. We audited this retrospectively on public Seagate time-series-2 data, whose 11 key quality indicator (KQI) endpoints are pass/fail outcomes of later wafer measurements (0.96–3.87% positives). Task-identifier-free validation gates, each selected on test outcomes of other endpoints, decided between a conventional reference and diffusion-augmented hybrids for the held-out endpoint. In the original archive (nine endpoints, LightGBM), the conservative objective selected no hybrid; the mean-seeking objective changed task-macro test precision–recall area under the curve by −0.0021. Because the archived generator had learned little, we rebuilt all candidates with a corrected generator and deterministic evaluation under a pre-specified protocol. On the nine original endpoints, the LightGBM effects were −0.0014 and −0.000029 (no descriptive interval above zero); with XGBoost on all 11 endpoints, the mean-seeking effect was +0.0017 (−0.0002 to +0.0042), and omitting one endpoint reversed the sign of three of four mean-seeking summaries. A historical gate family refined after test inspection (retrospective effect +0.0086) selected no hybrid in the rebuilt LightGBM archive. The evaluated gates gave no portable rule for using diffusion augmentation on these endpoints; this does not show that the augmentation itself is ineffective.