Use one class-balancing mechanism per training run

Choose either loss weighting with uniform example order or weighted sampling with an unweighted loss. Combining both applies the minority-class correction twice, makes the effective objective hard to reason about, and can make a receipt look balanced while optimization is overcorrected. Persist the chosen strategy, per-class counts, normalized weights, and example-order contract together, and reject artifacts that declare both. Effective-number weighting also requires every admitted class to have a positive count and a stable class-index mapping.

Applies when: training a classifier on imbalanced labels with deterministic, auditable optimization. Keywords: class imbalance loss weighting weighted sampler effective number objective reproducibility