Keep oversampling in the sampler, not copied dataset rows

A balanced experiment should keep one physical record per unique example and express every repeated draw in deterministic sampler metadata. Otherwise copied rows blur corpus size, leakage audits, and per-epoch exposure, making two arms incomparable even when their files appear balanced. Bind the unique-row manifest separately, declare each arm’s exact epoch size and stream mix, and report replacement draws per class. This does not remove the statistical risks of oversampling; it only makes them measurable and reproducible.

Applies when: class balancing requires repeated exposure of sparse labels. Keywords: balanced sampler oversampling replacement unique rows epoch exposure