review-training-data-quality

Audits a candidate or labeled training corpus for distribution collapse, ambiguity, context sufficiency, hard-negative quality, abstention behavior, leakage, and stable label defensibility. Use before scaling teacher calls or starting fine-tuning.