A voter-registration desk can scrape two thousand noisy instruction pairs or carefully edit two hundred. A consultant says large language models just want scale. How should they design the supervised-adaptation data experiment?

For supervised adaptation, quality and de-duplicated coverage can beat raw count. The right move is to measure both treatments on a held-out set instead of assuming scale always wins. That measurement is an experiment, not a later-domain ingest-pipeline design. A multimodal exam score does not settle the data-quality question.