Separate OCR transcription failures from identifier-selection and acceptance-policy failures.

When several OCR readers agree on the wrong region, adding votes can reinforce the error. Compare original and transformed views within each model family, then merge families; do not treat eight correlated image transforms as eight independent experts. Review disagreement cases alongside a small random agreement control sample, preserve production labels, and keep visual corrections in a separate sidecar. A policy requiring one favored model to support every acceptance can increase reviews without fixing errors; freeze it after development batches and test on unseen subsequent batches.

Applies when: Expanding an OCR ensemble benchmark incrementally using production-derived reference labels. Keywords: OCR ensemble transforms correlated votes production labels outliers permits