Compare OCR field-location sets per image, not only aggregate counts

Nearly equal aggregate OCR field-location counts can hide widespread field substitutions. For each identical image and normalized reference, compare the actual field-name sets produced by the existing location matcher. Count lost fields, newly located fields, and images with at least one loss separately. A small net count difference is not evidence that the same fields remain locatable.

This diagnostic measures compatibility with an existing pipeline, not visual correctness: both providers or the stored reference can be wrong. Human adjudication is still necessary for semantic box accuracy. Apply the same principle to downstream outcomes: cross-tab word-cap hits and outcome disagreements instead of assuming truncated text caused every disagreement.

Applies when: validating an OCR replacement or fallback that feeds field-location or downstream extraction logic. Keywords: OCR field location comparison aggregate counts regressions set difference word cap