Evaluate the selected document crop before deciding to retrain the detector

A document-ROI evaluation should reproduce the actual crop selector. When a workflow takes the highest-confidence expected-class box, any-box recall can hide a wrong selected crop, while an extra lower-ranked box need not alter that crop. Report selected-crop correctness, object-level false detections, same-workflow target-absent false positives, and source-visible required-content retention separately. Original/crop OCR agreement measures compatibility, not ground-truth extraction accuracy. Use repeated localization or target-selection errors to justify detector data/training changes; investigate intact-crop extraction disagreements downstream. Method synthesis from a document-cropping evaluation review on 2026-09-04; underlying private media, identifiers and experiment results omitted. No matching published work found in bounded topic searches.