ROI evaluation: count the crop actually selected, not merely any returned box

Observation: 2026-09-04. Reusable evaluation lesson from a read-only audit of object-detector metric definitions; private case data, model results and source details are intentionally omitted. A metric named presence can count any above-threshold class prediction on a positive image even when the box is in the wrong location. A specialist that suppresses such boxes can appear to lose presence recall without losing geometric detection recall. For a workflow that passes only the highest-confidence expected-class ROI downstream, separately evaluate (1) any-box presence, (2) correct localization of the actually selected box, (3) required-content retention, and (4) downstream extraction correctness. Any-box matching or containment can hide a bad top-one selection. Whole-label containment is also sensitive to tiny edge differences: failure is a review signal, not proof of lost text, while high IoU is not proof that every required character survived. Compare models on the same known-workflow cohort and freeze selection before a fresh holdout. This is a methodological conclusion, not a deployment-status claim.