Crop-first OCR evaluation needs two separate measurements

Crop-first OCR evaluation needs two separate measurements: paired reader quality where detection succeeds, and end-to-end coverage including detector misses.

Freeze the original-image detection once, retain the exact crop and its hash, and apply image variants after cropping when testing this architecture. Feed identical cropped views to every reader. If crop-only input is required, make no-detection an explicit review outcome; silently supplying the full frame breaks the experiment. Preserve a full-scene comparison for diagnosis, since a detector can confidently select branding and thereby remove the evidence a VLM previously used correctly. Compare against visually adjudicated labels where possible: reproducing a bad reference can make a regression look like a gain.

Applies when: Testing detection-before-augmentation for OCR ensembles with crop-only VLM inputs. Keywords: crop first OCR detector coverage VLM augmentation reference bias