Leave-one-out stability and image audits can overturn an apparently perfect OCR router

General lesson: A selective OCR router that removes every false accept on its tuning cohort can still depend on one influential error. Before treating its threshold as calibrated, select the threshold on N-1 rows and evaluate the excluded row, then visually audit the disagreements against source images, crops, and labels. This applies to small structured-OCR cohorts; leave-one-out remains development stability evidence, not production accuracy.

Applies when: A small OCR benchmark is used to choose abstention thresholds or reconcile deterministic readers with a strong VLM. Keywords: OCR threshold leave-one-out calibration image audit crop integrity VLM fallback