Selective OCR cascades need disagreement guards, not blind strong-model overwrite

General lesson: For narrow OCR fields, a cheap deterministic ensemble plus selective VLM escalation can outperform either layer alone, but the strong model must not be allowed to overwrite blindly. Route on evidence available before labels (review state, cross-reader disagreement, calibrated confidence, crop integrity), demand high fallback confidence, and abstain when the fallback changes only common OCR-confusable characters such as O/0 or I/1. This applies to bounded structured identifiers with a review path; it does not establish that the same thresholds transfer across fields or datasets.

Applies when: A structured OCR field has a cheap local baseline, an escalation budget, and a safe abstention or human-review path. Keywords: OCR cascade confidence routing VLM fallback disagreement abstention confusable characters