Separate OCR transcription quality from schema compatibility and acceptance calibration

General lesson: compare OCR/VLM systems in three layers—raw transcription on fields both systems actually produce, declared label-blind normalization into the product schema, and accepted-result precision/coverage. A model can look much worse because it stores different date/name/state formats or omits fields, while a high extraction score can still be unsafe when its accept rule is poorly calibrated. This applies to replacing a production OCR/VLM pipeline with a local model; it does not establish that any one model is universally better, and post-hoc field exclusions remain descriptive only.