OCR/VLM hybrids should be tuned per workflow, not globally

General lesson: A hybrid OCR/VLM architecture should share reliability plumbing but isolate extraction policy by workflow. Flat OCR evidence can improve exact transcription in one document class while distracting the VLM or breaking field association in another. Give each workflow its own image views, OCR evidence format, prompt, validators, abstention rule, and selection metric; tune on a small labelled development slice, freeze one candidate, then evaluate once on a disjoint holdout. Applicability: heterogeneous image-to-structured-data systems with plates, IDs, vehicle markings, or dense forms. Limits: a small development slice selects a candidate but cannot establish generalization, and shared source families can still leak across a nominal row split.