Problem
Dense bilingual bulletins must become structured matter fields without retyping — and without silent errors.
What happened
Text extraction alone still left structure wrong. Auto-merge fantasies create false confidence: the system looks done when the field is wrong.
Why it happened
Demos reward “upload → text.” Legal desks reward provenance and confirmation.
What changed
Staged pipeline with confidence and a mandatory human review gate before register merge.
What I’d do differently
Design the review UI before choosing an OCR vendor. Never ship a path that writes the system of record silently.