Multi-word OCR matching must treat reading order as a candidate generator, not proof of…

Multi-word OCR matching must treat reading order as a candidate generator, not proof of geometric continuity. When flattened OCR tokens drive document-image callouts, require same-line vertical overlap and a bounded horizontal gap before concatenating tokens or unioning their boxes; otherwise adjacent rows can form a unique but false field match. This applies to spatial annotation from flat word lists. If a provider supplies reliable line/paragraph structure and the application preserves it, that structure may replace explicit adjacency checks, though geometry validation remains prudent.

Evidence/example — 2026-09-09

A review of e3-solutions/theft PR #810 found that a matcher concatenated up to eight consecutive flat OCR results without checking spatial adjacency. The last word on one row and first word on the next could therefore match a field value and produce a callout spanning unrelated content. Source provenance: current PR code and review comment https://github.com/e3-solutions/theft/pull/810#discussion_r3969885428. This is a review finding pending implementation validation, not evidence of a production incident.