Decomposing OCR decisions requires field-specific grounding, not a universal quality gate

When decomposing a multimodal document evaluator into OCR, bounded semantic selection, and image checks, preserve the relationship between each required field and its visible source. A generic image-quality pass/fail gate can reject usable evidence because a harmless defect exists elsewhere, while flat OCR can combine values from unrelated records. Return visual observations and source-linked field or whole-record candidates; let task-specific policy compose the result. Treat missing candidate coverage and placeholder confidence as unknown, not assurance. A candidate selector's certainty does not measure transcription correctness.

Applicability and limits: useful for optimizing document/capture workflows that currently combine reading and acceptance in one model call. Parallelism only helps when the visual check is actually independent of the chosen text candidates. Otherwise latency stays serial, and fallback adds cost. Accuracy and latency gains remain hypotheses until measured end to end.

Evidence/example — 2026-09-17: constructed example, not customer data: a form contains an issue date and a registration-expiration date, with glare confined to a decorative header. OCR retains both dates and their boxes; a Choice model picks the expiration span or unknown; code normalizes the selected date; the visual check reports whether the required field is obstructed. The glare alone should not decide rejection. For multi-row insurance, select one coherent record instead of separately choosing its name, number, and dates. TypeSafe's public Choice API provides predefined options and uncertainty, not visual reading: https://docs.typesafe.ai/primitives/choice. Google's document OCR exposes layout hierarchy that should be retained: https://docs.cloud.google.com/vision/docs/fulltext-annotations.