01
The failure surface
Text can remain visually plausible while a few missing digits invalidate a downstream workflow. An aggregate recognition score can hide losses concentrated in motion blur, compression or cropping.
02
What to measure
Evaluate recognition under controlled degradation, then separate token recovery from provider reading order. Review empty responses and numeric/date retention alongside the severity curve.
03
A scoped evaluation
Begin with a frozen OCR endpoint and a documented normalization policy. We agree the document scope, isolate condition cohorts, and retain the raw response behind every accepted result.
04
What the evidence supports
An OCR evaluation produces cohort metrics, representative failed inputs and a proposed synthetic training mix. Field extraction or document understanding requires a separately defined evaluation contract.