Confidence belongs to each field
Verify critical fields individually and prefer independent captures over a single page-wide score.
Build evidence without a confidence score
Evidence unit
One important field and the exact image crop that produced it.
Stronger comparison
Source quality and agreement between readings provide evidence, not a magic score.
Escalation point
Escalate any disputed name, amount, date, dose, or instruction.
Confidence belongs to each field
An entire page can be easy to read while one faded digit remains uncertain. Treat names, dates, amounts, units, addresses, and negative words as separate fields. Mark each field as confirmed, plausible, or unreadable by comparing it with the pixels.
Do not let fluent surrounding text raise your confidence in a critical field. OCR models use patterns, so an output may look like a normal word even when the source contains a different name or code.
Use independent views
Make a second crop from the untouched original and, when possible, take a second photo with different lighting. Agreement between two genuinely different captures is stronger evidence than submitting the same compressed image twice.
View the source at high zoom and normal zoom. High zoom exposes character strokes, while normal zoom preserves line and label context. Keep both because either view can resolve a different kind of ambiguity.
Numeric fields need an additional audit because OCR can misread numbers even when the surrounding sentence looks correct.
Record uncertainty in the result
Use square brackets or a question mark for unreadable spans instead of silently guessing. Preserve the original spelling of names beside a transliteration. For structured data, attach the source crop to the extracted value.
If the field controls money, medicine, safety, identity, access, or legal rights, obtain a clearer source or human verification. Confidence is a decision about evidence, not a property created by a polished translation.
Use repeatable visual checks, independent captures, and field-level comparison to assess OCR text when no reliable confidence score is available. Continue with the image translation guide collection for adjacent workflows.