Separate image transcription errors from invalid source data and field assignment

Image-extraction evaluation needs separate checks for transcription fidelity, source-data validity, and semantic field assignment. An output can faithfully copy an invalid source value, or read a date correctly while assigning it to the wrong field. This method applies to document-to-structured-data systems; it does not establish any model's accuracy or replace domain validation.

Inspect the exact original image before attributing a discrepancy to OCR. Record both the literal visible value and the permissible normalized field value. Score invented text, invalid-but-faithfully-copied input, and wrong-field assignment separately. Preserve uncertainty when the image conflicts with newer explicit text; do not silently repair a source value from memory. Keep transport/API success separate from semantic success.

Evidence/example — 2026-09-11

Constructed example, not customer data or a reported benchmark: a screenshot labels a date as “Delivery” and prints a six-digit value in a US ZIP field. Copying that date into pickup is a semantic error. Copying all six digits may be faithful transcription but still fails field validation. Calling both cases OCR errors would point remediation at the wrong layer.

The linked prior work supports separating quality claims: OCR readability does not establish photo authenticity. This extends the same principle to transcription, source validity, and field meaning, without correcting or asserting current status of that earlier system.