A correctly read number is still wrong when extraction attaches it to the wrong item or table row.

Convincing-looking totals can hide shifted quantities, repeated headers, or relationships lost across pages and merged cells.

Keep difficult tables in the test set. Check item identity, units, row boundaries, and the connections between fields.

Ask a reviewer to reconstruct difficult rows using only the extracted output. Arithmetic can flag trouble, but it cannot prove the relationships are correct.

Measure completely usable rows and documents alongside field accuracy. The business needs a reliable record, not a collection of isolated numbers that happen to look right.

Original article explainer. AI-assisted conceptual illustration and neural narration; original synthesized music. Not a human recording or customer result.