How can you verify that AI extracted all data from a PDF?
By not taking the output on trust, but reconciling it against an independent reading of the original: every extracted value needs a location in the PDF, every number in the PDF must appear in the output, and row counts and totals must add up. Anything without a match is flagged as a gap and listed in a verification report.
1. Build an independent baseline
The PDF is read a second time – without AI: the text layer with conventional tools, scans via OCR with confidence scores. Pages, tables, rows, headings and footnotes are counted. This is the yardstick the AI output is measured against.
2. Provenance for every element
Every value in the JSON and every section in the Markdown points to its page and position in the original. That makes coverage measurable per page: which parts of the document have a counterpart in the output, and which don't?
3. Control totals and number reconciliation
Every number in the document must reappear in the output. Tables must keep their row count, totals must add up. This check is simple, deterministic – and in practice catches most gaps.
4. Literal grounding
Extracted values must be found verbatim in the source text. This exposes values the AI invented or “improved”.
5. A second, independent pass
A second model reads the same document. Where the two results differ, a human decides – not chance.
The result: a verification report
The output is not just Markdown and JSON, but a report with coverage per page, flagged items and an audit trail. Nobody can guarantee zero errors – but every gap becomes visible instead of silent.