Is your PDF extractor silently dropping pages?

Most extractors return empty text for a page and still report success. Certify Anything audits any engine’s output against the source PDF — page by page — and tells you exactly which pages came back empty. Free, MIT, runs on your machine.

pdfmux verifynew in 1.8.1

Drop a PDF. See what your extractor missed.

Upload a PDF and pdfmux re-derives the source text with its own audit pass, then scores every page: usable, silently‑empty, or recovered. Nothing is stored — the file is audited and discarded.

Drop a PDF here, or click to choose

Up to 100 pages · audited with pdfmux verify · never stored

This is what a silent failure looks like

A 12-page government RFP run through a typical extractor. It returned five pages of text, reported success, and shipped — while pages 6–12 came back completely empty. Here is the exact pdfmux verify report that caught it.

7 of 12 pages silently dropped

FAIL · overall confidence 39% · coverage 39% · engine: your-extractor

Page	Status	Coverage	Confidence
1	usable	100%	100%
2	usable	100%	100%
3	usable	100%	100%
4	usable	100%	100%
5	usable	100%	100%
6	silently‑empty	0%	0%
7	silently‑empty	0%	0%
8	silently‑empty	0%	0%
9	silently‑empty	0%	0%
10	silently‑empty	0%	0%
11	silently‑empty	0%	0%
12	silently‑empty	0%	0%

signature sha256:abf0738b3eb333bd4a73b7243734e0a5e2ce70c22f7f811f1873b553f9be073d — tamper-evident: any edit to the report changes it

Sample report — a real pdfmux verify run on the first 12 pages of a public government RFP, with an extraction that returned pages 1–5 and nothing for the rest. The verifier will not cry “silent drop” when a page’s content is merely re-paginated: it re-derives the source and checks that the content is genuinely absent before flagging it.

Your parser is the ground truth’s judge, not the source of it

1 · Re-derivepdfmux reads the source PDF with its own audit pass — the column-reorder + heading + text extraction that powers its multi-pass pipeline. That reading is the ground truth.
2 · AlignYour engine’s output is aligned to the source page by page. Content that is all there, just merged or re-paginated, is marked recovered, not lost.
3 · ScoreEvery page gets coverage, confidence, alignment, hallucination-risk, and table/heading integrity. Pages the source has but the engine returned empty are silent drops.
4 · CertifyThe result is a signed, timestamped manifest with a tamper-evident content signature and an honest list of what the check can and cannot judge.

We shipped silent garbage in our own pipeline

We ran pdfmux across 433 of our own customer PDFs. The first pipeline lost 16 of them — and 11 of those losses had no error at all: empty output, “all done,” a CSV that quietly went out missing 11 products. That is the exact failure this page detects. After routing every page through the audit gate, the second run processed all 433 with zero silent losses.

433

PDFs in the batch

lost on the first run

of those with no error

silent losses after the fix

Source: the ARK chemical-PDF batch retro (April 2026), the single real-world benchmark we stand behind. A page that reports success and returns nothing is worse than a page that errors — the error you catch; the silence you ship.

Audit a whole directory from your terminal

The web tool audits one PDF. The CLI audits a batch, gates CI on a clean verdict, and never leaves your machine.

pip install pdfmux && pdfmux verify yourfile.pdf copy

# audit one engine's output against the source
pdfmux verify --source report.pdf --extracted reducto_output.json

# audit a whole batch, fail CI unless every page is accounted for
pdfmux verify --source ./pdfs/ --extracted ./out/ --strict
# -> exit 3 if any document silently dropped a page

Six signals per page, one verdict per document

Silent dropsSource page has real text; the engine returned nothing while reporting success. The headline failure.
CoverageHow much of the source page’s content actually made it into the extraction.
AlignmentFraction of the source’s content tokens present in the extraction — catches partial drops a length check would miss.
Hallucination riskExtraction tokens with no support anywhere in the source page.
Table & heading integrityDid tables and headings the source had survive the extraction?
Recovery guardContent that moved to another page is marked recovered, not failed — so a re-paginated extraction is not falsely blamed.