Is your PDF pipeline lying to you?

Drop a PDF. See which pages your extractor silently dropped or got wrong — page by page. Free, no signup, nothing stored.

Audit your current extractor instead

Add the JSON / Markdown / text your existing tool produced. We'll check that against the source and show what it dropped.

Nothing you upload is stored — the PDF is deleted the moment the audit finishes.Measured accuracy: 6.7% false alarms and 0 missed page-drops on a pre-registered 24-document corpus; it misses most truncated tables (4 of 5 seeded cases).
What this catches — and what it misses
  • Catches reliably: dropped pages, severe content loss, and text with no support in the source — 0 misses across the validation corpus.
  • Misses often: truncated tables. It missed 4 of 5 seeded table truncations on the validation corpus — a 5-case sample, so read that as “most”, not as a precise rate. A table that is present but cut short can still pass. Do not rely on this tool for table completeness.
  • Declines to judge: pages with no digital text layer (scans) are marked unverifiable, never silently passed — 25% of the corpus. Some valid RTL/exotic-font PDFs also land here.
  • False alarms: ~1 in 15 correct extractions drew a failing verdict (6.7%), usually where a page's own text layer is degenerate.
Measured on a pre-registered, publicly published corpus — methodology and per-document results in VERIFIER-VALIDATION.md.