This page publishes our engine's measured results — and, beside each number, what that number does not prove. A gate being met is not the same as a result being reproducible outside this repository, and the two are stated separately.
100%
All 1,071 documents the corpus says should pass were accepted by the engine. The run started at 85.34%.
Gate: 99.5% — met.
This measurement stands on its own: the expected outcome is written into the corpus itself, so there is no second tool to install or to drift.
docs/quality/pdf-oracle-verapdf-2026-07.md
99.75%
The rate after policy classification. Only the "bug on our side" class counts against the gate.
Gate: 99.5% — met.
Measured and recorded — not yet publishable as a result. The reasons are listed below one by one; quoting this number as a finished claim would overstate it.
docs/quality/pdf-differential-oracle-2026-08-09.md
400
Real-world documents sampled from GovDocs1.
Measured and recorded — not yet publishable as a result. The reasons are listed below one by one; quoting this number as a finished claim would overstate it.
docs/quality/pdf-differential-oracle-2026-08-09.md
2,907
Total documents in the corpus; each filename states the outcome expected of it.
This measurement stands on its own: the expected outcome is written into the corpus itself, so there is no second tool to install or to drift.
docs/quality/pdf-corpus-coverage.json
The differential oracle ran against these comparators: qpdf, mutool. This comparator was absent from the environment and did not run: pdfium. The difference matters — "differential oracle" reads as more coverage than two tools, and the missing one is exactly what would buy that difference.
Every number here is stored alongside the verbatim line from the run document that produced it. A CI gate reads that document and fails the build if the line is gone, or if the number no longer appears in it. This page cannot outlive its own receipt.
scripts/ci/pdf-evidence-receipt-check.mjs