Where the third-party corpus documents come from

On 2026-09-24 the public internet was searched for the documents docs/corpus-contributions.md asks for (W01 to W12), and the test corpora of other PDF libraries were examined for the same purpose. On 2026-09-25 a fourth pass took the Open Preservation Foundation's format-corpus, file by file. This page records the rules applied, what entered tests/corpus/vendor and what entered the remote corpus — used without being redistributed —, what was turned down and why, and what the search taught us. tests/corpus/manifest.json holds the per-file detail: source URL, retrieval date, SHA-256 of the published bytes, and every expectation.

The rules a document had to meet

A file entered the corpus only if it met all four. A file that failed one was rejected, not repaired.

1. An attribution-only license that covers this file (ADR 23): public domain — including United States federal works under 17 U.S.C. 105 —, CC0, CC BY, the Licence Ouverte, the UK Open Government License, the EU's reuse decision 2011/833/EU, Japan's Public Data License, MIT, BSD or Apache-2.0. ShareAlike, copyleft, non-commercial, no-derivatives and "no license stated" were all refused. A code license on a repository does not cover the bug-report attachments committed into it. A file that fails this rule may still be used without being redistributed, in the remote corpus (ADR 32); In the remote corpus below says which terms that admits.

2. No real person's contact or identity data anywhere: no one's phone number, postal address or own e-mail address, and no bank, identity or health data tied to a person — in the visible text, /Info, XMP, annotations, form values, XFA packets, signing certificates, embedded files, and the objects of superseded revisions that incremental updates keep in the file. A person's name alone is acceptable, and so are a Windows user ID in a path and a photograph. Fabricated, fake and specimen data of any kind is fine — John Doe, Contoso, Entenhausen, a 555 number. Organizations' addresses and switchboards are fine.

   The rule was decided with the maintainer, and relaxed during the search: it began as "no real person's name anywhere", with typeface-designer credits, public figures and historical figures as the only exceptions. The first 38 files were admitted under that stricter rule, and every file turned down for a name alone was then examined again.

   The third pass brought four more rulings. An image of a named person's handwritten signature counts as the name: acceptable, even in a committed file. A software library author's own e-mail address inside that library's copyright string, compiled into the file, is a library credit and acceptable, as a typeface designer's credit inside a font is. An anonymous photograph whose caption describes a health condition discloses no one's health. A fabricated e-mail address used as a test constant is fictitious data, even at a real domain.

3. The publisher's own bytes, never a re-saved copy: each file was downloaded again and its SHA-256 compared, from a permanent URL where one exists (a repository commit, the EU's Cellar URIs, the IRS prior-year archive), or from the Internet Archive's unmodified id_ copy when the publisher has withdrawn it.

4. Small: at most 2 MB, usually a few hundred kilobytes. W11, where size is the point, is recorded below by reference and not committed.

   The maintainer has since turned this rule into a threshold. Size no longer turns a document down: a file over 2 MB that meets the other rules goes to the remote corpus (ADR 32) instead of being committed. Five files had been refused on size alone, and were then reconsidered on that basis. Three went to the remote corpus: the Open Preservation Foundation's signed 3D portfolio (CC0, 4.7 MB), and two members of the PDF/UA Reference Suite, a tagged scan (10.6 MB) and a textbook chapter (2.3 MB), both a publisher's content, which the suite's CC BY cannot be shown to cover. Two were refused again, for personal data this time: a NIST request for quotation scanned on a Canon SC1011 prints two named staff members' direct phone lines and own e-mail addresses — and Acrobat had re-saved it anyway, so it was not the untouched copier file W03 wants —, and the suite's Danish magazine gives the health conditions of named people, children among them.

Embedded fonts stay inside the documents under their own embedding permission: subsets always, complete fonts when their fsType allows embedding.

How the search ran

• Survey: some ninety PDF libraries and public corpora, in every language, were examined for test PDFs, the license that actually covers them, and how provenance is tracked. The findings are below.
• Hunt: 69 candidates were downloaded and inspected — producer chain, cross-reference shape, signatures and certificate subjects, form fields, fonts, image filters, extracted text, qpdf --check.
• Adversarial verification: 47 shortlisted files went to independent verifiers told to reject on doubt. They re-read the full text of every page and every revision, re-fetched each license statement, and re-downloaded each file. 12 were rejected; 4 of those were readmitted after the maintainer's rulings on font credits, complete fonts and historical figures; one more was dropped by a final scan of every revision's metadata. 38 entered the corpus, 6.6 MB in all.
• Reopening, after the rule on names was relaxed: every file turned down for a name alone was examined again, and a critic went through the 117 earlier rejections for the ones missed. Two rounds re-downloaded, re-screened and re-licensed 70 candidates against the revised rule. 38 more entered, 6.1 MB, filling most of what the first pass had left open.
• Third pass, the leads held back by their license: that evening, eleven research groups took up the leads this page listed as held back by their license, and looked for more of their kind. They examined 113 candidates and recorded 158 further leads they left, each with its reason. Verifiers, again told to reject on doubt, upheld 79 of the 80 files proposed, and the integration set aside two more, one of which the maintainer then admitted. With the files once refused for size and a PDF/UA brochure reconsidered, 82 entered: 6 committed and 76 remote. The committed corpus then held 112 documents, 16.8 MB.
• Fourth pass, the OPF format-corpus: all 284 PDFs at the pinned commit that the first passes had not taken, screened one by one, then every admitted file re-screened by one verifier told to reject on doubt. 124 entered: 56 committed and 68 remote — one of them, a fact sheet too damaged to be screened whole, admitted afterwards by the maintainer's decision, with the exception written into its entry —; 160 were refused, 89 of them the OPF's copy of a hand-built set whose bytes git had altered, and seven after the verifier's objections. The committed corpus then held 168 documents, 23.0 MB, and the remote corpus 154.
• The iPRES 2017 set, through ADR 33: the 88 test files of that hand-built set, fetched out of the BagIt tar its authors deposited on RADAR, once fetch_remote.py could copy a member out of a pinned archive. The committed corpus still holds 168 documents, 23.0 MB; the remote corpus now holds 242.
• Referees: page counts and attachments from pikepdf in the first two passes, from qpdf --show-npages in the last two; the qpdf --check verdict from qpdf 11.9.1 in the integration tests' container; text from poppler's pdftotext 24.02; conformance from veraPDF 1.30.2.
• Our reader opened every one of the 239 real PDFs downloaded along the way: 209 opened clean, 15 encrypted ones were refused with the typed exception M16 will replace, 15 opened with repairs or warnings, and none crashed or hung. At least one of those warnings is a false alarm of the reader's own — a stream whose endstream falls just past its 8 KB window is reported truncated (T21 in docs/status.md, fixed since) — and the others were not all checked against qpdf: a Cerfa reported truncated may be the same defect. In the third pass the reader was run over every file taken in, against expectations qpdf had set. None crashed or hung it. It fell short on five, recorded as unsupported rather than fixed on this branch: an object longer than its 8 KB window is cut at the edge (T23, two files, fixed since), each cross-reference section is read through a window of up to 64 KB whatever its size (T24, one file), and two files leave it silent where qpdf reports damage. In the fourth pass none of the 131 first admitted crashed or hung it either; of the 124 kept, it fell short on ten, recorded as unsupported — seven since T21 and T23 were fixed: T21 (one file), T23 (two), T24 (one: a 273 KB table read again each time its window grows), a new T25 (a /Prev 12 bytes off, dropped without a word), and five where the page tree or the catalog is wrong and the reader is silent or counts differently from qpdf, until M02. Over the 88 iPRES files it falls short on 17, all until M02: seven page trees it counts differently from qpdf, four faults it reads without a word, and six recoveries that differ from qpdf's — in four, the reader rebuilds a sound index to find a catalog the trailer no longer leads to. Two references to an object the file lacks, after which it rebuilt its whole index where qpdf takes null, read as null since T27 was fixed on 2026-09-27.

What entered the corpus

First pass

File under tests/corpus/vendor/	W	What it brings
us-federal/xerox-workcentre-treasury-imf-report-scan.pdf	W03	Copier firmware output: JBIG2 with global segments, no text layer
us-federal/hp-mfp-acrobat-ocr-nih-report.pdf	W03	An HP MFP scan with an invisible OCR layer added by Acrobat 11
us-federal/acrobat3-import-irs-1040-1988-scan.pdf	W03, W10	PDF 1.2, CCITT G4, imported by Acrobat 3.0 in 1999, NUL-terminated /Info strings
us-federal/distiller2-irs-9465-1996-rc4-40.pdf	W10	Distiller 2.0, 40-bit RC4 with an empty user password, an AcroForm from 1998
us-federal/distiller3-irs-ss4-1995-form.pdf	W10	Distiller 3.0, an Acrobat 3 AcroForm, LZW streams
us-federal/distiller4-opm-fehb-election-form-2000.pdf	W10	Distiller 4.0, TrueType turned into Type 1 subsets, Arial not embedded
us-federal/indesign-irs-pub1-arabic.pdf	W09, W02	Right-to-left Arabic in Type 1 (CFF) subsets, /Lang wrongly en-US
us-federal/indesign-irs-pub1-russian.pdf	W09	Cyrillic through /Differences encodings, six incremental updates, a hybrid index
us-federal/indesign-irs-pub1-chinese-traditional.pdf	W09	CIDFontType2 and CIDFontType0C subsets
us-federal/livecycle-uscis-ar11-xfa-form.pdf	W08	An XFA form with calculate scripts, Reader usage rights, AES-128
us-federal/itext-govinfo-us-code-certified.pdf	W05	A GPO certification signature (DocMDP) in an incremental update, written by iText 7 (Java)
opf-format-corpus/distiller3-dea-cfr-damaged.pdf	W06, W10	Real damage from GovDocs1: a wrong startxref and wrong stream lengths
opf-format-corpus/imagemagick-false-pdfa1b-jpx.pdf	W12	A false PDF/A-1b claim over a JPEG 2000 image
uk-ogl/word2019-ccs-contract-schedule.pdf	W01	Word 2019's hybrid cross-reference, non-embedded Arial
uk-ogl/indesign-acrobat-hmcts-n208-form.pdf	W02, W08	InDesign 15.1, Acrobat usage rights, optional content, a CCITT image
uk-ogl/distiller10-hmrc-gift-aid-letter.pdf	W02	Distiller 10.1.8 behind the PScript5 driver
uk-ogl/pagemaker-distiller5-hmrc-iht205-form.pdf	W06, W08	JavaScript form calculations, RC4-128; it broke pdf.js (issue 13132)
fr-licence-ouverte/word2010-meae-paris-call-survey.pdf	W01	Word 2010, /Lang fr-FR over English text
fr-licence-ouverte/word365-dila-text-drawn-as-images.pdf	W01	Word 365 drew the text as soft-masked images, having failed to embed the font
fr-licence-ouverte/print-to-pdf-excel-agec2025.pdf	W01	Microsoft Print to PDF output published untouched, /Author empty
fr-licence-ouverte/libreoffice-cerfa-13983-form.pdf	W08	A Cerfa: a LibreOffice form re-saved four times by other tools
eu-publications/pdflib-oj-exchange-rates-greek.pdf	W09, W12	Greek, PDF/A-2a from PDFlib, complete EUAlbertina fonts
eu-publications/antenna-house-oj-exchange-rates-2019.pdf	W12	PDF/A-1a from Antenna House: empty /Lang, references to objects the xref lacks, malformed font XMP
eu-publications/distiller10-eu-consolidated-regulation-2015.pdf	W12, W02	A PDF/A-1a claim with a CMYK output intent, which veraPDF rejects
eu-publications/3heights-eu-consolidated-regulation-2026.pdf	W12	PDF/A-2a from a converter rather than a producer
lu-legilux/antenna-house-legilux-memorial-pades-lta.pdf	W05	A qualified electronic seal: PAdES B-LTA, DSS, a document timestamp, three incremental updates
jp-nta/indesign-distiller18-nta-gift-tax-vertical.pdf	W06, W09	Vertical Japanese without ToUnicode, AES-128; it broke pdf.js (issue 11526)
pikepdf/scanner-ccitt-endofline.pdf	W06, W03	A scanner's CCITT G3 image with /EndOfLine true; it broke pikepdf (issue 601)
pikepdf/handwritten-cyclic-toc.pdf	W06	Cyclic destinations, no trailer /Size, a page without /Parent, /MediaBox or /Resources; it crashed pikepdf (issue 677). Reported since M02's third slice
pdf-association/handwritten-type3-recursion.pdf	W06	A Type 3 recursion cycle: a denial-of-service probe for invariant 4
pdf-association/handwritten-dict-is-stream.pdf	W06	A page object given a stream body: qpdf 11.9's walk counts it as a page, libqpdf 12 drops it. Reported since M02's third slice
pdf-association/handwritten-indexed-color-out-of-range.pdf	W06	Indexed color addressed out of range
pyhanko/acrobat-reader-signed-twice.pdf	W05	Two Acrobat Reader signatures with timestamps, test certificates
zugferd/mustang-zugferd2-en16931-invoice.pdf	W07, W12	ZUGFeRD 2.0 EN 16931, PDF/A-3u, an /AF attachment
zugferd/mustang-zugferd1-basic-invoice.pdf	W07, W12	The legacy ZUGFeRD 1.0 namespace and attachment name
zugferd/fop26-xrechnung-visualization.pdf	W04	Apache FOP 2.6 with base-14 fonts
docentric/fop-factur-x-visualization.pdf	W04, W12	Apache FOP, a PDF/A-3b claim without the XML attached
bfo/bfo-pdfa2b-embedded-pdf.pdf	W12	PDF/A-2b carrying another PDF, whose %%EOF sits inside a stream

Second pass, under the revised rule

File under tests/corpus/vendor/	W	What it brings
zugferd/itext-pdfbox-weclapp-facturx-en16931-invoice.pdf	W07, W04	A real ERP's Factur-X (weclapp, a test tenant): iText 2.1.7 then PDFBox, /MarkInfo aliasing the page tree, a one-cent gap between page and XML
docentric/aspose-d365-facturx-extended-invoice.pdf	W07	Dynamics 365 through Docentric AX and Aspose.PDF for .NET: Factur-X EXTENDED, PDF/A-3B
zugferd/gnuaccounting-mustang10-zugferd-rc-invoice.pdf	W07, W12	An accounting package's ZUGFeRD from 2014, in the release-candidate namespace
pdf-association/pdflib-pps-kraxi-pdfa2a-pdfua1-invoice.pdf	W04, W12	PDFlib's Java invoice, PDF/A-2a and PDF/UA-1, both upheld by veraPDF
jasper-modular/openpdf-jasperreports-financial-statement.pdf	W04	JasperReports 7.0.6 on OpenPDF, fabricated data
swissqrbill/pdfbox-swiss-qr-bill-a4.pdf	W04	A Swiss QR-bill from the SwissQRBill library through PDFBox 3, with SIX's specimen data
fr-licence-ouverte/fop-dictao-dila-signed-joafe-notice.pdf	W05	A Journal officiel notice signed by DILA's Dictao service: SHA-1/RSA, an organization certificate
lu-legilux/fop22-legilux-memorial-seal-renewed-timestamps.pdf	W05	A 2018 qualified seal kept alive by document timestamps added in 2018 and 2020
us-federal/xerox-workcentre-5335-ocr-hud-fonsi-linearized.pdf	W03	A Xerox copier's own OCR (render mode 3, horizontal scaling up to 2000 %) over a JBIG2 page
us-federal/xerox-workcentre-5755-ocr-hud-fonsi-mrc.pdf	W03	Mixed raster content with the copier's own OCR: a JPEG background, JBIG2 masks, a MediaBox with a negative origin
uk-ogl/print-to-pdf-word-cspl-agenda.pdf	W01	Print to PDF from Word, untouched, the printing account in /Author
us-federal/print-to-pdf-excel-dod-fcf-rates-2021.pdf	W01	Print to PDF from Excel on a government workstation, untouched
opf-format-corpus/quartz-word-mac2011-lorem-ipsum.pdf	W01	Word for Mac through the macOS Quartz writer: indirect stream lengths, a MacRoman subset
uk-ogl/pdfmaker10-word-dh-care-bill-factsheet.pdf	W02	PDFMaker 10 for Word, then an Acrobat incremental save that redefines the pages
uk-ogl/pdfmaker25-home-office-eia.pdf	W02	PDFMaker 25 for Word, linearized then updated twice
opf-format-corpus/pdfmaker9-word-distiller-pdfa1b-test-document.pdf	W02, W12	PDFMaker 9.1 and Distiller 9.5 with a PDF/A-1b claim veraPDF upholds
opf-format-corpus/pdfmaker9-word-distiller-linearized-font-not-embedded.pdf	W02	8 KB of Adobe linearization: a first-page startxref 0, a stale hint table
opf-format-corpus/pdfmaker5-distiller5-va-select-agents.pdf	W10, W06	PDF 1.2 from PDFMaker 5 (2002), linearized then updated in 2006; a GovDocs1 error file
us-federal/illustrator-irs-pub1-english.pdf	W02	Adobe Illustrator CS6: the English source of the vendored Arabic, Russian and Chinese editions
opf-format-corpus/reader10-openoffice32-annotated-object-streams.pdf	W02	An OpenOffice file fully re-saved by Adobe Reader X: Adobe's streams around a third party's content
opf-format-corpus/openoffice32-pdfa1a-embedded-lucida.pdf	W12	PDF/A-1a from OpenOffice.org 3.2, upheld by veraPDF
pdf-association/indesign13-pdfua1-german-book-chapter.pdf	W02, W12	InDesign CC 2018 claiming PDF/UA-1: a rich tag tree, three incremental updates
pdf-association/indesign15-pdfua1-form.pdf	W08, W12	A tagged InDesign form claiming PDF/UA-1, widgets without appearance streams
us-federal/livecycle-irs-1040-2022-xfa-ur3.pdf	W08	An XFA form the reader can open today: not encrypted, 136 fields, usage rights
pikepdf/livecycle-dod-dd293-aes128-xfa.pdf	W08	An XFA form under AES-128, linearized with a faulty hint table
us-federal/designer-distiller23-uscis-i9-javascript-form.pdf	W08	A 2025 AcroForm driven by JavaScript validation, format and keystroke actions
fr-licence-ouverte/pdfmaker-acrobat-cerfa-12156-form.pdf	W08, W02	PDFMaker 21 re-saved by Acrobat Pro DC: 406 fields, a calculation order, 494 stale hint entries
fr-licence-ouverte/livecycle-es9-cerfa-14880-xfa-form.pdf	W08	LiveCycle Designer ES 9 XFA, usage rights applied twice
us-federal/omniform-usda-rd1924-5-hidden-widgets.pdf	W06, W08	OmniForm output whose hidden widgets pdf.js drew outside the page (issue 4914)
us-federal/pdfwriter3-copyright-office-dmca-summary-1998-rc4-40.pdf	W10	Acrobat PDFWriter 3.02 (1998), RC4-40, a malformed creation date
us-federal/pdfwriter4-usda-dry-whey-standard-2000.pdf	W10	Acrobat PDFWriter 4.05 (2000): page thumbnails, ASCII85 over LZW
us-federal/distiller3-mac-msha-crusher-dust-card-1997.pdf	W10	Distiller 3.0 for Power Macintosh (1997): non-embedded Type 1 fonts, eight content streams on a page
us-federal/distiller6-cdc-west-nile-chinese-traditional.pdf	W09, W10	Traditional Chinese from Word through Distiller 6 (2004), ArialUnicodeMS as Identity-H
pdf-association/handwritten-compacted-syntax.pdf	W06	Every token pairing with no whitespace, and an empty object; xpdf 4.00 extracts nothing from it
pdf-association/handwritten-dual-startxref.pdf	W06	Two startxref lines, of which only the last is right
pdf-association/handwritten-utf16le-strings.pdf	W06	Text strings in UTF-16LE, which the standard forbids and real files contain
pdf-association/handwritten-content-stream-indirect-refs.pdf	W06	Indirect references inside a content stream
pdf-association/handwritten-inline-image-abbreviations.pdf	W06	Inline images whose abbreviated and full keys disagree

Third pass: the leads held back by their license

The third pass went through the leads the first two had held back for their license, in eleven groups: signatures (pdfcpu's test data, EU DSS's test resources, Foxit's files in pdfium_tests, node-signpdf, the BOE's sealed gazette, and a hunt for commercial e-signatures), real invoices and statements, vendors' Factur-X and ZUGFeRD samples, the government leads, files that broke pdf.js, PDFBox, PDFium and pdfplumber, and the ShareAlike sets. Most of what it found may be used but not redistributed, and joined the remote corpus (below). Six files could be committed after all: two that node-signpdf made itself, which its MIT license does cover; the ZUGFeRD corpus maintainer's own factur-x output; two government leads whose license was verified at the source; and, from the hunt, a GSA contract modification completed in DocuSign — the first commercial e-signature the committed corpus holds. All six passed a last, independent review of their bytes, license and personal data. Three files proposed for committing went remote instead: the USDA's Hebrew fact sheet, whose banner carries a symbol the USDA reserves; the PDFpen file from OCRmyPDF, whose REUSE.toml entry the issue record contradicts; and a truncated FEMA form, public domain but not the publisher's bytes.

File under tests/corpus/vendor/	W	What it brings
node-signpdf/skia-chrome74-node-signpdf-reason-contains-trailer.pdf	W05, W06	node-signpdf's own Chrome 74 print, signed in an incremental update whose /Reason contains the keyword trailer; startxref on the line end before xref
node-signpdf/pdfkit-node-signpdf-unsigned-placeholder.pdf	W05	An unsigned signature placeholder from PDFKit: name tokens in /ByteRange, 8,192 zero bytes of /Contents
us-federal/docusign-pdfkit-gsa-sf30-contract-modification.pdf	W05, W03	A commercial e-signature: DocuSign's envelope seal with a FieldMDP reference (MD5) over a JBIG2 scan on a page rotated 270°, validation data appended in an update, a creation date in year 0001
zugferd/pypdf2-facturx-python-false-pdfa3b.pdf	W07, W12	Factur-X BASIC from the factur-x Python library through PyPDF2, with a PDF/A-3b claim veraPDF rejects on five rules — no trailer /ID, no binary comment, no output intent, a bare metadata stream
us-federal/finereader8-frb-sr0115-examiner-guidance.pdf	W03, W02	ABBYY FineReader 8 OCR beneath 300 dpi CCITT G4 page images, tagged, in a linearized hybrid-reference file
uk-ogl/pdfmaker21-ozev-sample-invoice.pdf	W02, W04	PDFMaker 21 for Word, tagged, a "Sample" watermark in an Acrobat optional-content group set in a complete Arial

Fourth pass: the Open Preservation Foundation's format-corpus, file by file

The survey had marked the OPF format-corpus usable, and the first three passes took nine files from it, each to fill one W line. On 2026-09-25 the maintainer asked why the rest had not followed, and then for all of it that could come in: committed where possible, remote otherwise, because producer diversity is what the corpus is for. The record held no answer: none of the other 284 PDFs at the pinned commit had been examined, turned down or listed as a lead.

Every one of the 284 was screened against the four rules — text of every page, /Info, XMP, annotations and their authors, form values, XFA datasets, embedded files (a Quattro Pro and a Lotus 1-2-3 sheet, two Word files, videos, a sound), signature dictionaries, the literal strings of every object and every inflatable stream including superseded revisions, and page images viewed wherever a file has no text layer. One independent verifier, told to reject on doubt, then went over the 131 files admitted, OCR included: it turned down six, moved a seventh to the remote corpus for the copyrighted tables in it, and flagged two more that were then excluded on doubt — one of which, the SAMHSA fact sheet, the maintainer later admitted to the remote corpus with a written exception (In the remote corpus). Four of its finds were in text the first screen could not reach — an e-mail address with a space after the @, a named nurse's extension split across two lines, contact details printed only inside a scan. Expectations come from the referees below, established on the files; the reader was then run against them.

56 committed, 6.2 MB, all under 1.1 MB:

Under tests/corpus/vendor/opf-format-corpus/	W	What it brings
The Cabinet of Horrors, 16 files: pdfmaker9-word-distiller-* (one byte missing, embedded QuickTime, external link, a Quattro Pro attachment, fonts in full and as subsets, three AES-128 permission variants and one under an open password nobody published), pdfmaker10-word-file-attachment-annotation, pdfmaker11-word-pdf17-header-says-18, distiller7-pscript5-calisto-mt-not-embedded, distiller952-pscript5-kb-pdf-risk-inventory, make-pdf-javascript-openaction	W02, W06, W12	One feature isolated per file from the same Word document: AES-128 (R4) under three permission sets, an open password that cannot be supplied, JavaScript, attachments in the name tree and in an annotation, a movie, a link to another file, a header naming PDF 1.8, a byte removed so every offset is one out, fonts left out
acrobat11-image-conversion-*, three files	W06, W12	A PDF/A-1b page from Acrobat 11 Image Conversion, and two copies damaged on purpose: the image's /Height changed, the image pointed at an object that does not exist. veraPDF upholds all three claims
The Save As corpus, 11 files: openoffice32-writer-* (a one-line document, Lucida Sans, PNG and TIFF images, PDF/A-1a, and 128-bit RC4 (R3) with and without an open password), openoffice33-writer-simple, libreoffice35rc3-writer-simple	W12	OpenOffice.org 3.2 and 3.3 and LibreOffice 3.5 on one document, and OpenOffice's own RC4 handler with the password documented
quartz-pages09-lorem-ipsum, openoffice32-writer-lorem-ipsum and -pdfa1	W01	Pages '09 through Quartz, beside the Word for Mac variation already vendored
ibooks-author11-quartz-* and ibooks-author20-quartz-*, six files; calibre0857-pypdf-lorem-ipsum-ebook, calibre090-podofo-lorem-ipsum-ebook	—	iBooks Author through Quartz, three of them under 40-bit RC4 (R2) with an open password; pyPdf and PoDoFo, two writers nothing else in the corpus uses
indesign-cs-puppet-guild-event-flyer	W02	InDesign CS (3.0) through Adobe PDF Library 6.0, waived to the public domain by its author
GovDocs1 error files, 15: nine from the Department of Veterans Affairs (PDFMaker 5 to 8 for Word and PowerPoint, Distiller 5.0.5 behind ADOBEPS4.DRV, PScript5; one with eight incremental updates), one from the Census Bureau, two from the USGS (one typeset by groff), an enrolled bill, a House committee markup whose content streams pdftotext cannot finish, a Law Library of Congress report	W06, W10	Federal staff's own work from 1999 to 2008, out of the Distiller 4 and 5 generation and a few producers no one else ships

68 remote, 216 MB, listed under In the remote corpus. 160 refused:

• 89 for their bytes: the OPF's copy of the iPRES 2017 hand-built set (pdf-handbuilt-test-corpus/), 87 test files, each derived from one page with one deviation from ISO 32000-1's structure, and two reference files. Its authors published it on RADAR (doi:10.22000/53) under CC BY-SA 4.0 — so it would have been remote at best — but the OPF copies are not those files: each test file lost the five carriage returns of its content stream when it was committed to git, so its /Length and every later offset are wrong and every file shows damage its authors did not put there. The RADAR tar was downloaded and compared: all 87 differ by exactly those bytes. So does hello_world.pdf, the page every test file derives from: with its five carriage returns put back, it has the MD5 its authors list. minimal_test.pdf is intact, but holds only a header and %%EOF. The tar has neither reference file, and one test file the OPF lacks, T04_019, a trailer pointing at the wrong cross-reference offset — the authors' own checksum list, inside the tar, names exactly the OPF's 89 files, so the copy predates it. The 88 test files are now fetched from the tar itself (ADR 33; In the remote corpus).
• 68 for personal data. 51 of the JHOVE files, journal articles, theses and posters from bug reports: 49 print an author's, a contact's or a student's own e-mail address or direct line — three only inside a page image, two with a space after the @ —, one is a repository deposit agreement with its author's phone number, e-mail and signature, and one is an essay in which its named author describes her own health and her partner's. 16 of the GovDocs1 files: named staff's direct lines or extensions and own e-mail addresses, from the VA (a nurse's extension among them), the CDC (in the XMP only), the GAO, a state agency and contractors; directories of private e-mail addresses; a cadet newsletter naming minors beside a phone number of unclear owner. And the rich-media sample from fully-featured-pdf/, excluded on doubt: its embedded iPhone video carries the position and time at which an identifiable contributor filmed it.
• Two for a restriction of purpose: a JSTOR copy whose terms allow personal, non-commercial use only, and a SHRP 2 report the National Academies let be reproduced for classroom and not-for-profit purposes only.
• One not a PDF: a 213-byte macOS AppleDouble resource fork saved under a .pdf name.

Our own, from Windows

Our own additions for W01 are in documents/invoice/: the same fictitious invoice through Word's Save as PDF, the Microsoft Print to PDF driver and the PDF24 printer (build/build_word.ps1). The PDF24 file draws a box for every capital E of the regular face — the driver's PostScript lacks the glyph, while ToUnicode still says E — which makes it a W06 file as much as a W01 one.

W11, by reference only

Document	Size	SHA-256 on 2026-09-24
United States Code, 2023 edition, Title 42 — 9,302 pages, a GPO signature in an incremental update (govinfo)	37,632,691 bytes	9384648a70cf217b7dfff872ebd699b24a59dce55c76cdc06c802810782d0350
USGS Professional Paper 1 (1902) — 125 pages of JPEG 2000 up to 29 megapixels each, with an OCR layer (USGS)	146,708,859 bytes	50b5ea3fecbbda5f74fcd4dd53f348d697c18d44f1a6ff1a42fd1f0031c5ee3d
US Topo map, Washington West, 2023 — one page, a 109.5-megapixel Flate image and 31 layers (USGS)	63,128,959 bytes	33cfc95870f64560dcdedc616d3e666f6f99e485125f290cc9963e1db2d9f6d5

All three are US federal works. Publishers reissue such files, so the hash says which bytes a measurement was taken on. All three are now in the remote corpus (below), fetched and tested every night at these hashes: the heavy scan joined the other two in the third pass, so M23 can state its memory budgets against every one of them.

What other PDF libraries ship

Almost every PDF library keeps a corpus. Very few could be reused here, and the reason is nearly always the same: the files are attachments to bug reports, committed under the repository's code license by someone who did not own them.

Library or corpus	Test PDFs	What covers the PDFs	Usable here
veraPDF corpus	~2,900	CC BY 4.0; the Isartor folder forbids redistribution	Yes, already vendored; Isartor excluded
BFO PDF/A suite	34	CC BY 3.0	Yes
PDF/UA Reference Suite (PDF Association)	19	CC BY 4.0	Three members committed (2-02, 2-05, 2-10); three remote — 2-06, whose cover photograph is a third party's, and 2-08 and 2-09, publishers' textbook chapters, both over 2 MB as well; 2-01 refused for named people's health data, the papers and presentation for their authors' e-mail addresses
PDF Association SafeDocs artifacts, pdf-differences	~60	Apache-2.0, CC BY 4.0	Yes, the hand-written files; several others name their author
OPF format-corpus	293	CC0 for what OPF's contributors made — the Cabinet of Horrors by its own statement, the Save As corpus, the ebooks and the Lorem Ipsum variations by the repository's default; GovDocs1 crawls and JHOVE bug attachments carry only their publishers' terms; the hand-built set is CC BY-SA 4.0 on RADAR	OPF's own files, and the GovDocs1 files that are federal staff's work: 64 committed since the fourth pass, 69 remote. The hand-built set's copies lost bytes to git; its originals are fetched from RADAR instead (ADR 33)
pikepdf	36	Per file, in REUSE.toml	The CC0 and public-domain files; not the CC BY-SA ones
qpdf	~720	Apache-2.0 for the maintainer's own files; nothing for issue attachments	The maintainer's own files
pyHanko	~130	MIT; files generated by pyHanko against a fictitious test PKI	Yes
Mustang, ZUGFeRD corpus, quba-viewer	~190	Apache-2.0 for the repositories; nothing specific for vendors' samples	The maintainer's own output
Apache FOP	26	FOP's own output, Apache-2.0	Yes, but no invoice among them
pdf.js	~980 committed, 459 linked	Nothing stated; mostly bug attachments	Only the government files its links point to
PDFium, Chromium	~1,100	BSD for hand-written .in files; nothing for bug files and the Foxit QA suite	Hand-written files only; Foxit files name employees
Apache PDFBox, Apache Tika	~250	Nothing for JIRA attachments and GovDocs1 extracts	No
PdfPig, pypdf, pdfminer.six, pdfplumber, camelot, tabula, hayro, pdf-rs, lopdf, pdfcpu, Docnet, endesive, pdfrw	a few to ~250 each	Nothing stated, or "can be distributed freely" without a license	No
Poppler, MuPDF and PyMuPDF, Ghostscript, iText, OpenPDF, EU DSS, HexaPDF, Origami, borb, img2pdf	up to ~7,000	GPL, AGPL, LGPL or MPL	No
py-pdf/sample-files, PDF Association pdf20examples	~40	CC BY-SA 4.0	No: ShareAlike
Syncfusion samples, UniPDF	—	Commercial terms, or private	No
QuestPDF, ReportLab, WeasyPrint, Xpdf	none	—	—
GovDocs1, SafeDocs CC-MAIN-2021-31, SafeDocs issue-tracker corpus, Common Crawl	millions	None: each file is its publisher's	Only files whose publisher's license qualifies, found one by one
Internet Archive, Wikimedia Commons	millions	Declared per item by uploaders, unverified	With the publisher's original bytes only

"Usable here" means committable. Since ADR 32, the remote corpus also uses files from pdfcpu, EU DSS, pdfium_tests, node-signpdf, pdf.js, PDFium, PDFBox, PdfPig, pdfminer.six, pdfplumber, Mustang, the ZUGFeRD corpus, OCRmyPDF, py-pdf/sample-files, pdf20examples, pdf-differences, the PDF/UA Reference Suite and the OPF format-corpus: see In the remote corpus.

Practices worth borrowing

• pikepdf's REUSE.toml gives every test file its own SPDX license and copyright line, often with a link to the issue where the owner granted it. It is the only corpus examined whose provenance is complete, file by file.
• pdf.js commits a .link file holding a URL — often an Internet Archive copy — for files it may not redistribute, downloads them before the tests run, and checks each against an MD5 in its manifest. PDFBox does the same at build time, pinned by SHA-512. That is the model for W11.
• py-pdf/sample-files records the producer and creator of every file in a files.json, beside the source each was made from.
• veraPDF encodes the expected verdict in each file's name, so a fixture is self-describing.

The manifest now records the source URL, date and SHA-256 of every third-party file, which gives this corpus the provenance pikepdf has.

What was turned down, and why

Candidate	Reason
A Journal officiel des associations notice	The registered office is a named couple's home
Another such notice	The association's purpose describes a named child's medical care
PDF/UA Reference Suite papers and presentation	Their authors' own e-mail addresses are printed on the page
CDC Emerging Infectious Diseases article (1997, PDF 1.1)	The author's own e-mail and fax, in an "Address for correspondence" block
HUD charge of discrimination (Xerox, the copier's own OCR)	Health data tied to an identifiable complainant
Hetzner invoice (Apache FOP), in the ZUGFeRD corpus	A private customer's home address
GnuAccounting invoice of May 2014	A sole proprietor's tax number in the attached XML
JasperReports invoice sample (iText 2.1.7)	Third-party template images the author's license does not cover
National Park Service lesson plan (Canon iR-ADV scan)	The NPS arrowhead is excluded from the Service's public-domain statement
European e-Justice page (iText 5)	The EU reuse decision excludes logos, and the page embeds the portal's
IRS Form 1040 for 1994 (Net Distiller 1.02)	Complete Type 1 fonts whose embedding permission cannot be checked
NIST request for quotation scanned on a Canon SC1011 (3.9 MB)	Two staff members' direct phone lines and own e-mail addresses; and Acrobat had re-saved it, so it was not the untouched copier file W03 wants
PDF/UA Reference Suite 2-01, a Danish magazine	Health conditions of named people, children among them
California Geological Survey Note 17, linked by pdf.js	Reproduction for classroom or public education only: a restriction of purpose, which the remote corpus refuses too
Mustang's Oracle Reports invoice (ZUGFeRD EXTENDED)	Its attached XML gives an employee's direct line and e-mail
A Business Central invoice from Aspose.Words, among Mustang's tests	Its attached XML gives a real person's e-mail
An Oracle PL/PDF insurance invoice, in PdfPig's tests	A named contact person's mobile number
An AWS invoice from Apache FOP 0.95, in invoice2data's tests	The billing address is, in all likelihood, a person's home
Symtrax's Factur-X from SAP demo data, with a false PDF/A-3a claim	A seller account that passes every IBAN check and whose owner is unknown: excluded on doubt
Adobe's well-known signed sample, in pdfcpu's test data	The signer certificate carries an employee's own e-mail, printed again in the visible signature
A qualified signature among pdfcpu's samples	A real person's qualified certificate, with a persistent holder identifier
Most of the signed files in EU DSS's test resources	Real people's certificates: national identity or birth numbers, personal e-mail addresses and phones, a home address; and a Commission seal whose pages give an official's direct line
A notarial mortgage deed under a commercial e-seal, in EU DSS's test resources	A named person's home address and passport number, perhaps fabricated: excluded on doubt
Public bodies' contracts signed through DocuSign or Adobe Sign, with the audit report	The report lists each signer's e-mail and IP address
Ten damaged files from pdf.js, PDFBox and pdfplumber bug reports	Personal or health data: employees' e-mails on orphaned pages, a private car's registration, a household's electricity bill, a patient's blood test

Under the first, stricter rule, a dozen more had been turned down for a name alone — a civil servant in a superseded /Info, an employee's Windows path in an XFA packet, a photographed sales representative, a /Creator nobody could show to be fictitious. All of them entered in the second pass. Four files had already been readmitted in the first pass after the maintainer's rulings: the two EU Official Journal notices whose EUAlbertina fonts are embedded in full under an editable-embedding permission, the 2026 consolidated regulation whose only name is a typeface designer's credit inside a font, and the 1786 engraving under ImageMagick's false PDF/A claim, whose only names are historical figures.

Leads for what is still missing, held back by their license

These files would fill a gap the corpus still has, and are in it neither committed nor remote. They are listed so that a later session, or a rights holder asked directly, can reopen them. The third pass worked through what this table used to hold: most of it now sits under In the remote corpus (ADR 32), six files turned out to be committable after all (Third pass above), and the Oracle Reports invoice was turned down for personal data. What remains open is a note the remote corpus may not use, an invoice addressed to a private person, and files no one has yet screened or checked — most of them in the bug-report corpora.

W	Document	Where	What holds it back	What would clear it
W06, W10	California Geological Survey Note 17: PageMaker and PDFWriter 4 for Power Macintosh (2002), re-secured with RC4-128, 28 CCITT masks tiling one map	pdf.js test/pdfs (issue 4926, a link to an Internet Archive capture)	Reproduction permitted "for classroom or public education purposes" only: a restriction of purpose, which the remote corpus does not accept either	The Department of Conservation's permission
W04	A real Hetzner invoice, Apache FOP 1.0	ZUGFeRD/corpus unstructured/	Hetzner's document, addressed to a private person	Not recoverable
W06	The rest of the bug-report files in pdf.js (~980), PDFium, pdfplumber, PDFBox's JIRA downloads, and the SafeDocs issue-tracker corpus (>32,000); nineteen went remote in the third pass	see the survey table above	Attachments, licensed by nobody: the remote corpus can take them, but each needs its own personal-data screen — of the 31 the third pass examined closely, it turned down 10 for personal data and 2 whose damage was made by hand	A later pass. The third pass named reserves: PDFBOX-4490 and PDFBOX-3977, screened and clean; pdf.js issue5549 and issue5567 for JPEG 2000, issue8702 and issue14814
W06, W08	The rest of the files pdf.js links to: US state documents, Canadian forms, forms attached to Bugzilla	pdf.js test/pdfs	Terms that are not attribution-only, so remote at best; the third pass took three, and the attached forms are often filled in by real people	A personal-data screen each
W08, W05	A filled form with Reader usage rights and a signature, in EU DSS's resources (pades-signed-filled-form.pdf)	EU DSS dss-pades resources	LGPL-2.1, so remote at best; its field values were not screened in full	A full screen
W06	The PDF Association's Brotli prototype, a /BrotliDecode filter qpdf does not know, among pdf.js's test files	pdf.js test/pdfs	Not a bug-report file, but its license was not checked	A look at the PDF Association's own repository

Two good files first held back for a name rather than a license — the PDF Association's CompactedPDFSyntaxTest.pdf and Docentric's Factur-X EXTENDED sample from Dynamics 365 — entered in the second pass, once a name alone no longer disqualified a file.

In the remote corpus

242 documents are used without being redistributed (ADR 32): the manifest pins each one, with origin remote, to an immutable URL, a SHA-256 and a size — or, for 88 of them, to a member of a pinned archive (ADR 33) —, and the Remote corpus workflow fetches them — some 500 MB, nearly a third of it one scan — and runs both suites over them every night. Their expectations were established like everyone else's — pages and verdicts by qpdf 11.9.1 in the integration tests' container, text by poppler's pdftotext 24.02, conformance by veraPDF 1.30.2 — on 2026-09-24: ten during the day, 76 in the evening's third pass. For those 76, whether a file opens clean and whether its index must be rebuilt were derived from qpdf's warnings, not from our reader; the reader was then run against them, and where it falls short the entry is marked unsupported, with the reason and the milestone that owes the fix.

Rule 2 holds here with one exception: a remote document is screened for personal data exactly as a committed one. The exception was made by the maintainer's decision of 2026-09-25 and is written into its entry (feature partially-screened): a SAMHSA fact sheet from GovDocs1, sent at some point through a text-mode transfer that turned each of its CR and LF bytes into CR LF — 1,904 pairs, not one lone CR or LF left — and dropped every 0x1A byte, of which none is left where every other byte value occurs 398 times or more. Every stream's /Length, every offset and every Flate stream broke; as the file stands, its three content streams stop decoding after 18 to 96 bytes. Undoing the damage is a search with many answers: put back 0x1A bytes and single line ends, and the opening of each page and the third page's references read with certainty, but no attempt reached a stream's own checksum, and past those openings the reconstructions drift into text that cannot be trusted. What reads is clean — screening guidance and questions, citations of published authors, the metadata, the annotations, a link to a SAMHSA center —; the rest could not be recovered with certainty and was not screened. The entry publishes nothing from the document beyond its publication name. What it adds is a whole file broken by a text-mode transfer, which nothing else in the corpus is. Rules 1, 3 and 4 are what the remote corpus relaxes, within limits the maintainer set in the third pass. Since nothing is redistributed, terms that restrict reuse do not keep a document out — conditions beyond attribution, non-commercial reproduction only, nothing beyond fair use, no modification or commercial use. A restriction of purpose does: a note that may be reproduced "for classroom or public education purposes" only was left out.

W	Document	Why it is remote	What it adds
W01, W08	OPM Optional Form 306, from an Internet Archive capture, since opm.gov serves whatever edition is current	No OPM reuse statement, so public domain rests on 17 U.S.C. 105 alone	A Word for Microsoft 365 document made into a tagged AcroForm by Acrobat, Producer left as Word: 40 fields with date JavaScript, two unsigned signature fields
W01	A Border Force poster, from GOV.UK	OGL v3, except for three other bodies' logos it carries	Microsoft Print To PDF with the text as outlines and no font at all, then an Adobe update that rewrote /Info and added XMP
W03	Ricoh MP C3003 scan through 3-Heights (pdf.js issue5747.pdf)	A bug-report attachment, licensed by nobody	A copier's CCITT page with an OCR layer in a non-embedded Identity-H font; a valid PDF/A-1b
W03	A Konica Minolta bizhub C554e scan of a regional council's letter, untouched (Environment Canterbury, fetched from the Internet Archive's capture)	Reuse for personal, informational and non-commercial purposes only	The untouched copier file W03 lacked: mixed raster content — a JPEG background under CCITT image masks — and the copier's own invisible OCR in a non-embedded Arial
W03	A Kodak Capture Desktop scan and an Epson device scan, from OCRmyPDF's test resources	CC BY-SA 4.0 and CC BY-SA 3.0	One CCITT G4 page each, and structural slips: an outline whose /First names a missing object and whose /Last points at itself; a Pages node nested in another, /Rotate as an indirect object, no /Info
W03	PDF/UA Reference Suite 2-09, a scanned textbook chapter remediated by AbleDocs (10.6 MB)	Over 2 MB, and the pages are a publisher's that the suite's CC BY cannot be shown to cover	A tagged scan under a PDF/UA-1 claim veraPDF upholds: 82 pages whose OCR text is drawn first and the page image painted over it as a pagination artifact, 480 formulas with spoken alt text, JBIG2 without globals, a catalog /Lang of (English)
W04	SAP NetWeaver 7.40 form output (pdf.js bug1727053.pdf)	A Bugzilla attachment, licensed by nobody	A real statement from an ERP, the one kind of W04 file the committed corpus still lacks
W04	An SAP NetWeaver 7.40 invoice and an AFP Batch Processor bank statement excerpt, from pdfminer.six	A supplier's and a bank's documents, kept under a code license that does not cover them	SAP: indirect /Filter arrays, ArchiveLink comments inside /Info, one logo embedded twice. AFP: bitmap Type 3 fonts with a named encoding and no ToUnicode, later re-saved by an Adobe application
W04, W06	An Axapta credit note and a Scoro invoice, from PdfPig	Suppliers' documents attached to bug reports; PdfPig's Apache-2.0 covers its code, not them	Axapta's own writer: an /Info object without endobj, non-embedded Identity-H fonts, a ToUnicode that turns the minus into a soft hyphen. Scoro: a Chromium page re-serialized by qpdf, then an ExifTool update that adds /Info and XMP
W04, W06	A card statement page from PDFlib+PDI on IBM z/OS (pdf.js issue6605.pdf)	A bug-report attachment, licensed by nobody	EBCDIC character codes in bitmap Type 3 fonts; a Nitro Pro update leaving 115 orphaned objects
W04	A central bank's rates bulletin from JasperReports on iText 2.1.0	The bank's output in an MIT repository whose grant does not reach it; the bank's terms forbid modification and commercial use	A real seven-column table in non-embedded Helvetica
W05	Three PAdES levels on one Word 2019 page — B-B, B-LT, B-LTA — signed with the Ukrainian DIIA test CA, from pdfcpu's test data	pdfcpu states no license for its test data	The baseline levels one at a time: CAdES detached, a signature timestamp, a DSS of certificates and OCSP responses without VRI, a document timestamp lacking /Filter; every update glued to the %%EOF marker
W05	Avow Systems' certified sample and Tecxoft's signed web capture, from pdfcpu's test data	Vendors' samples, no license	DocMDP P=1 with FieldMDP and legacy MD5 object digests, embedded CRL and OCSP. The legacy adbe.x509.rsa_sha1 subfilter: a raw PKCS#1 value, the certificate in /Cert, over a three-page Acrobat Web Capture 8.0 capture; unsupported for T24 (see Traps)
W05, W02, W12	The Slovak National Security Authority's supervision scheme under its qualified seal, from EU DSS	LGPL-2.1, and no reuse license from the authority	A certification signature (DocMDP P=2 with FieldMDP, MD5 references) saved by Acrobat 11 into a linearized file; a PDF/A-1a claim veraPDF rejects
W05	Two ETSI plugtest files, from Hungary and France, in EU DSS	LGPL-2.1	Two qualified seals, RSA then ECDSA, each followed by a document timestamp, and a Producer written as a byte-order mark before single-byte text; a PAdES-EPES signature with claimed roles under SHA-512, then two document timestamps
W05	A Notepad printout re-signed over seven years, from EU DSS	LGPL-2.1	A 2013 signature, document timestamps in 2013 and 2019, a 2020 signature in RSASSA-PSS: five incremental updates, two startxref one byte early
W05, W11	An Excel sheet with 24 signatures and a document timestamp, from EU DSS	LGPL-2.1	48 incremental updates; a DSS of 31 certificates, 29 OCSP responses and 24 VRI entries. Its 10,112-byte /VRI dictionary crosses the reader's 8 KB window, which made it unsupported for T23 until that was fixed on 2026-09-26
W05, W06	EU DSS's infinite-loop regression file (DSS-1872)	LGPL-2.1, over a chart under CC BY-SA	A signed update nobody can reach: startxref past the end of the file, wrong offsets, a signature widget that is its own /Parent; it hung xpdf 4.00
W05, W02	The Spanish Official State Gazette: a 2015 law, a 2026 royal decree, and a decree re-signed among EU DSS's resources	The BOE's reuse terms add conditions beyond attribution; the third is LGPL-2.1 as well	The gazette's seal — adbe.pkcs7.sha1, no signed attributes, an RFC 3161 timestamp as an unsigned attribute — over PDF/A-1a claims veraPDF upholds; the law's 36 structure arrays of about 8.7 KB each cross the reader's 8 KB window, which made it unsupported for T23 until that was fixed on 2026-09-26; the re-signed decree adds a test signature, and veraPDF rejects its claim
W05, W03	Commercial e-signatures: a DocuSign envelope export (EU DSS), a Maine contract amendment completed in DocuSign, an Adobe Sign test agreement (pdfcpu issue 389), a Yousign-sealed test page (qpdf issue 1469)	LGPL-2.1 over DocuSign's own form; a state work with no reuse statement; two issue attachments	Beside the committed GSA file: DocuSign's seal extended by a DSS and a Czech demo TSA's document timestamp; a Ricoh copier scan on rotated pages sealed through iTextSharp; an ETSI.CAdES.detached certification under AES-128 with an empty user password; a qualified e-seal written by a full rewrite, over read-only fields
W05	Three Foxit PhantomPDF signatures, from pdfium_tests	Foxit's QA files; the repository's BSD license covers the PDFium authors' own work	Certification and approval signatures written in a single full save, the byte range covering the whole file, CMS without signed attributes; in one, a BBox of garbage reals inside the signed bytes leaves a string open to the end of the file
W05	An OpenOffice.org page signed twice by node-signpdf	The base document is W3C's test file, whose license is not node-signpdf's	Two signatures in two updates, CMS signed attributes not in DER order, startxref on the line end before each xref
W05, W02, W08	Acrobat 9's signed 3D portfolio, from the Open Preservation Foundation (4.7 MB)	CC0, but over 2 MB	The first PDF portfolio: a collection schema, folders, a Flash navigator and five embedded PDFs, among them an XFA form filled with fictitious values under a certification signature, and U3D and PRC 3D models
W06	Four files from pdf.js's tests that break structure: an Aspose.CAD drawing written over an older, longer file; a Firefox print through cairo; a Canon scanner page; Aspose.Pdf for Java output	Bug-report files, licensed by nobody	A stale tail whose trailer names /Info as /Root, so qpdf and poppler find no page; 65,542 objects in one object stream, their indexes wrapped at 16 bits, so poppler loses the catalog; junk after %%EOF and a corrupt JPEG under Flate; /DecodeParms [null] on every stream and a /ToUnicode that is a name
W06	Five more from pdf.js's tests, on fonts and images: typeset.sh, Oracle Outside In, Esri ArcMap, Acrobat 8 image conversion, iText 5	Bug-report attachments, licensed by nobody	A FontFile2 whose deflate data is corrupt, and text before any Tf; a PDF/A-1a veraPDF passes over a Calibri subset with a broken loca; a CFF FDSelect starting at glyph 1; JPEG 2000 with several precincts per level; a CCITT G4 stencil as an explicit /Mask
W06, W11	The Massachusetts COVID-19 dashboard of 25 October 2020, linked by pdf.js	Mass.gov forbids any copying beyond fair use	A Power BI export through PDFium, linearized and then saved nine more times by Acrobat, which kept PDFium as Producer: 24 pages at the end of a long /Prev chain
W06	A wine merchant's product sheet, from pdf-differences' UnknownFilter set	A third party's document under the PDF Association's CC BY	A PNG stored under /DCTDecode; Identity-H subsets without ToUnicode; no /Info
W06	Four files from PDFBox's JIRA downloads, damaged at their end: Amyuni PDF Converter output cut 1,950 bytes short, a Distiller 8 file with 7.6 KB of UTF-16LE text after %%EOF, a FrameMaker 10 budget book and a Word 2010 FEMA form each cut at 1 MB	Web-crawled bug attachments; the FEMA form is public domain, but a crawl cut short is not the publisher's bytes	A lost main cross-reference section and a /Prev past the end; startxref hidden from a 1 KB tail search; and no page count asserted for the two cut at 1 MB, since qpdf 11.9.1 finds no /Root in one, and qpdf 11.9.1, PDFium and qpdf 12 count 11, 8 and 4 pages in the other
W06, W10	Three more from PDFBox's JIRA: a 1998 PDFWriter 3.02 file whose CR LF line ends were stripped, a Distiller 6 report with a 32 KiB block zeroed, an Acrobat form with #00 in names	Bug attachments; the report is a public-domain USGS work, but exists damaged only there, and weighs 2.1 MB	Offsets 3 to 175 bytes off and startxref 184 off; one object stream destroyed, which the reader detects only when it reaches it, so the entry expects no rebuild at opening and the rebuild diagnostic after a full read; a NUL in names used as values and keys, which the reader accepts silently until M02's third slice reports it
W06, W09	PDFium's bug_182.pdf, and two pdfplumber reporters' files	Bug-report files, licensed by nobody	UTF-16BE /Info strings opening with a language escape; an inverted MediaBox over a Flate-over-DCT scan; Chinese font names written as GBK bytes in #xx escapes
W06	A PDFpen page of text imprints, from OCRmyPDF	Its contributor agreed to CC BY-SA 4.0; OCRmyPDF's REUSE.toml now says MIT, with no MIT grant on record	Seven content streams whose q and Q fall in different streams; PDFpen's private keys and plist streams
W07	The FNFE-MPE's official Factur-X example, BASIC WL, in French	All rights reserved	factur-x.xml under /AF, a valid PDF/A-3b, written by the factur-x Python library through PyPDF2
W07	intarsys's ZUGFeRD 2.0 EN 16931 sample	A vendor's sample, no redistribution license	A second vendor's toolkit, zugferd-invoice.xml, a valid PDF/A-3b
W07, W12	Two DWC FX Generator invoices through WeasyPrint, among Mustang's tests	"© DWC 2025" on the page, and no redistribution license	Factur-X EXTENDED in PDF 2.0 under a valid PDF/A-4f claim, two RDF blocks in one XMP packet, factur-x.xml listed twice in /AF; a sibling whose XMP has no Factur-X schema, under a valid PDF/A-3u claim
W07	The official Order-X example, COMFORT profile, among Mustang's tests	FNFE-MPE's and FeRD's example, no redistribution license	A purchase order rather than an invoice: order-x.xml, from the factur-x Python library through PyPDF4
W07, W12	Four vendors' samples from the ZUGFeRD corpus: intarsys (ZUGFeRD 2.2, XRECHNUNG profile), Symtrax Compleo through iTextSharp 4.1.0 (MINIMUM), 4s4u's additional-data library (ZUGFeRD 1.0 EXTENDED), Konik (ZUGFeRD 1.0 BASIC)	Copies of vendors' packages with no redistribution license; Konik's is AGPL-3.0	xrechnung.xml attached by an incremental update; a valid PDF/A-3a with a whole Times New Roman embedded; a second XML attached as /Supplement, CR-only line ends; one subset tag shared by four font programs
W07, W04	An ERP's test invoice made hybrid by iText Core 9 (quba-viewer issue 143)	An issue attachment, licensed by nobody	A factur-x.xml that holds a UBL invoice rather than CII
W08, W05	Two Canadian dynamic XFA forms, linked by pdf.js: an immigration form (IMM 1344) and a fish export license application	Government of Canada terms: non-commercial reproduction only	The first dynamic XFA, whose only PDF page is a "please wait" placeholder; AES-128 with an empty user password; a certification with a timestamp, then UR3 usage rights in an update; legacy /UR beside /UR3
W09	WeasyPrint 54.1 Arabic (py-pdf/sample-files)	CC BY-SA 4.0	Arabic shaped into CID subsets by an HTML-to-PDF engine — this project's own kind of producer
W09	US Census Bureau 2020 language guide in Hebrew	No reuse statement found; possibly a contractor's translation	Hebrew right to left in Adobe Hebrew Type 1 subsets, tagged with /Lang he
W09	USDA Title VI fact sheet in Hebrew, from an Internet Archive capture	The USDA reserves the symbol on its banner; the translation's provenance is unknown	Hebrew from Word through PDFMaker 23, in TrueType and CID subsets beside the Census guide's Type 1; Hebrew outline titles
W11	United States Code 2023, Title 42	37.6 MB, too large to commit	9,302 pages and a GPO signature in an incremental update; opening, decoding every stream and walking the page tree took 0.8 s on the development machine
W11	US Topo map, Washington West, 2023	63.1 MB, too large to commit	One page with a 109.5-megapixel image. The false truncated-stream report it first exposed (T21) is fixed. The image decodes to 328,608,000 bytes, past the 256 MB a stream decodes to by default: the defaults keep the first 256 MB and report limit.decoded-stream, and the entry is opened with readerLimits at 512 MB, under which it reads whole and clean (ADR 34)
W11, W03	USGS Professional Paper 1 (1902), scanned in 2017	146.7 MB, too large to commit	125 JPEG 2000 page images under an invisible OmniPage 19 OCR layer, tagged; linearization hints that point past the end of the file
W11, W06	A tiff2pdf image of 35,000 × 35,000 pixels, from OCRmyPDF	CC BY-SA 4.0	10.5 KB of CCITT G4 that decodes to 153 MB, on an 8,400-point page under a PDF 1.1 header
W12	Ghostscript 10 output claiming PDF/A-1b (py-pdf/sample-files)	CC BY-SA 4.0	Type 1C subsets without ToUnicode, word spaces only as TJ kerning, an AdobeRGB output intent; veraPDF upholds the claim
W12, W06	A Photoshop CC 2015 page claiming PDF/X-3, from OCRmyPDF	CC BY-SA 4.0	CMYK JPEGs, a SWOP output intent, a CID subset whose ToUnicode values are byte-swapped
W12	PDF/UA Reference Suite 2-08, a textbook chapter remediated by AbleDocs (2.3 MB)	Over 2 MB, and the chapter prints a publisher's and photo agencies' copyrights the suite's CC BY cannot be shown to cover	A PDF/UA-1 claim veraPDF upholds; PDF/A and PDF/E output intents with no claim; CMYK photographs, JPEG 2000 icons with JPEG 2000 soft masks; 1,018 stale hint-table warnings from qpdf
W12	PDF/UA Reference Suite 2-06, the PDF Association's 2012 brochure	Its cover photograph carries a third party's copyright, and the second photograph no credit	InDesign CS6, linearized in 2019 then updated twice, the first update's /Prev pointing at the first-page xref stream; alt texts ending in NUL
—	pdfLaTeX with hyperref (py-pdf/sample-files)	CC BY-SA 4.0	The TeX family, absent until now: Computer Modern Type 1 subsets, outline, xref stream
—	A Google Docs download (py-pdf/sample-files)	CC BY-SA 4.0	Skia's Google Docs renderer, a Type 3 font beside CID subsets
—	wkhtmltopdf 0.12.5 on Qt 5 (py-pdf/sample-files)	CC BY-SA 4.0	Another HTML-to-PDF engine: DejaVu CID subsets, one glyph per text operator in a y-down text matrix
—	The PDF Association's hand-written PDF 2.0 example of UTF-8 strings (pdf20examples)	CC BY-SA 4.0	UTF-8 text strings, which only PDF 2.0 allows, in /Info, outline titles, layer names, a page-label prefix and alternate text

From the OPF format-corpus

The fourth pass added 68, all fetched from the OPF repository at its pinned commit — the SAMHSA fact sheet among them, the exception described above. Their expectations were established the same way, pages by qpdf --show-npages.

Source	Files	Why remote	What they add
The Cabinet of Horrors	2	The AVI file is 2,053,552 bytes, just over the threshold; the Web Capture file quotes pages of the OPF website whose posts the folder's CC0 cannot be shown to cover	An uncompressed AVI behind a Screen annotation; Acrobat Web Capture pages appended to a Word document
GovDocs1 error files	21	Six are federal work over 2 MB — the VHA coding handbook, the VA Kernel guide (464 pages, a PDF/A-1b claim veraPDF rejects on seven rules), a USGS earthquake map, a USFWS recovery plan, the 1994 Transportation Statistics report whose /Producer names Distiller 1.0.2 for Macintosh, a Reclamation EA —; one is a Census Bureau section whose tables are partly copyrighted by the firms that supplied them; thirteen are not shown to be federal staff's work: contractors (ORNL, JPL), PIARC, WARDA, the EU's delegation, IBM, Scholastic, an unnamed consultant; one, the SAMHSA fact sheet, is screened in part	IBM ID Workbench and XPP, Xyvision's Parlance Publisher, WordPerfect through PDFWriter 4, PageMaker 6.5, a /Prev 12 bytes off (T25), a whole file broken by a text-mode transfer
JHOVE error files	45	Files attached to JHOVE's issue tracker, each filed under the JHOVE error it raised (PDF-HUL-n, kept in the file name): articles, theses, reports, posters, scans and a blank IRCC visa form, under their publishers' or authors' terms	A second independent verdict on each file; producers nothing else supplies — tiff2pdf, Apex PDFWriter, Pixel Translations, wPDF, activePDF, FreeHEP, cairo, Skia m89, SignNow, Atypon PDFplus, dvipdfm with PDFStamp, Acrobat 7 Paper Capture, a French Distiller 3.0 —; a MacBinary header and a data: URI prefix before %PDF; a page tree with a null kid; kids pointing at objects the file lacks; a catalog without /Type; a reference to object 0; a file whose tail was lost

Two remote entries are recorded as unsupported: the signed web capture and the VHA handbook (T24). There were twenty-six until M02's second slice made eight of the iPRES files supported on 2026-09-27, eighteen until the first half of its third made twelve more supported on 2026-09-28, and six until the second half made the last four supported: two catalogs without /Type, a page tree root typed /Pagez and a page typed /Font. What follows says what each was recorded for. From the first three passes: the signed web capture (T24), the Axapta credit note and the #00 form (M02); the topographic map was too, for T28, until ADR 34 made its decoding bound an option. From the fourth: the VHA handbook (T24), two catalogs without /Type, the wPDF chapter's null kid, and two page trees whose missing kids qpdf counts as pages (M02); IBM's manual was too, until T25 was fixed on 2026-09-26 — its section is found 12 bytes from where /Prev names it. Five more were, until T21 and T23 were fixed on 2026-09-26: the hospital-bed guidance (T21), the 25-signature sheet, the 2015 BOE law, the VA Kernel guide and a poster (T23). From the iPRES 2017 set, the 17 described below (M02); two more were, until T27 was fixed on 2026-09-27. The laziness test honors that mark, and since 2026-09-29 it alone does for the two documents left (unsupportedTests): every other acceptance test holds them to their expectations. It skips encrypted documents visibly until M16 brings decryption, and it no longer applies to a document whose index must be rebuilt, since a rebuild scans the file by definition.

From RADAR: the iPRES 2017 hand-built set

88 files, fetched out of one BagIt tar — the deposit Michelle Lindlar, Yvonne Tunnat and Carl Wilson made on RADAR in 2017 (doi:10.22000/53, CC BY-SA 4.0), whose archive checksum RADAR publishes — under ADR 33. Each derives from one page, "Hello PDF-world!" in Times-Italic, with one deviation from ISO 32000-1's structure: header 7, catalog 7, page tree 9, page object 12, page resources 6, content stream 18, cross-reference table 10, trailer 19. They hold no metadata and no other text, so the screen was short. Titles are our own words, from each file's difference with the base page; the entries take from the authors' spreadsheet only each case's number, category and JHOVE 1.16.5's 2017 verdict, kept as features (ipres2017-t04-019, ipres2017-trailer, jhove-1.16.5-not-well-formed) and never copied.

Their expectations come from qpdf and pdftotext, as for the rest of the corpus; the 15 files with a space before %PDF expect the offset adjustment qpdf makes without a word. The edits often did more than their authors meant: in 45 of the 59 header and body files the table no longer leads to its objects — qpdf rebuilds it in 43 and adjusts to a leading space in the other two —, and the operator faults of the content-stream cases came with broken lengths or offsets the reader does meet. One file holds no catalog at all, and its entry says so (catalogRecoverable: false): the reader must open it and hand back none.

19 were recorded as unsupported, all until M02, which names them in its acceptance conditions; none is since M02's third slice (2026-09-28):

• Seven page trees counted differently: a root listing itself, a root listing three kids of which one page, a kid the file lacks, a root without /Kids, a root claiming /Count 9 for one page (qpdf trusts the count), a root without /Count, a missing page object.
• Four faults read without a word: a root typed /Pagez, a page typed /Font, a table entry with generation 10000, a trailer without /Size.
• Six recoveries that differ from qpdf's: an entry 60 bytes early, which the reader finds nearby where qpdf rebuilds; a trailer without >>, which the reader reads anyway where qpdf finds no trailer; and four trailers whose /Root is missing, lacks its R or its generation, or points at the content stream — qpdf keeps its sound index and gives up, the reader rebuilds the index to find the catalog, where it could look among the indexed objects first.
• Two references to an object the file lacks (T27): a catalog's /Pages and a page's /Contents. qpdf takes each as null, as the specification says; the reader rebuilt its whole index looking for the object and reported a repair on a file qpdf calls clean, until T27 was fixed on 2026-09-27. The acceptance test found them only once it walked each page's contents and resources before judging a clean file's diagnostics, which it now does. Both are supported; their finding is M02's third slice.

M02's second slice made eight of the 17 supported on 2026-09-27: the table entry with generation 10000 and the trailer without /Size of the second item, and the six recoveries of the third. Each earns its finding, and the reader, where only /Root is broken, now finds the catalog among the indexed objects without rebuilding the index, as qpdf keeps its own; their expectations say what the reader does, where they had copied qpdf's rebuild. Nine remained, the page trees and the two misnamed types, for the third slice.

M02's third slice made seven of the nine supported on 2026-09-28, the seven page trees: each earns its page tree finding, and its page count is qpdf's walk of the tree — each missing kid a page with nothing on it, a kid that loops back nothing —, where the manifest had held the root's /Count as qpdf --show-npages gives it: t02-02-002 and t02-02-005 hold none, t02-02-007 and t02-02-008 one. The two misnamed types wait for the object-shape rules.

Traps met along the way

• Printers stamp who printed. Microsoft Print to PDF writes the Windows account's display name into /Author whatever the document says; PDF24's Ghostscript copies the account name from the PostScript %%For comment into /Author and the XMP; Word's Save as PDF copies the document's Author property, which defaults to the Office user. build/build_word.ps1 handles each, and deletes any output that still names the account.
• Incremental updates remember. An /Info dictionary replaced by a later revision is still in the file, and so is the name in it.
• XFA remembers too: a LiveCycle form's configuration packet keeps the designer's file paths.
• Some URLs serve whatever is current: irs.gov/pub/irs-pdf/, the Cerfa .do links and opm.gov's forms change with each edition, while irs.gov/pub/irs-prior/ and the EU's Cellar URIs are permanent. EUR-Lex itself answers robots with a challenge page; the Cellar serves the same bytes.
• The Internet Archive's id_ URLs return the original bytes, byte-identical to what pdf.js and others recorded; without id_, the archive rewrites the file.
• pdfa.org refuses a bare user agent, and several government sites answer 403 to one.
• A registered office can be someone's home. Notices about associations give their seat, and a small association's seat is often its founder's house — one gave it as "chez M. et Mme …". An association's stated purpose can even describe a named child's illness. Each notice had to be read, not sampled.
• A copier's name survives a later save. Two Xerox scans still carry the copier as Producer after an unidentified tool re-serialized them; the OCR layer is the copier's (its own font names, no Acrobat trace), but the bytes are not the copier's alone. The Canon scan NIST published had been re-saved by Acrobat too.
• Some suites exist only as a zip. The PDF/UA Reference Suite is published as one 26.8 MB zip, served only to a named user agent, and the fetcher takes a URL, not a zip member. Byte-identical copies of its members sit at pinned commits of other projects — the veraPDF regression tests and opendataloader-pdf — so the manifest pins those, each hash checked against the zip's own member.
• Well-known samples are signed with an employee's own certificate. Adobe's widely copied signed sample carries the signer's own e-mail in the certificate subject, printed again in the visible signature. Most of EU DSS's signed files carry real people's qualified certificates, with identity numbers, personal e-mail addresses and phones in the subject, and a public body's DocuSign or Adobe Sign audit report lists every signer's e-mail and IP address. A signed file is screened certificate by certificate, not only page by page.
• Personal data hides where extraction does not look. One pdf.js file keeps two page objects outside the page tree, holding employees' names and e-mail addresses in hex glyph codes: invisible to pdftotext and to a pattern scan, found only by grafting the pages back into the tree and extracting them. In two hybrid invoices, the personal data was in the attached XML.
• A license file can be contradicted by the record. OCRmyPDF's REUSE.toml says MIT for its PDFpen file; the issue thread in which the contributor granted it says CC BY-SA 4.0. The primary record wins.
• Truncated files make referees disagree. On a Word file cut at 1 MiB, qpdf 11.9.1 counts 11 pages, PDFium 8, qpdf 12 and PDFBox 4; on a budget book cut at 1 MB, qpdf 11.9.1 finds no /Root at all. Neither entry asserts a page count.
• The reader's fixed windows. Real files found three bounds the synthetic ones had not reached. The object parser reads through an 8 KB window: a stream whose endstream falls just past it is reported truncated (T21), and an indirect object longer than the window is cut at its edge (T23) — a 10,112-byte DSS /VRI dictionary in a 25-signature file, 36 structure arrays of about 8.7 KB in a 2015 BOE law — where qpdf reads both whole. And each cross-reference section is read through a window of up to 64 KB, whatever the section's size (T24): opening a 218 KB signed web capture with three sections reads 117 KB, bounded, but proportional to the number of sections. T24 is for M23; T21 and T23 were fixed on 2026-09-26 — what an attempt through a window too small for its object notices is now dropped with it, and the file is asked for the bytes after a stream's data. docs/status.md tracks all three. The fourth pass met T24 from the other side: a section larger than its first window is read again from its start each time the window grows, so a 273 KB table costs 683 KB of reads at opening.
• Git can change a PDF's bytes. A PDF made only of ASCII, with no NUL in its first 8,000 bytes, looks like text to git, and a repository that normalizes line endings strips its carriage returns: the /Length of each stream and every offset after it go wrong. The OPF's copies of the iPRES 2017 hand-built set lost five bytes each that way — all but minimal_test.pdf, which had none to lose — which the authors' own archive on RADAR shows; the repository has since added * text=binary to its .gitattributes. Every other OPF file admitted has a NUL early on or still holds its CR LF pairs, and passed qpdf --check where its producer left it sound.
• Referees count pages differently on a damaged page tree. A kid that points at an object the file does not contain is a blank letter-size page to qpdf, poppler and PDFium, and nothing at all to pikepdf and our reader: 65 pages or 53 for one JHOVE file. pikepdf, which the first two passes used for page counts, also counts one page where qpdf, poppler and the linearization dictionary count six, on a damaged tiff2pdf scan. The fourth pass counts with qpdf --show-npages, the command the integration tests run, and records the disagreement where the reader sides with pikepdf.
• qpdf is silent about bytes before the header. A MacBinary header of 128 bytes, or a data: URI prefix of 28 left by a browser, shifts every offset, and qpdf adjusts without a word. Both files are recorded as not clean, with the offset adjustment the reader must report.
• Damage can wait for the object that holds it. qpdf checks every cross-reference entry when it opens a file; our reader, lazy by design, trusts the index until it reads an object that is not where the index says, and rebuilds then. Two JHOVE files therefore expect no rebuild at opening and the rebuild diagnostic after a full read, like PDFBOX-3947 before them.

What the milestones can take from it

• M01: none of the 239 real PDFs downloaded in the first pass crashed or hung the reader, nor any of the 82 the third pass or the 131 the fourth pass first admitted, and every vendored file opens as its entry says. Two hand-written files show it silent where qpdf reports damage; they are recorded as unsupported until M02. On the empty object of the compacted-syntax file, the reader and qpdf agree.
• M02: the vendored files carry anomalies every reader tolerates and a validator should name without calling them errors — stale linearization hint tables (some twenty files), a /Size one too large, an xref stream without its own entry, references to objects the xref lacks, font-level XMP that is not well-formed XML, an empty /Lang, a /Lang that contradicts the text, UTF-16LE text strings, two startxref lines, a /MarkInfo pointing at the page tree, /Info and XMP that disagree. docs/milestones/M02.md lists them among its acceptance conditions. The third pass adds T23, an indirect object longer than the parser's 8 KB window cut at its edge (fixed on 2026-09-26); two more files where the reader is silent while qpdf reports damage — an /Info object without endobj, names containing #00; an object stream destroyed by a zeroed block, which the reader finds only when it reaches it; and more anomalies to name — startxref on the line end before xref, an update glued to the %%EOF marker, startxref past the end of the file, a text string whose byte-order mark precedes single-byte text, UTF-16BE strings opening with a language escape, a catalog /Lang of (English), /DecodeParms [null], a /ToUnicode that is a name, an inverted MediaBox, an indirect /Rotate, an outline whose /Last is itself. The fourth pass adds T25, a /Prev 12 bytes off that drops a whole section in silence, to be fixed before M02 closes; five files the reader must stop being silent or divided about — two catalogs without /Type, a null kid in a page tree, two trees whose kids point at missing objects —; bytes before %PDF (MacBinary, a data: URI), a header naming PDF 1.8, a reference to object 0, a trailer /Size equal to the highest object number, a byte missing after the header, an image whose /Height contradicts its data, a resource pointing at an object that does not exist; and JHOVE's own verdict on 45 files, a second independent opinion to compare the validator with. And, through ADR 33, the iPRES 2017 hand-built set: 88 files each derived from one page with one deviation its authors describe, 19 of them recorded as unsupported until M02 — the one external test suite the structural profile can be measured against.
• M03, M05: signatures to keep intact through an incremental update — a GPO certification, DILA's Dictao signature, a qualified seal renewed by timestamps over two years, PAdES B-LTA, Acrobat signatures. The third pass adds each PAdES baseline level on one page (B-B, B-LT, B-LTA); DocMDP P=1 and P=2 with FieldMDP and legacy MD5 object digests; the legacy adbe.x509.rsa_sha1 and adbe.pkcs7.sha1 subfilters; plugtest seals each followed by a document timestamp; a file re-signed over seven years, ending in RSASSA-PSS; 24 signatures in 48 incremental updates; certified and signed files under AES-128; and commercial e-signatures — DocuSign (one committed, two remote), Adobe Sign, Yousign's seal written by a full rewrite. And files a signer must neither mistake nor break: an unsigned placeholder whose /ByteRange holds name tokens, a /Reason containing the word trailer, damage inside a signed byte range, a signed update that startxref cannot reach.
• M06: Acrobat 9's signed portfolio — a /Collection with a schema and folders, a Flash navigator, five embedded PDFs among them a certified XFA form and 3D models — is the first portfolio a case-file assembly will meet; beside it, the Massachusetts dashboard's chain of ten revisions.
• M15: extraction will meet text drawn as images, OCR layers — Acrobat's and two copiers' own, with horizontal scaling up to 2000 % —, vertical Japanese without ToUnicode, Arabic on which xpdf and poppler disagree — xpdf reads "الضرائب", poppler "الرضائب" —, the PDF24 file where ToUnicode says E and no glyph is drawn, a compacted-syntax file from which xpdf 4.00 extracts nothing while poppler and PDFium do, indirect references inside a content stream, and inline images whose keys contradict each other. The third pass adds tagged scans whose OCR text lies under the page image — ABBYY FineReader 8's (committed), and an 82-page PDF/UA-1 remediation —, so extraction must prefer the structure over what a renderer shows; OCR layers from OmniPage 19 and a Canon scanner; Hebrew in TrueType and CID subsets beside the Census guide's Type 1; a ToUnicode that turns the minus into a soft hyphen, and one whose values are byte-swapped; EBCDIC codes in bitmap Type 3 fonts, and bitmap Type 3 fonts with no ToUnicode at all; word spaces present only as TJ kerning; UTF-8 strings in PDF 2.0; a real seven-column table; and a poster whose text is outlines, which must yield no text rather than noise.
• M16: the fourth pass brings the security handlers side by side on one document: Acrobat 9's AES-128 (R4) under three permission sets and under an open password nobody published, OpenOffice's 128-bit RC4 (R3) with and without the open password password, and Quartz's 40-bit RC4 (R2) behind it. Before them, XFA forms from Designer 6.4, 6.5 and ES 9, one of them readable today; AcroForms driven by JavaScript; a calculation order; widgets without appearance streams. The third pass adds dynamic XFA, in two Canadian forms whose only PDF page is a placeholder; an XFA form filled with fictitious values and certified, inside the portfolio; OPM's OF-306, with date JavaScript and two unsigned signature fields; radio buttons whose export value holds a NUL; read-only fields under a Yousign seal.
• M20: veraPDF 1.30.2 upholds thirteen of the fifteen PDF/A claims among the new files, and PDFlib's PDF/UA-1 claim. It rejects two: the 2015 consolidated text (rules 6.4-3 and 6.2.3.3-1) and ImageMagick's claim (6.7.3-8, 6.1.8-1, 6.7.3-1, 6.2.3.3-1). It passes the 2019 notice despite its malformed font XMP, which is worth an opinion of our own. With the third pass, the corpus holds Factur-X in every profile — MINIMUM, BASIC WL, BASIC, EN 16931 and EXTENDED, the last three among the committed files — and ZUGFeRD 2.2's XRECHNUNG profile, beside ZUGFeRD 1.0 BASIC and EXTENDED with a supplementary attachment, an Order-X purchase order, and a factur-x.xml holding UBL rather than CII. veraPDF rejects the committed factur-x Python file's PDF/A-3b claim on five rules, and the PDF/A-1a claims of the Slovak seal and the re-signed BOE decree; it upholds PDF/A-4f in PDF 2.0, Symtrax's PDF/A-3a, the two gazettes' PDF/A-1a, Oracle Outside In's PDF/A-1a despite a broken loca, and three more PDF/UA-1 claims from the reference suite. A textbook chapter carries PDF/A and PDF/E output intents with no claim at all: a case for D16. The fourth pass adds OpenOffice.org 3.2's PDF/A-1a, twice upheld; Acrobat 11's Image Conversion PDF/A-1b, upheld on the intact page and — worth an opinion of our own — on both copies damaged on purpose; a PDF/A-1b claim on a 464-page VA guide, rejected on seven rules; and two claims veraPDF cannot parse at all, one on a file missing a byte after its header.
• M23: the W11 references above, all three fetched every night — 9,302 pages, one 63 MB page, 125 JPEG 2000 pages in 147 MB — and beside them a 35,000-pixel square CCITT image that decodes to 153 MB from 10.5 KB, 65,542 objects in one object stream, 48 incremental updates, an 82-page tagged scan of 10.6 MB; and T24, which ties the bytes read at opening to the number of cross-reference sections.

What the public sources could not provide in a file we may commit — a real invoice or statement from a supplier or bank (six are now remote), a copier file untouched since the copier wrote it (one, from Konica, is now remote), Hebrew (two remote), Korean — is what docs/corpus-contributions.md asks contributors for. The commercial e-signature it used to ask for is now committed: the GSA's contract modification completed in DocuSign.