We Ran veraPDF on 774 .gov & .edu PDFs (3.89M Failed Checks): 為何「已標記」PDF 仍無法通過 ISO 14289-1,以及如何自動修復
We Ran veraPDF on 774 .gov & .edu PDFs (3.89M Failed Checks): Why 'Tagged' PDFs Fail ISO 14289-1 & Auto-Fix
veraPDF 對 774 份來自 .gov 和 .edu 網站的 PDF 進行 PDF/UA-1 驗證,771 份(99.61%)未通過,累計有 3,891,504 項檢查失敗;595 份已有結構標籤的 PDF 中,592 份仍未通過。
When engineering teams audit websites for DOJ 28 CFR Part 35 (ADA Title II) or the European Accessibility Act (EN 301 549), running axe-core against HTML templates is straightforward. Auditing—and repairing—legacy PDF repositories is a completely different binary puzzle.
To quantify what actually breaks inside real-world public documents, we ran veraPDF (ISO 14289-1:2014 / PDF/UA-1 profile) across two reproducible datasets:
-
Public-Sector & Higher-Ed Crawl Corpus: 774 unique PDFs (19,022 pages) across 124 domains (
28.gov,89.edu,7.org/public portals). -
In-the-Wild Internet PDF Torture Benchmark: 31 adversarial PDFs (228 pages) across municipal budgets, university syllabi, dense federal/campus AcroForms (
335form fields), scanned clerk ordinances, and EU hardware Declarations of Conformity.
Here is the raw engineering breakdown—and why honest tooling must separate machine_checks_pass: true from human_review: PENDING.
1. The Empirical Baseline: 99.61% Failure Rate (and the "Tagged PDF" Myth)
Out of 774 unique SHA-256 PDFs crawled from public .gov and .edu websites:
-
771 / 774 (
99.61%) failedveraPDFPDF/UA-1 validation (5,453failed rules;3,891,504failed checks). - 225 / 225 (
100.0%).govPDFs failed. -
524 / 527 (
99.43%).eduPDFs failed (all 3 passing documents originated from a single campus,csub.edu).
The biggest surprise wasn't untagged scans (179 files). It was that 595 of the 774 PDFs (76.87%) already had /StructTreeRoot tags—and 592 of those 595 (99.50%) still failed.
2. The Low-Level Binary Traps Inside ISO 14289-1
Why do documents exported straight from Word, PowerPoint, InDesign, or Acrobat still trigger thousands of veraPDF clause checks?
| Clause | Affected PDFs (n=774) |
What Breaks at the PDF Object Level (ISO 32000-1 / ISO 14289-1) |
|---|---|---|
Clause 5 |
650 (84.0%) |
Catalog /Metadata XMP stream omits pdfuaid:part=1. |
Clause 7.1 |
622 (80.4%) |
Vector paths/headers sit in page content streams outside both MCIDs and /Artifact BMC ... EMC, or /ViewerPreferences omits /DisplayDocTitle true. |
Clause 7.18.1 / 7.18.5 |
358 (46.3%) / 320 (41.3%) |
/Subtype /Link annotations in page /Annots lack /Contents strings, forcing screen readers to spell raw URLs. |
Clause 7.2 |
337 (43.5%) |
Missing BCP-47 /Lang, list items (<LI>) missing <LBody>, or ragged table ColSpan counts (e.g., a 33-page municipal budget where 522 table cells had mismatched spans). |
Clause 7.21.4.1 / 7.21.4.2 |
327 (42.2%) / 314 (40.6%) |
Unembedded fonts, or subsetted CIDFonts whose /CIDSet stream bitmap disagrees with the embedded font program's glyph table. |
Clause 7.18.4 |
129 (80.6% of forms) |
AcroForm /Subtype /Widget annotations exist on the page without <Form> OBJR structure references in /StructTreeRoot. |
Clause 7.21.7 |
84 (10.9%) |
Missing or invalid /ToUnicode CMap streams (including illegal U+FEFF/U+FFFE entries). |
3. What Deterministic Local Python (pikepdf + RapidOCR + veraPDF) Can Repair
In our 31-PDF In-the-Wild Torture Benchmark (228 pages), running a 100% local workstation pipeline (zero cloud API uploads) took total veraPDF failures from 212 failed rules (57,205 checks) down to 0 failed rules (0 checks) across 31 / 31 PDFs in 877.96s (3.85s/page), while verifying pixel-identical rendering (max_page_mean_pixel_diff == 0.0) on every page:
- Rebuilding missing
/ToUnicodeCMaps and pruning broken subset/CIDSetstreams without mutating rendered glyph metrics. - Wrapping untagged decorative background vectors in
/Artifactmarked-content sequences while preserving visual z-order. - Wiring orphaned AcroForm
Widgetannotations into<Form>structure nodes and setting/Tabs /Sacross annotated pages. - Running local
RapidOCRon scanned ordinances (Segment D) to inject an invisible searchable text layer + tag hierarchy without altering the archival raster image.
4. Why machine_checks_pass: true Still Leaves human_review: PENDING
A PDF passing 106 / 106 veraPDF rules is syntactically valid PDF/UA-1, not automatically human-accessible. Our workflow engine explicitly enforces:
{
"machine_checks_pass": true,
"human_review": "PENDING",
"incomplete_delivery_blocked": true
}
Automated checkers cannot tell whether:
- Two-column reading order flows down Column A before Column B instead of jumping across rows.
- A
<Figure>/Altattribute describes the actual trend in an EPA water-quality chart rather than repeating a filename. - Complex financial tables associate multi-level column headers semantically when navigated in NVDA or PAC 2026.
Use local automation to eliminate the 85% binary PDF surgery in seconds—and save your human accessibility team's hours for real screen-reader verification.
Full Benchmark Data & Open-Source / Local Software Tools
- Full 2026 Benchmark Report (774 PDFs + 31 Torture-Test Dataset): hadesloc.github.io/pdfua-partner-desk/benchmark-report-2026/
-
Official Mozilla-Approved Firefox Add-on (
v1.3.0): PDF/UA & ADA Page Scanner on Mozilla Add-ons -
Open-Source MIT Validator CLI & GitHub Action (
npx pdfua-ada-scanner): github.com/hadesloc/pdfua-ada-checker -
Windows Desktop Studio
v1.6.1(100% Local Pure Software — Free 5-Page Trial &EARLY5050%-Off Solo$9.50/ Agency$32.50Licenses): hadesloc.github.io/pdfua-partner-desk/ (run Auto-Fix, RapidOCR, AcroForms & ZIP export free on<= 5pages, or unlock unlimited6+-page local batches with an instant license key)
Technical Software Disclaimer: All tools linked above are standalone pure-software utilities and open-source validators intended for local developer/operator engineering workflows—not legal advice, regulatory compliance certification, or manual human remediation services. Automated ISO 14289-1 (
veraPDF) validation does not replace human screen-reader (NVDA/JAWS) andPAC 2026semantic verification prior to public distribution.
來源:DEV Community · dev.to