The Document Accessibility Census — University of Texas System (Pilot)
Internal pilot. Aggregate, machine-verifiable findings only (Tier 0 + Tier 1). No Tier 2 judgment work.
We have a commercial interest in these findings: we sell document remediation. The methodology, sampling frame and clause-to-class mapping are published so a hostile reader can check the numbers.
Headline
- Almost nothing passes. Only 2 of 1,500 sampled documents (0.1%) clear PDF/UA-1 machine validation. Of the 1,386 veraPDF could actually check (the rest are encrypted, image-only or too large), still only 0.1% pass. 99.9% fail even the automated floor — and machine-conformance is necessary, not sufficient, so the truly-accessible share is lower still.
- But most of the failure is automatable. Among failing documents, 33% fail only on structural defects (Class A) a machine can repair with no author involvement. The problem is near-universal; the fix, mostly, is not human-authored.
- Counting everything a machine can finish alone, 33% of documents need no author’s time (Class A + D), and a further 59% are near misses — Class A but for one or two authored items (e.g. a single alt text).
- 8 of 13 UT System institutions had samples large enough for the weighted headline (≥30 documents); 19,854 PDFs discovered across all 13. The other 5 are bot-blocked or have thin public sitemaps — see coverage.
- Institution-weighted across the well-covered set, the no-author-fix share is 32% and conformance 0.1%.
What a PASS means (and what we did not test)
These are machine-checkable PDF/UA-1 results only. A passing clause means veraPDF could confirm a required structure or attribute is present — not that it is meaningful. We did not test alt-text quality, reading-order sanity, heading logic, table semantics, link-text usefulness, colour contrast (a WCAG matter), or real screen-reader behaviour. A document with zero failures is machine-conformant — necessary but not sufficient for accessibility. We never call a document “accessible.” Read the numbers as a lower bound on the problem and a precise map of the automatable fixes.
The hypothesis under test
Half of the inaccessible documents in American higher education can be fixed without anyone writing a word.
Among failing documents in this sample, 33% are structural-only (Class A). The pilot number differs from the 50% prior — reported as found.
Class distribution (document-pooled)
| Class | Meaning | Documents | Share |
|---|---|---|---|
| D | Machine-conformant (not verified accessible) | 2 | 0% |
| A | Structural only (no author) | 483 | 33% |
| A+B | Structural + short author task (near miss) | 890 | 60% |
| C | Source/rebuild (image-only) | 80 | 5% |
| oversized | Too large to validate | 31 | 2% |
Tier-0 quick wins (trivially fixable, near-universal)
- 35% have no tag tree at all (untagged).
- 38% have no document language set (
/Lang) — a one-line fix. - 67% do not set
DisplayDocTitle. - 5% are image-only scans with no text layer (need OCR — Class C).
Estimated exposure at prevailing rates
Failing documents in the sample span ~17,611 pages. At market rates of $5–8/page this is $88,055–$140,888 of remediation for the sample alone. Of that, $27,502 (the structural-only pages) is avoidable through automated repair. Presented as prevailing-rate context, not a quote.
Failure taxonomy (discovery)
| State | Count |
|---|---|
| http-403 | 244 |
| http-404 | 55 |
| not-pdf | 44 |
| fetch | 29 |
| oversized | 24 |
Coverage (institutions excluded from the headline)
| Institution | Discovered | Analyzed | Why excluded |
|---|---|---|---|
| UTHealth Houston | 1,258 | 6 | downloads blocked (HTTP 403 bot-detection) |
| University of Texas at Dallas | 4 | 4 | thin public sitemap (few PDFs discoverable) |
| University of Texas at Arlington | 3 | 2 | thin public sitemap (few PDFs discoverable) |
| University of Texas at Austin | 1 | 1 | thin public sitemap (few PDFs discoverable) |
| University of Texas Medical Branch | 28 | 1 | thin public sitemap (few PDFs discoverable) |
These are reported for honesty, not dropped silently. Bot-detection (HTTP 403) and long crawl-delays are themselves accessibility-adjacent signals worth noting.
Method & limitations
- Sample frame: PDF URLs discovered by a polite crawl (robots.txt honoured, per-host rate limits) of each institution’s public web, drawn from both fetched PDFs and the crawl frontier.
- Coverage is uneven: institutions with thin public sitemaps or aggressive bot-detection (and those enforcing long crawl-delays) yield smaller samples. Discovered vs sampled counts are reported per institution so this is visible.
- Tier 0 = structural probe (pikepdf). Tier 1 = veraPDF 1.30 PDF/UA-1. No Tier 2 (alt-text quality, reading order) at this scale.
- Class A means machine-fixable without new content. It does not mean verified accessible: reading order, heading logic and existing alt text still need human review after repair.
- Near-miss counts distinct content clauses, not instances. A single failed ‘figure needs alt text’ clause can stand for one image or a hundred. Near-miss therefore measures how few kinds of authored input a document needs, not the total authoring effort. Instance-level counts (veraPDF failedChecks) are a planned refinement.
- Source documents were never retained: fetched, probed, validated, discarded. Only findings, hashes and metadata persist.