Kill Log

TeardownsTeardown 1 of 3

A validator passed 20 of 20 PDFs. Only 8 were usable.

The idea: public bodies must make documents accessible, so an automated pipeline could fix PDFs at a price under human vendors. We built the pipeline from free tools and ran it on real government and school PDFs.

Pilot files dated 2026-11-19, Asia/Seoul time.

Verdict

Killed

Strict pass rate 40% of documents (47% of pages) against a kill line of 50%. Estimated human time 12.5 minutes per page against a kill line of 8.

40%
Strict pass, 8 of 20 documents
Kill line 50%
12.5
Estimated human minutes per page
Kill line 8
0 of 7
Forms that passed
27 pages
20 of 20
Passed the validator alone
veraPDF

Why this idea looked good

The US Department of Justice sets web and mobile accessibility deadlines for state and local governments. On ada.gov the compliance dates are April 26, 2027 for entities with a population of 50,000 or more, and April 26, 2028 for smaller ones and special districts. The same page exempts older PDFs that were posted before the deadline, unless they are used to apply for or use a service. That shrinks the market to new documents and forms.

Visible prices from vendors: ClearLayer from $6 per page, Allyant at $5 to $8 per page (Syracuse University page), and an automated vendor, CASO Comply, from $0.30 per page.

The pipeline

PyMuPDF extracts structure, RapidOCR reads scanned pages, python-docx builds a structured document, LibreOffice exports a tagged PDF, pikepdf fixes metadata and annotations, and veraPDF 1.28.2 checks the PDF/UA-1 profile. All free. PAC, the other common checker, runs only on Windows and was not used.

Round 1: six samples, 15 pages

  • veraPDF passed 5 of 6 files. Usable passes were 4 of 6 files, 7 of 15 pages (47%).
  • A fillable form passed the validator after losing all 7 of its form fields. The validator was satisfied because the fields were gone.
  • Our first table extraction turned 4,210,000 into 4210000 and -16.7% into 167%. veraPDF did not flag either. We fixed the extraction and added a text comparison.
  • A real IRS W-9 was already tagged. Rebuilding it lost 23 fields, so we switched to in-place repair, which still failed 3 checks (heading level skip and font embedding).

Round 2: 20 real public PDFs, 110 pages

We took the first 8 pages of each document. Sources: IRS forms (W-4, 8822, 4506-T, 2848, 4868) and Publication 15-T, a USCIS I-9, the federal voter registration form, the Federal Register copy of the ADA rule, three NIST special publications, two enrolled bills from govinfo, four city, county and school board documents, and two scanned USGS reports from 1885 and 1974. A GAO site blocked our download, so those were left out and not worked around.

We defined pass strictly: veraPDF passes, and word and number recall are each at least 98%, and every original input field is restored (signature and button fields excluded), and reading order similarity is at least 0.90.

GroupDocsStrict passPagesEst. human min/page
Forms702736.2
Simple text and scans53234.9
Long multi-column reports42324.8
Tables and meeting packets43284.9
All208 (40%)11012.5
Figure 1Strict pass by document, grouped by type
Forms
0 of 7
Simple text and scans
3 of 5
Long multi-column reports
2 of 4
Tables and meeting packets
3 of 4

passed every strict check failed at least one

Source: pdf-accessibility-pilot-round2.csv, column pass_strict. 20 documents.

Figure 2Estimated human minutes per page, against the 8-minute kill line
Forms
8 min
36.2
Simple text and scans
4.9
Long multi-column reports
4.8
Tables and meeting packets
4.9
All 20 documents
12.5

Forms run at 36.2 minutes per page, 4.5 times the line. Every other group sits under it.

Source: pilot-pdf-a11y round 2 scoring (est_human_min_per_page, weighted by pages). Minutes come from a formula, not a timed person.

  • veraPDF alone: 20 of 20 pass after processing. Before processing, all 20 originals failed, with 193 to 7,927 failed checks each.
  • Form field restoration: W-4 31 of 48, I-9 99 of 130, Form 2848 20 of 92. Restored fields were attached to nearby instruction text, not the real labels, and forms grew 2 to 13 times in page count.
  • Machine time: 70.6 seconds total, about 0.64 seconds per page including OCR.
  • Without forms: 8 of 13 documents pass (62%) at about 4.9 minutes per page.
Figure 3Form fields restored on three forms
Form W-4
31 of 48
Form I-9
99 of 130
Form 2848
20 of 92

Bar length is the share of original fields restored.

Source: round 2 pipeline output. Fields were counted in the original PDF and in the rebuilt one.

Where the market comparison breaks

At $30 per hour (an assumption) 4.9 minutes per page costs about $2.45, which sits between the $0.30 automated vendor and the $5 to $8 human vendors with no clear reason for a buyer to pick us. Forms cost about $18 per page to fix by this estimate, against ClearLayer's listed $12 per page for complex tables and forms. The business turned into hourly labour.

Limits of this test

Only the first 8 pages of long documents were processed. Human minutes are computed from a formula (2 minutes of review per page, 0.75 per image, 3 per table, 4 per scanned page, plus extra for failed content, reading order and forms), not timed with a person. No screen reader or PAC check was run. Reading order similarity is a rough word-sequence comparison. Images got placeholder text, so 57 images in one budget packet still need real descriptions. OCR accuracy on the scans could not be scored without a ground-truth text. The Federal Register copy of the rule that we processed lists April 24, 2026 and April 26, 2027 as dates, which differ from the dates on ada.gov; we did not trace why.

Files

Per-document results (CSV).