audit.icjia.app · built for Title II ADA · WCAG · Illinois IITAA
Built to be checked.
See for yourself.
Every claim on this page can be verified without taking anyone's word for it — including ours.
Every number below was pulled live from the running system onOCT 6, 2026
12,659
check-ups run for internal and external agencies — the quick per-file check-up that has caught on beyond ICJIA: easier than SiteImprove, clearer than PAC. 146 in the last 30 days
4,118
self-checks the software runs on itself before anything goes live — all must pass, or nothing ships
How it’s tested →881
saved, dated, public changes over 31 weeks — every step on the record
what it does
One job: is this document readable by everyone?
Many people who are blind or can't see well use a screen reader — software that reads what's on the screen out loud. A document is accessible when a screen reader can read it correctly. Upload a PDF, Word, PowerPoint, or Excel file and in about a minute this tool reports everything that would trip a screen reader up — and exactly how to fix each one. Title II of the ADA and the Illinois Information Technology Accessibility Act (IITAA) require this of every public body — and both point to the same rulebook: WCAG.
the law · three names, one idea
Title II. IITAA. WCAG.
Government information must work for everyone. Two laws say so; one rulebook defines "works."
Federal law
Title II of the ADA
The Department of Justice rule for state and local government. It names WCAG 2.1 Level AA as the standard — and its compliance dates are April 26, 2027 for entities serving 50,000 people or more and April 26, 2028 for smaller ones and special districts. This is not optional, and it is close.
Illinois law
IITAA
The Illinois Information Technology Accessibility Act — our state's own accessibility law, older than the federal rule, also built on WCAG 2.1 AA. It applies to every Illinois public body, including this one.
The rulebook
WCAG
The Web Content Accessibility Guidelines — the international rulebook both laws point to. This checker tests WCAG 2.1 AA — the exact version both laws name. WCAG 2.2 exists and adds to it, but nothing 2.2 added can move your grade here.
Every finding on every report names the WCAG rule behind it, so a reviewer can trace any grade straight back to the law's own standard.
the tests
What every check-up looks at.
Up to nine graded areas per document, matched against the PDF industry's official 31-point checklist — the same one professional human reviewers work from. Every finding cites the WCAG rule behind it — plus an honest list of what only a person can judge.
Real, selectable textwords, not pictures of words
Title & languagewhat a screen reader announces first
True headingslevels, order — and whether they say anything
Image descriptionsevery picture a reader can't see
Table structureheader cells, and which way they point
Reading orderis it read aloud in the order a person would read it?
Links & navigationbookmarks, link text that means something
Formslabeled fields, keyboard order
Word · PowerPoint · Exceltheir own check suites, same scoring rules
Every report also states plainly what automation cannot check — about 60–70% of accessibility is human judgment, and the report says so on every grade.
The 31-point checklist has a name: the Matterhorn Protocol. It is published by the PDF Association — the same industry body that builds the veraPDF referee — and it is the test model professional evaluators and tools like PAC work through. Here is how the pieces fit: Title II and IITAA name the law’s standard, WCAG — rules for what must be true of any content. Matterhorn translates those rules into 31 PDF-specific, testable checkpoints — where inside a PDF to look. So the law, our checker, veraPDF, and a human evaluator’s checklist are all reading from connected pages, and every report maps its findings onto all 31:
01Real content tagged02Role Mapping03Flicker04Color and Contrast05Sound06Metadata07Dictionary08OCR-generated content09Appropriate Tags10Character Mappings11Declared Natural Language12Stretchable Characters13Graphics14Headings15Tables16Lists17Mathematical Expressions18Page Headers and Footers19Notes and References20Optional Content21Embedded Files22Article Threads23Digital Signatures24Non-Interactive Forms25XFA26Security27Navigation28Annotations29Actions30XObjects31Fonts
the referee · PDF/UA & veraPDF
It doesn't grade its own homework.
PDF/UA is the international rulebook for PDFs a screen reader can use — an official standard, like a building code for documents. veraPDF is a separate referee program that checks files against that rulebook, built by the PDF industry's own association — not by us, and used by national libraries worldwide. It runs beside our checker on every single report, as an independent second opinion.
Your document
Checked, never stored. Deleted the moment the report is built.
Built in-house
Our checker
Structure, headings, image descriptions, tables, forms, reading order.
PDF Association
veraPDF
The industry's own PDF/UA validator. We did not write one line of it.
Our verdictThe 0–100 score, the letter grade, and the fix-it list
Their verdictveraPDF’s independent pass / fail against the PDF/UA rulebook
Two independent referees on every report. If ours were wrong, theirs would say so — in public, on every audit. And one precision the experts will appreciate: Title II and IITAA require WCAG, not a PDF/UA badge — the two overlap heavily by design, and this report checks both, telling you plainly which rulebook each finding belongs to.
the trap test · modeled on canva, indesign, and word
We built 222 documents designed to fool it.
Real documents can't prove a checker is right — nobody knows their ground truth. So we built trap PDFs where the correct answer is known in advance, including the exact tricks Canva, InDesign, and lazy shortcuts produce. Anyone can re-run the whole battery; the traps are rebuilt from scratch every time.
synthetic-02Perfect title & language — secretly one big photoCaught · scored 0
synthetic-03Image "description" of nothing but blank spacesFound a real bug
synthetic-07Whole paragraphs disguised as headingsCaught
synthetic-13Big bold text instead of real headingsCaught
synthetic-14InDesign's custom naming soup, one typo deepRead correctly
synthetic-15A page built the way Canva sometimes builds them — unreadable insideSaid "can't judge" instead of guessing
synthetic-16A form with no labels on any fieldCaught
synthetic-12Hostile computer code hidden in the textStayed harmless
synthetic-05Internal structure that loops foreverNo crash, no hang
…and 213 more — modeled on what Canva, InDesign, and Word actually export (flat untagged posters, decorative-shape swarms, naming soup, picture-of-words titles, “Print to PDF”), plus native Word, PowerPoint, and Excel traps — unmarked header rows, alt panels never opened, slides with no titles, “Sheet1” — and, for each flawed file, a matching corrected copy (the same document with the mistake fixed) that the checker must NOT flag. All 222 held, and for every such pair the broken version may never outscore the corrected one.
27 real bugs in this checker, found and fixed — every one on this page, because that is exactly what all this machinery is for.
Each one is now a trap document that re-proves the fix on every run, wearing a FOUND A REAL BUG chip in the inventory below. In the spirit of the rest of this page: here is every one, who found it, and what it wrongly cost.
- Blank spaces counted as a description. An image whose “description” was nothing but spaces passed the alt-text check. Found by the battery's own first full run; fixed the same day, and no real document in the test set had the defect, so no score changed. synthetic-03
- The third legal value of a header direction, rejected. The PDF standard lets a corner header label its row and its column (“Both”); this checker accepted only Row and Column, so the most carefully marked-up tables read as broken — a correct Illinois DoIT reference file was graded 89/B instead of 100/A. The agency was right, the checker was wrong, and that file's exact score is now pinned in the public ledger. synthetic-116
- Values hidden behind references, unread. Word sometimes stores a header's direction or a cell's span as a reference to another object instead of writing it in place. The checker read only in-place values and invented two false failures on a genuinely accessible syllabus (85 → 100 after the fix). Found by real files submitted as “100% accessible” — and they were. synthetic-117
- A language tag present — and wrong — sailed past everyone. English text declaring itself French means a screen reader pronounces the whole document with French rules. No tool flagged it, this one included; the silence was the bug. The check that closed it carries four guards against false accusations, and the trap holds both directions: the mislabeled file is caught, its correctly-labeled twin is not. synthetic-118
- The class-map route, unread — caught before any file arrived. The standard's third way to attach header directions (by named class) was invisible to the parser. Found by this project's own encoding gate, which rewrites one document every legal way and demands identical grades — the first bug caught preemptively instead of by an agency's file. synthetic-120
- One-row tables misread as crosstabs. A single-row table puts its only header in the first row and the first column at once, and the exactly-one-axis logic called that a two-axis grid and docked it. The trap was written first, watched fail at 89/B, and forced the fix — the battery catching the checker, by design. synthetic-124
- A header row marked the way Microsoft says, called unmarked. Microsoft tells Word authors to mark a table’s header row with the Header Row box on the Table Design tab, and Microsoft’s own checker accepts it; this checker accepted only an older setting, so a correctly built agency meeting agenda was graded 69/D. The agenda was right and the checker was wrong — it now grades 100/A. Tracing it also showed the same unmarked table costing a D in Word, a C in a PDF and a B in Excel; it now costs the same in all four formats. synthetic-157
- “No shading” read as shading. Text pasted from a web page carries a marker that means no shading, and the checker took that marker as the styling of a data table — so a borderless grid used only for layout could be graded as a data table missing its header row. Found while tracing the agenda bug above; no file in the test set was affected. synthetic-162
- A memo’s title, graded as missing sections. A Word document whose only heading-like line was its bold title — no Heading styles anywhere — lost 70 points and graded D, while the same memo saved as a PDF graded A: the PDF check had already learned that one such line is a title and two or more are sections. Word and PowerPoint now apply the same rule, and an agency quick-reference guide in the test set stopped being accused of hiding sections it does not have. Found by this project’s own comparison of the same defect across all four formats. synthetic-163
- PowerPoint’s default title, checked nowhere. A deck still titled “PowerPoint Presentation” — PowerPoint’s own default, inherited by every deck made from a template that carries it, including an agency template in the test set — names the program, not the presentation: a WCAG 2.4.2 failure (F25). No format caught it. The PDF check’s list of tool defaults lacked it, and Word, PowerPoint and Excel titles were never checked at all. Every format now applies the same title check. synthetic-167
- A PowerPoint layout grid, graded as a data table. Authors line up an agenda in a table stripped of every style, border and fill. Word has never counted a bare grid like that as a data table; PowerPoint did, so the same agenda lost its header-row points and capped at C as a deck. PowerPoint now applies Word’s rule. Found by this project’s own comparison of the same defect across all four formats; no deck in the test set was affected. synthetic-170
- A damaged properties part, reported as a missing title. When the part of a PowerPoint or Excel file that holds its title could not be read, the checker reported “no title” — a WCAG 2.4.2 failure about a title it never saw — while Word had long said “could not be read” and left that half unscored. All three now say so. Found by the same comparison; no file in the test set was affected. synthetic-175
- The same share of described images, a letter apart. A report with 16 of its 23 images described scored 69 as a PDF (a C ceiling) and 70 as a Word file (a B ceiling): PDF rounded the share down, while Word, PowerPoint and Excel rounded it to the nearest point and then capped any failing share at 85. Every format now scores alt text and link names by one rule. Found by the same comparison; two real decks in the test set moved a few points within their band, and no grade changed. synthetic-178
- Untagged section headings, reported as “no issues found”. Two agency annual reports in the test set tag only three of their headings — seventy-odd section titles are tagged as plain paragraphs — and their heading score read 100, “No issues found”: the PDF check for headings that are only visual ran only on documents with no heading tags at all. Word has always counted the same lines. PDF now does too, with guards measured against the test set first (cover pages, letterheads, pull quotes and captions all look like headings and are not); four real reports lose a score they never earned, and all four stay D for other problems. Found by this project’s own comparison of the same defect across all four formats. synthetic-179
- A language set the way this report advises, not recognised. A Word file with no document-wide language setting whose every word is marked English — exactly what this report’s own advice (Set Proofing Language on the selected text) produces — was still accused of declaring no language, while PowerPoint read the same marks. Both now take the language marked on most of the text. Found while checking that the advice actually works; no file in the test set was affected. synthetic-184
- Word tables judged by their markup, not by what they draw. Whether a Word table counts as a data table (scored, needs a header row) or a grid used only for layout (not scored) was decided by which elements the table carried. So a grid whose borders were all explicitly switched off — how Google Docs, LibreOffice and pasted web content write an invisible table — or one carrying a table style that draws nothing was accused of a missing header row, while a real data table drawn with borders on its cells instead of the table was never checked at all. The checker now asks what the table actually draws. Found by the first table traps written before any fix; no file in the test set was affected. synthetic-186
- Line breaks in a description, kept as code. PowerPoint stores a line break inside a picture’s description as a short code (

) that every program reading the file is required to turn back into a line break. The checker kept the code as text, so descriptions carried stray codes, and a description made of nothing but line breaks passed as a description. Found by the new Office encoding gate, which rewrites one Word document, one deck and one workbook every legal way and demands identical grades; no file in the test set changed score. synthetic-199 - A file saved as UTF-16, refused. The Office file format lets each part of a file be stored as UTF-8 or UTF-16 text, and Word opens both. The checker read every part as UTF-8, so a perfectly good UTF-16 file was turned away as “not a supported document”. Found by the new Office encoding gate, which rewrites one Word document, one deck and one workbook every legal way and demands identical grades; no file in the test set changed score. synthetic-200
- “Bold: off” read as bold, in Word. A Word file can say outright that text is not bold, and the python-docx library writes it that way whenever a script switches bold off. The checker saw the bold setting and never read its value, so a short line of plain 14-point text was reported as a heading typed by hand, costing points on a document with nothing wrong. Found by the new Office encoding gate, which rewrites one Word document, one deck and one workbook every legal way and demands identical grades; no file in the test set changed score. synthetic-201
- “Bold: off” read as bold, in Excel — a failure missed. The same misreading in a workbook made 14-point grey text count as large text, which needs less contrast, so text too faint to read passed the contrast check. Missing a real failure is the more damaging of the two errors. Found by the new Office encoding gate, which rewrites one Word document, one deck and one workbook every legal way and demands identical grades; no file in the test set changed score. synthetic-202
- “True” and “false” not understood. PowerPoint, and the drawing format inside all three Office formats, let a setting be written as 1 or “true”, 0 or “false”. The checker knew only the digits, so a slide hidden with “false” was judged for its missing title, and a picture marked decorative with “true” counted as an undescribed image, in Word, PowerPoint and Excel alike. Found by the new Office encoding gate, which rewrites one Word document, one deck and one workbook every legal way and demands identical grades; no file in the test set changed score. synthetic-203
- The wrong slide named. A deck lists its slides in the order they are shown, and that order need not match the order the slides are stored in. When the list was written in a different but equally valid way, the checker silently fell back to the storage order, and its report named the wrong slide. Found by the new Office encoding gate, which rewrites one Word document, one deck and one workbook every legal way and demands identical grades; no file in the test set changed score. synthetic-204
- Cells found only by a written address. Each cell in a workbook may state its own address (“B7”) or leave it out, in which case its position gives it. The checker read only written addresses, so in a workbook that leaves them out it never found a link’s text and never checked the link. Found by the new Office encoding gate, which rewrites one Word document, one deck and one workbook every legal way and demands identical grades; no file in the test set changed score. synthetic-205
- Bullets a slide’s layout switched off, never read. PowerPoint decides whether a line gets a bullet through a chain — the line itself, its text box, the slide’s layout, then the master — and the checker read only the master. So lines whose layout turns bullets off were counted as real list items: bullets typed by hand with a “•” character passed as a real list, and a title slide’s subtitle or a slide number counted as list items too. Found by building test decks the way PowerPoint’s and python-pptx’s templates write them, then confirmed on a real agency deck, which drops from 100/A to 89/B for four hand-typed bullets. synthetic-206
- A shaded band read as its plain colour. PowerPoint can shade a theme colour lighter or darker, and the checker read the shading off. White titles on a band shaded 25% darker than the theme’s blue were judged against the plain, lighter blue and failed at 2.96:1 — six failures on a real survey-results deck in the test set that were not there (the shaded band gives about 4.85:1). A shaded colour is now left unassessed rather than guessed at, the rule the checker already applied to slide backgrounds, and that deck’s contrast score rises from 50 to 100. Found while teaching the checker to follow text colours through PowerPoint’s layouts and masters. synthetic-219
- Link text judged in a colour it is not drawn in. PowerPoint draws link text in the theme’s link colour, whatever colour the text itself is set to, unless the link carries a newer mark asking for the text’s colour. The checker judged links by their text colour, so a link set to white on a navy slide passed at 11.6:1 while what is actually drawn — blue on navy — is 1.97:1: a real failure, missed. Links are now judged in the colour they are drawn in, which LibreOffice’s rendering of two real decks in the test set confirms. Found the same way; no file in the test set was affected. synthetic-220
- Highlighted text judged against the slide. A highlight paints a colour behind the text, so that colour is the text’s background. The checker ignored highlights: black text on a yellow highlight, on a navy slide, was judged black on navy and failed at 1.81:1 — a failure that is not there (19.6:1 against the yellow). Highlighted text is now judged against its highlight. Found the same way; no file in the test set was affected. synthetic-221
And the battery is not just spelling tests. It re-proves, on every run, the hard cases real documents taught this checker the hard way: Canva's phantom leftovers that once cost a clean report a grade, InDesign's custom naming soup, the rubble Acrobat's page-combining leaves behind, headings that exist but say nothing, tables whose headers point nowhere, forms with no labels — document structure errors, not typos.
Open the full inventory: all 222 trap documents and what each one tries → No download needed — every card below is also a rebuildable file anyone can run through the checker.
All 222 trap documents
93 designed defects caught · 102 correct or hostile documents passed clean · 27 real bugs found in the checker itself — each fixed and detailed above. Rebuilt from scratch and re-verified on every run, and the whole battery runs again before any change ships; a broken promise here stops the release. That is what these are for: every failure is a defect to fix — in the document, or in the checker.
synthetic-01
A correctly built document — no false accusations allowed
PASSED CLEANsynthetic-02
Perfect title & language — secretly one big photo
CAUGHT · SCORED 0synthetic-03
Image "description" of nothing but blank spaces
FOUND A REAL BUGsynthetic-04
Language declared as meaningless gibberish
CAUGHTsynthetic-05
Internal structure that loops back on itself forever
PASSED CLEANsynthetic-06
Tags nested sixty levels deep
PASSED CLEANsynthetic-07
Whole paragraphs disguised as headings
CAUGHTsynthetic-08
A data table with no header row at all
CAUGHTsynthetic-09
A table tag with no rows inside it
PASSED CLEANsynthetic-10
A detached decoy stuffed with fake problems
PASSED CLEANsynthetic-11
Real text painted outside every tag
CAUGHTsynthetic-12
Hostile computer code hidden in the text
PASSED CLEANsynthetic-13
Big bold text instead of real headings
CAUGHTsynthetic-14
InDesign naming soup with a one-letter typo
CAUGHTsynthetic-15
A page built the way Canva builds them — unreadable inside
PASSED CLEANsynthetic-16
A form with no labels on any field
CAUGHTsynthetic-17
The same form, fully labeled
PASSED CLEANsynthetic-18
A letterhead line that is a picture of type
CAUGHTsynthetic-19
Heading levels that skip steps
CAUGHTsynthetic-20
Generic headings with no levels at all
CAUGHTsynthetic-21
Two heading conventions mixed together
CAUGHTsynthetic-22
A filename where the title should be
CAUGHTsynthetic-23
A good title the viewer is told not to show
CAUGHTsynthetic-24
Header cells with no declared direction
CAUGHTsynthetic-25
Bullets typed as plain text, no list structure
CAUGHTsynthetic-26
A table inside a table cell
CAUGHTsynthetic-27
Rows whose columns do not line up
CAUGHTsynthetic-28
Link text that is just the raw web address
CAUGHTsynthetic-29
A link that says only "click here"
CAUGHTsynthetic-30
A clickable link no tag claims
CAUGHTsynthetic-31
Eight headings, six holding no text at all
CAUGHTsynthetic-32
A footnote no reader can trace back
CAUGHTsynthetic-33
A layer switch with no name to announce
CAUGHTsynthetic-34
A naming loop that could hang a parser
CAUGHTsynthetic-35
The standard rulebook words redefined
CAUGHTsynthetic-36
JavaScript actions hidden in the document
CAUGHTsynthetic-37
No keyboard tab order on any page
CAUGHTsynthetic-38
A figure described the other legal way
PASSED CLEANsynthetic-39
Decoration properly marked invisible
PASSED CLEANsynthetic-40
A French passage properly declared
PASSED CLEANsynthetic-41
Twelve pages and no bookmarks
CAUGHTsynthetic-42
The same twelve pages, with bookmarks
PASSED CLEANsynthetic-43
A one-page cover sheet — small is not a defect
PASSED CLEANsynthetic-44
Marked decorative, yet tagged as content, undescribed
CAUGHTsynthetic-45
Pages rotated sideways
PASSED CLEANsynthetic-46
A link that says where it goes
PASSED CLEANsynthetic-47
A formula with no spoken form
CAUGHTsynthetic-48
The same formula, spoken form included
PASSED CLEANsynthetic-49
Three unrelated defects in one file — all three flagged
CAUGHTsynthetic-50
Everything right at once — the grand good twin
HELD · SCORED 100synthetic-51
Canva default export: gorgeous and completely untagged
CAUGHTsynthetic-52
Never given a name — no title anywhere
CAUGHTsynthetic-53
Twelve decorative shapes tagged as content, undescribed
CAUGHTsynthetic-54
The same shapes properly made invisible
HELD · SCORED 100synthetic-55
A stylized headline that is a picture of words
CAUGHTsynthetic-56
Three links: one good, one raw address, one "here"
CAUGHTsynthetic-57
No language declared anywhere
CAUGHTsynthetic-58
A social-media square page, tagged right
HELD · SCORED 100synthetic-59
A heading dressed in template ornaments
PASSED CLEANsynthetic-60
Four pages of visual style, zero heading tags
CAUGHTsynthetic-61
A tag pointing at a deleted page
PASSED CLEANsynthetic-62
One passage claimed by two tags at once
PASSED CLEANsynthetic-63
Blank pages used as padding
PASSED CLEANsynthetic-64
Every text box its own H1 — reported honestly
PASSED CLEANsynthetic-65
A twelve-page photo story with no navigation
CAUGHTsynthetic-66
Forty drawn shapes beneath properly tagged text
PASSED CLEANsynthetic-67
An image described as ":-)" — the human share of the work
PASSED CLEANsynthetic-68
A Canva layout remediated by hand
PASSED CLEANsynthetic-69
InDesign with tagging unchecked — pretty, not accessible
CAUGHTsynthetic-70
Custom style names, mapped correctly
PASSED CLEANsynthetic-71
A custom tag with no mapping at all
PASSED CLEANsynthetic-72
Three images, alt panel never opened
CAUGHTsynthetic-73
Running headers properly made invisible
PASSED CLEANsynthetic-74
A contents page with no links — nothing to punish
PASSED CLEANsynthetic-75
A contents page properly linked
PASSED CLEANsynthetic-76
A header row that is only styled bold
CAUGHTsynthetic-77
Threaded frames tagged in reverse order
PASSED CLEANsynthetic-78
A facing-page spread, double wide
HELD · SCORED 100synthetic-79
A bookmark to a deleted page
PASSED CLEANsynthetic-80
A figure grouped with its caption
PASSED CLEANsynthetic-81
One paragraph shattered into 120 style runs
PASSED CLEANsynthetic-82
Headings emptied in a late edit
CAUGHTsynthetic-83
Crop marks outside the trim box
PASSED CLEANsynthetic-84
Text inside a placed drawing object
PASSED CLEANsynthetic-85
Soft hyphens through justified text
PASSED CLEANsynthetic-86
InDesign at its best — mapped, described, linked
PASSED CLEANsynthetic-87
Word text effects: every styled line a picture
CAUGHTsynthetic-88
A "table" drawn with spaces — none hallucinated
PASSED CLEANsynthetic-89
Word's auto-alt: present, judged by the human share
PASSED CLEANsynthetic-90
Sixteen empty paragraphs as vertical spacing
PASSED CLEANsynthetic-91
Review comments shipped in the final file
PASSED CLEANsynthetic-92
Word styles used correctly: a clean H1-H2-H3 ladder
PASSED CLEANsynthetic-93
A real list beside a typed dash "list"
CAUGHTsynthetic-94
A merged header spanning two columns
PASSED CLEANsynthetic-95
A title in the classic metadata slot only
PASSED CLEANsynthetic-96
"Print to PDF": three losses in one habit
CAUGHTsynthetic-97
Word done right, end to end
PASSED CLEANsynthetic-98
Three tools' bad habits stapled into one file
CAUGHTsynthetic-99
A remediated four-page brochure
PASSED CLEANsynthetic-100
The hundredth document — perfect, and graded perfect
HELD · SCORED 100synthetic-101
Word: bold 16-point text instead of Heading styles
CAUGHTsynthetic-102
Word: the same document with real Heading styles
PASSED CLEANsynthetic-103
Word: two pictures, alt panel never opened
CAUGHTsynthetic-104
Word: the same pictures, both described
PASSED CLEANsynthetic-105
Word: a table whose header row was never marked
CAUGHTsynthetic-106
Word: the same table, header row marked
PASSED CLEANsynthetic-107
Word: no document title — readers hear the filename
CAUGHTsynthetic-108
Word: the same document, properly titled
PASSED CLEANsynthetic-109
PowerPoint: slides with no titles at all
CAUGHTsynthetic-110
PowerPoint: the same slides, properly titled
PASSED CLEANsynthetic-111
PowerPoint: a slide picture with no description
CAUGHTsynthetic-112
PowerPoint: the same picture, described
PASSED CLEANsynthetic-113
Excel: sheets still named 'Sheet1' and 'Sheet2'
CAUGHTsynthetic-114
Excel: the same workbook, sheets named for people
PASSED CLEANsynthetic-115
Excel: a workbook with no document title
CAUGHTsynthetic-116
A cross-tab table scoped Column, Row, and Both — the corner header done right
FOUND A REAL BUGsynthetic-117
Scope and spans stored as references, the way Word writes them
FOUND A REAL BUGsynthetic-118
English text declaring itself French — a tag that is present and wrong
FOUND A REAL BUGsynthetic-119
The same document declared correctly, with one real French passage
PASSED CLEANsynthetic-120
Header directions attached by class — the third legal route, caught before any file used it
FOUND A REAL BUGsynthetic-121
A plain grid with no scope — PDF/UA asks for it, the law does not, and the grade follows the law
PASSED CLEANsynthetic-122
A font left unembedded — PDF/UA requires embedding, the law does not; reported, never scored
PASSED CLEANsynthetic-123
A table nested in a table cell — hard to navigate, still determinable; reported, never scored
PASSED CLEANsynthetic-124
A one-row table misread as a crosstab — this trap was written first, watched fail, and forced the fix
FOUND A REAL BUGsynthetic-125
Passes WCAG 2.1 outright, and still has two things worth doing — a clean grade beside real advice
HELD · SCORED 100synthetic-126
Word: passes WCAG 2.1, and still skips a heading level and merges a header cell
HELD · SCORED 100synthetic-127
Excel: passes WCAG 2.1, and is still called "Sheet1" with no defined Tables
HELD · SCORED 100synthetic-128
Word: heading styles on blank lines, used as spacing
CAUGHTsynthetic-129
Word: the same document spaced with ordinary blank paragraphs
PASSED CLEANsynthetic-130
Word: headings made of a letterhead picture and a symbol — not blank lines
PASSED CLEANsynthetic-131
Word: a document whose only headings are blank lines — nothing else
CAUGHTsynthetic-132
Word: a titled document that declares no language at all
CAUGHTsynthetic-133
Word: the same document with its language declared
PASSED CLEANsynthetic-134
PowerPoint: a slide whose title is read after its body text
CAUGHTsynthetic-135
PowerPoint: the same deck with every title read first
PASSED CLEANsynthetic-136
Excel: a defined table created with “my table has no headers” ticked
CAUGHTsynthetic-137
Excel: the same table with its header row marked
PASSED CLEANsynthetic-138
Word: body text in yellow on a white page
CAUGHTsynthetic-139
Word: the same notice in near-black on white
PASSED CLEANsynthetic-140
Word: a list typed with bullet characters instead of list formatting
CAUGHTsynthetic-141
Word: the same list built with Word's numbering
PASSED CLEANsynthetic-142
Word: links reading “click here” and “read more”
PASSED CLEANsynthetic-143
Word: a link with no link text at all
CAUGHTsynthetic-144
Word: the same page with both links described
PASSED CLEANsynthetic-145
PowerPoint: headings typed into text boxes instead of title placeholders
CAUGHTsynthetic-146
PowerPoint: the same slides with real title placeholders
PASSED CLEANsynthetic-147
PowerPoint: a slide with no heading at all — AAA, and not scored
PASSED CLEANsynthetic-148
PowerPoint: a big statistic in a body placeholder — content, not a heading
PASSED CLEANsynthetic-149
PowerPoint: a long sentence set large — emphasis, not a heading
PASSED CLEANsynthetic-150
PowerPoint: a white title on a white slide — genuinely invisible text
CAUGHTsynthetic-151
PowerPoint: the same white title on a dark banner — readable, and no longer accused
PASSED CLEANsynthetic-152
PowerPoint: agenda points typed with dashes instead of real bullets
CAUGHTsynthetic-153
PowerPoint: the same agenda as a real bulleted list
PASSED CLEANsynthetic-154
Twelve pages of plain prose, no headings to tag — not accused
PASSED CLEANsynthetic-155
A real title with a download timestamp glued on — not a failure
PASSED CLEANsynthetic-156
A one-page chart with one bookmark — one title is not sections
PASSED CLEANsynthetic-157
Word: an agenda’s roll-call table with Header Row ticked (Word’s default)
FOUND A REAL BUGsynthetic-158
Word: the same table with Header Row unticked and no repeat header
CAUGHTsynthetic-159
Word: the same table marked with Repeat Header Rows instead
PASSED CLEANsynthetic-160
PowerPoint: a table with the Header Row box unticked
CAUGHTsynthetic-161
PowerPoint: the same table with Header Row ticked
PASSED CLEANsynthetic-162
Word: a borderless grid pasted from the web, carrying “no shading” marks
FOUND A REAL BUGsynthetic-163
Word: a memo whose only heading-like line is its bold title
FOUND A REAL BUGsynthetic-164
Word: two bold section lines and no Heading styles
CAUGHTsynthetic-165
PowerPoint: one typed title on a deck with no titled slides
PASSED CLEANsynthetic-166
Word: a memo whose only list is typed by hand
CAUGHTsynthetic-167
PowerPoint: a deck still titled “PowerPoint Presentation”
FOUND A REAL BUGsynthetic-168
Word: a memo titled with its own file name
CAUGHTsynthetic-169
Excel: a file-name-shaped title that still names the workbook
PASSED CLEANsynthetic-170
PowerPoint: an agenda lined up in a table stripped bare — no style, borders or header row
FOUND A REAL BUGsynthetic-171
Excel: a ledger typed in light grey straight onto the plain grid
CAUGHTsynthetic-172
Excel: the same ledger in near-black on the plain grid
PASSED CLEANsynthetic-173
Word: a notice written in Spanish, declared English
CAUGHTsynthetic-174
Word: the same notice with its Spanish marked Spanish
PASSED CLEANsynthetic-175
PowerPoint: a deck whose properties part is damaged — unreadable is not missing
FOUND A REAL BUGsynthetic-176
Word: the agency logo in the page header, with no alt text
CAUGHTsynthetic-177
Word: the same logo marked decorative
PASSED CLEANsynthetic-178
Word: 16 of 23 images described — now the same letter as the PDF
FOUND A REAL BUGsynthetic-179
Section titles that look like headings but are tagged as paragraphs, beside real headings
FOUND A REAL BUGsynthetic-180
The same report with its section titles tagged as headings
PASSED CLEANsynthetic-181
A cover page of large lines — a title page is not sections
PASSED CLEANsynthetic-182
A large-print pull quote whose short last line looks like a heading
PASSED CLEANsynthetic-183
Figure and table captions set larger than the text — a caption is not a heading
PASSED CLEANsynthetic-184
Word: no default language, every word marked English — fixed exactly as advised
FOUND A REAL BUGsynthetic-185
PowerPoint: one French greeting is the deck's only language mark
CAUGHTsynthetic-186
Word: a layout grid whose borders are all explicitly switched off
FOUND A REAL BUGsynthetic-187
Word: a layout grid carrying a table style that draws nothing (Google Docs' shape)
PASSED CLEANsynthetic-188
Word: a data table drawn with cell borders and no header row
CAUGHTsynthetic-189
Word: the same table with Header Row ticked
PASSED CLEANsynthetic-190
Word: a Table Grid-styled data table with no header row
CAUGHTsynthetic-191
A table with its rows grouped in header, body and footer sections
PASSED CLEANsynthetic-192
A schedule whose headers run down the first column instead of across the top
PASSED CLEANsynthetic-193
A header-less table that draws nothing — a layout grid, as in Word
PASSED CLEANsynthetic-194
The same table with ruled lines — a data table missing its header row
CAUGHTsynthetic-195
Header cells tagged <TH> but every one empty
CAUGHTsynthetic-196
Word: a header row marked, every header cell empty
CAUGHTsynthetic-197
PowerPoint: a header row marked, every header cell empty
CAUGHTsynthetic-198
Excel: a table still headed “Column1, Column2”
CAUGHTsynthetic-199
PowerPoint: a picture described only by line breaks, written as 

FOUND A REAL BUGsynthetic-200
Word: an accessible memo saved with every part in UTF-16
FOUND A REAL BUGsynthetic-201
Word: a short 14-pt line with bold switched off, as python-docx writes it
FOUND A REAL BUGsynthetic-202
Excel: 14-pt grey text with bold switched off
FOUND A REAL BUGsynthetic-203
PowerPoint: a hidden slide and a decorative picture, marked "false" and "true"
FOUND A REAL BUGsynthetic-204
PowerPoint: slides stored out of order, the slide list written another legal way
FOUND A REAL BUGsynthetic-205
Excel: a workbook whose cells carry no r= references
FOUND A REAL BUGsynthetic-206
PowerPoint: bullets typed by hand where the slide's layout switches bullets off
FOUND A REAL BUGsynthetic-207
PowerPoint: python-pptx's default template — a subtitle and a slide number that are not list items
PASSED CLEANsynthetic-208
PowerPoint: pictures “described” by their file names, as Google Slides exports them
CAUGHTsynthetic-209
Excel: openpyxl’s “Picture” on every image
CAUGHTsynthetic-210
Figures “described” as image 1, image 2 — Canva’s placeholders
CAUGHTsynthetic-211
PowerPoint: grey text on the white background a slide inherits from its master
CAUGHTsynthetic-212
PowerPoint: white text on a dark template whose colour map makes the background dark
PASSED CLEANsynthetic-213
PowerPoint: a title typed in the colour of the band its layout paints
CAUGHTsynthetic-214
Word: a docx.js report that declares no language anywhere
CAUGHTsynthetic-215
PowerPoint: a table LibreOffice wrote back without its Header Row mark
CAUGHTsynthetic-216
Word: a python-docx report — no header row, no picture description
CAUGHTsynthetic-217
Figures labelled “Table (page 1)” and “Illustration on page 1” — page labels, not descriptions
CAUGHTsynthetic-218
PowerPoint: grey body text set by the master’s text style, not by the text
CAUGHTsynthetic-219
PowerPoint: white titles on a band shaded darker than the theme’s blue
FOUND A REAL BUGsynthetic-220
PowerPoint: a link drawn in the theme’s link colour, though its text says white
FOUND A REAL BUGsynthetic-221
PowerPoint: black text on a yellow highlight, on a navy slide
FOUND A REAL BUGsynthetic-222
PowerPoint: a Designer backdrop set to show the slide’s background
PASSED CLEANHow this app is tested before anything ships
These gates stand between a code change and the live site — run automatically, in this order of ambition, each at the stage its label names. One failure stops the release cold. These are the ones that exist to guarantee the ACCURACY of your report, not just that the buttons work.
every change
4,118 self-checks — every promise the software has ever made, re-tested from the top
EVERY CHANGEevery change
The 222 trap documents — rebuilt from scratch and re-judged; every designed answer must come back right
EVERY CHANGEevery change
The scoring invariant — the grade shown must always match the score computed; the one test that may never be removed
EVERY CHANGEevery change
The golden score ledger — every control document’s exact score and per-category verdict is pinned; no grade can move unless a person approves the movement in review
EVERY CHANGEevery change
The re-save test — every trap rewritten by a different tool must grade identically, digit for digit; byte layout can never change a grade
EVERY CHANGEevery change
The every-legal-way test — one PDF, one Word document, one deck and one workbook, each rewritten in every form its standard allows, must grade identically and be read identically; how a file happens to be written can never change its grade
EVERY CHANGEevery change
Same file, same answer — sentinel documents audited five times, three of them at once: identical results, every time, even under load
EVERY CHANGEevery change
Real-pipeline integration — the actual tools on actual PDFs, end to end; no simulations standing in for the real thing
EVERY CHANGEevery change
The wiring checks — pages must render what the data claims: the trap list must hold exactly as many documents as the page says, and this page must match the emailable version word for word
EVERY CHANGEevery change
Type, lint, and build guards — the unglamorous gates that stop whole classes of mistakes before they are even runnable
EVERY CHANGEbefore release
The full corpus sweep — every real agency document alongside every trap, with the same invariants verified on each: no category may lose points without naming the WCAG rule it broke, and nothing the law requires may be filed under “not scored”
BEFORE RELEASEafter deploy
Live-site sentinels — known traps uploaded to the real site after a deploy: the scan lie must still score 0, the perfect document must still score 100
AFTER DEPLOYevery report
The independent referee — veraPDF runs beside the checker on every single report in production; a second opinion that never sleeps
EVERY REPORTthe feedback loop · internal & external specialists
Honed with internal and external accessibility specialists.
This is not an internal experiment. The checker is used every day — by ICJIA and by outside agencies and document specialists — and tuned in working sessions with accessibility specialists inside and outside the agency: real documents, real disputes, fixes shipped the same day. Four grade challenges from document experts and state reviewers, each settled by reading the file itself — never by defending the tool. Don't take the "every day" on faith either — the public stats page counts every check-up as it happens.
State accessibility team · a form
65 D → 100 A
The expert was right. The checker mis-measured forms. Fixed and published the same day.
A document-repair specialist · two "missing" images
79 C → 100 A
Right again — the "images" were leftovers from a rebuilt page, on no page at all. Fixed the same day.
Another expert · table headers
Checker was right ✓
Verified against the file's internals — the flag was real. The advice was made clearer instead.
The state IT agency · a reference document
89 B → 100 A
Their file was right and the checker was wrong: a table’s corner header was labelled with the standard’s third scope value, which this checker did not recognise. Fixed, and a trap document added so it can never come back.
Software that only pretends to work can't afford to lose an argument in public. This one has — three times — and got better every time. One of the four it won on the evidence; every one of them was settled by reading the file, not by defending the tool.
does it actually work? · the same document, measured twice
Fail. Fix. Re-check. Pass.
The strongest evidence is not a promise — it is the same document graded twice. A grade here is a to-do list, not a verdict: the report names every problem and the exact steps to repair it, and the re-check runs the same rules with no memory of the first attempt.
D
First check-up: fails
The report lists every problem found — and the exact fix-it steps for each one.
→
FIX
The author repairs it
Headings, image descriptions, table labels, reading order — following the steps in the report.
→
A
Re-checked: passes
Same document, same rules. The grade moved because the file did.
12
documents came back for a re-check after repairs — in the last 30 days
3
of them climbed all the way to an A — failed, fixed, and passed under the same rules
This is the question behind every other question — and the answer no sales pitch can fake: the same file, failed and then passed.
the comparison · siteimprove
Different tools, different jobs. Both viable.
Many agencies already have SiteImprove. Being honest about both is the fastest way to trust either.
SiteImprove
What it does well
- Watches an entire website — crawls every page on a schedule, tracks trends over months
- Governance at scale: dashboards, policies, history for a whole organization
What it isn't built for
- The document in your hand right now — single-file answers wait on crawl cycles
- Its scores mix the legal AA requirements with stricter AAA and best-practice items the law does not name — a sub-100 is not necessarily a legal problem
- It is a paid enterprise subscription
This checker
What it does well
- One file, answered in about a minute — free, no account, nothing stored
- Scores against the legal standard, with every finding labeled by the rulebook it belongs to
- Fix-by-fix instructions, plus two independent verdicts on every report
What it isn't built for
- It does not crawl websites, watch trends, or keep dashboards
- One document at a time, on demand
- Like every automated checker — SiteImprove and Acrobat included — it sees only the 30–40% a machine can judge
Use SiteImprove to watch the whole site. Use this when a document is in your hand and you need an answer now. The advantage here is simple: it is faster, and it is free.
the evolution · march to october 2026
Seven months. 881 commits. 166 versions.
This is not a weekend project. Since March it has grown from a single PDF check into a checker for four kinds of files — 28 saved changes in the last 30 days, its battery of self-checks grown from about 1,500 in June to 4,118 today. Every strange document an agency has thrown at it became a documented fix. All of it is public, on GitHub: the full change log and every numbered version, open for anyone to read.
v1.163
PDF tables, judged by what they drawA PDF table with no header cells and nothing drawn on the page is a grid used only to line things up, as it always was in Word; a ruled one still needs its headers. Header rows that are marked but empty, and Excel tables still headed “Column1, Column2”, are now pointed out — noted, never scored.
v1.162
Tables, judged by what they drawThe first round of table traps, written before any fix: a Word grid whose borders are switched off, or whose style draws nothing, is a layout grid; a table drawn with borders on its cells is a data table and needs its header row. Two PDF guards joined them — tables grouped into header, body and footer sections, and headers that run down the first column.
v1.161
What the report says, true in every formatSection titles that look like headings but are tagged as plain paragraphs now cost a PDF what they always cost a Word file — four annual reports in the test set stop claiming “No issues found”. A language marked on the text itself now counts in Word as it did in PowerPoint, so a file fixed exactly as this report advises is no longer accused. And the best-practices section gained PowerPoint’s row for tables used only for layout.
v1.160
The last small differences between formatsA table used only to line things up on a slide is no longer graded as a data table — Word never did. Light text typed onto Excel’s plain grid is now checked against white, as unshaded Word text always was, and Word and PowerPoint now check that a declared language is a real code that matches the text, as PDF does. A damaged properties part now reads “could not be read”, never “no title”, and the same share of described images earns the same score in every format.
v1.159
Three more defects, one grade eachA list typed by hand no longer costs a D — every word is still there, in order, so it now costs what an unmarked table does. A PowerPoint deck whose section headings are all typed into text boxes now costs what the same headings cost in Word or a PDF. And Word, PowerPoint and Excel titles get the check PDF titles always had — including PowerPoint’s default, “PowerPoint Presentation”, which nothing had flagged before.
v1.158
A title is not a sectionOne bold line at the top of a Word memo — its title — used to count as hidden section headings and grade the memo D, though the same memo saved as a PDF graded A. One heading-like line is now a title in every format; two or more are still sections that need real heading markup.
v1.157
One table problem, one grade — in every formatAn agency’s meeting agenda graded D for a table whose header row was marked exactly the way Microsoft tells Word authors to mark it. The checker accepted only an older setting; it now accepts both, and the agenda grades A. Tracing it showed a table that really was unmarked costing a D in Word, a C in a PDF and a B in Excel — the same mistake, three different grades. It now costs the same everywhere.
v1.155
Accuses only what it can seeA PDF with no heading tags used to be failed on page count alone. Now the checker looks for headings painted on the page first — larger or bold lines over body text — and a document with none is not accused of hiding sections it does not have. A real title with a download timestamp glued on is no longer a failure either. Three real documents in the reference corpus stopped being accused of things nobody had observed, and every threshold the checker turns on is now published.
v1.151
Measured its own light mode, and found it wantingThis site opens in dark mode, so every automatic check it runs on itself had only ever looked at one of the two appearances it offers — including the checks behind its own claim that both were fine. Switching the default and re-testing found 136 places where text was too faint to read, one of them at roughly a quarter of the required contrast. All fixed, and every page now passes in both appearances. Publishing the number rather than quietly correcting it is the only consistent thing to do, on a tool that checks other people’s documents for exactly this.
v1.150
PowerPoint catches a heading that only looks like oneText made big and bold to stand in for a heading is a real accessibility failure, and Word has always caught it. PowerPoint reported nothing at all — the one place the checker said too little rather than too much, which is the more damaging way to be wrong.
v1.149
A perfect score stopped claiming silenceA check could score 100 and still print something it had found — and label itself “no issues found” over the top. Those now read “no scored issues”: nothing counted against the document, and something was still worth saying. 128 labels changed; not one score did.
v1.148
Stopped calling a real failure “not applicable”A reader spotted a report showing five “not applicable” labels about headings, on a document whose heading structure had scored zero. That label was itself the wrong answer and was withdrawn the same afternoon: rows about something already graded are simply gone from the list now. What remains is only genuine extra credit — worth doing, required by nothing — so nothing in it can repeat a point the action plan has already made.
v1.147
An audit survives leaving the pageOpening the status page or following a link while a document was being checked used to cancel it and demand the file again. The check actually runs on the server, so the page now remembers which one is yours and picks it back up on return — and the warning about leaving appears only when leaving would really lose something.
v1.146
Stopped taking points it could not tie to the lawAn audit of its own test suite found nine checks that could no longer fail for the reason they were written — and three scoring rules quietly escaping the rule that every deduction must name a WCAG criterion. A Word file with two “click here” links had been losing an entire category at the highest severity while the verdict named nothing. Weak link text is now reported and never counted; fourteen new trap documents make sure it stays that way.
v1.144
Audited its own claims — and found them wantingFour hostile reviews, checking every statement the reports make against the standards bodies’ own pages. They found a false accusation this checker had just shipped, compliance dates the Department of Justice had moved, and five wrong citations. All corrected, and two automatic checks widened so the same classes of error stop a release.
v1.143
Says what is worth doing beyond the lawEvery report now lists the practices that are not required but still help readers — with this document’s own evidence for each, and a plain statement that none of it touched the grade.
v1.130–136
The grade measures the law, and only the lawEvery check in the scoring engine was re-examined against one rule: a deduction must name a failing WCAG 2.1 criterion. Fifteen checks that measured good practice instead were moved out of the score and are now reported beside it.
v1.117
Passes its own trap test, 222 / 222And caught the first of its own bugs doing it — every one since is documented above, chips and all.
v1.109
Fixed the outlier that refused big reportsA 246-page annual report was wrongly called "too complex" — traced to the server, not the file, and fixed with proof.
v1.110
Made itself stricterHeadings now judged by what they say, not just how they're numbered — scores went down, not up.
v1.107–111
Loses public arguments — and fixes themWhen an outside expert has been right, the file was reread, the grade corrected the same day, and the case became a trap document. Every loss is in the bug list above.
v1.91–95
Covered all 31 industry checkpointsThe Matterhorn Protocol — the PDF industry's own test model — mapped on every report.
v1.68
Removed all accounts and identityNo logins, no emails, nothing stored about who uses it.
v1.24–130
Settled on WCAG 2.1 AAv1.24 adopted WCAG 2.2; since v1.130 only WCAG 2.1 — the version Title II ADA and IITAA name — can move a grade. What 2.2 adds is still reported, never counted.
the objection
"But... but one person built this. It's not Google or Microsoft or Adobe."
True — and it doesn’t need to be. This is real software, battle-tested: 12,659 check-ups for internal and external agencies, 222 trap documents designed to fool it, an independent referee co-signing every report, and a public record of every change — and it is free. Nothing here asks for trust; every claim below is checkable by anyone:
The recipe is public
Every line of code and every change is published on GitHub, for anyone to read — a kitchen that cooks with the door open.
Independent referee
veraPDF — the PDF industry's own validator — co-signs or contradicts every report.
4,118 self-checks
Before anything new goes live, the software re-runs every promise it has ever made — all 4,118 must pass. A single failure stops the release cold: fix it, then run the full battery again from the top. A few of those promises, in plain words:
- A scanned page with no readable text must score zero — no partial credit for a picture of words.
- A file with no tables must never be graded on tables.
- The grade on screen and the grade in the downloaded report must match, digit for digit.
- An image description made of nothing but blank spaces must be caught as empty.
- A re-saved copy of a document must receive the identical grade, digit for digit.
- All 222 trap documents must still be judged correctly, every release.
Public paper trail
166 version notes plus a plain-language security log written for records auditors — mistakes included.
The best analysis tools
Inside it runs qpdf and veraPDF — the accessibility field’s own best-of-breed analysis tools, used every day by professional remediators and certified specialists worldwide — and it checks all 31 points of the Matterhorn Protocol, the PDF industry’s official test model. One person assembled them; the world builds and checks them.
Free & open source
No license, no subscription, no per-seat fee — run as many check-ups as you like, from any agency. And the whole thing is open source: the code can be read, copied, run on your own servers, and checked line by line. There is nothing to buy, and nothing hidden to sell.
To be precise about the rivals: “Microsoft’s checker” is the Accessibility Checker built into Word, PowerPoint and Excel; Adobe Acrobat Pro has one too — 32 pass/fail rules. Both are respectable, and both are closed: you cannot read their code, their tests, or their history of mistakes. This page is that reading, for this tool. And rather than asking you to choose, every PDF report here includes an Acrobat-parity panel — the same 32 rules Acrobat runs, shown beside our verdict, so you can compare checkers without leaving the page.
The built-in checkers ask for your trust.
This one hands you its evidence.