06-reference/research

pdf ua1 vs wcag22 aa residual delta

2026-11-01·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)
accessibilitywcagpdf-uatypstverify-stack

The residual delta is real, it is six criteria wide, and the Typst flag is both stronger and narrower than the parent brief claimed

The question

"Does PDF/UA-1 (ISO 14289) conformance actually reach WCAG 2.2 AA for a static document, or is there a residual delta the Typst ua-1 flag would leave uncovered?"

Context: [[2026-09-03-wcag-gaps-verify-pdf-output]] recommended a three-part fix whose load-bearing first part is flipping typst compile --pdf-standard ua-1, on the stated assumption that PDF/UA-1 conformance closes WCAG 1.1.1, 2.4.2, 3.1.1 and the structure-tree half of 1.3.1. That brief did not verify the mapping. If a delta exists, the recommended split under-covers.

What we already know (from the vault)

What the web says

Three layers, kept separate on purpose. Layer (a) is what the standard requires, (b) what the Matterhorn Protocol can mechanically check, (c) what the Typst flag actually enforces.

Convergences and contradictions

Synthesis for RDCO

A residual delta exists, it is six criteria wide for a realistic RDCO artifact, and the parent brief's mapping claim holds for the four criteria it named. --pdf-standard ua-1 on Typst 0.14.2 does close 1.1.1 (alt presence), 2.4.2 (document title), 3.1.1 (document language) and the structure-tree half of 1.3.1, and it closes two more the parent did not claim: consecutive heading levels and non-empty heading titles, both inside 1.3.1. I verified the whole boundary by compiling fixtures on the installed 0.14.2 and by extracting the validator's own error-string table out of the binary. So the headline is not "the recommendation was wrong." It is "the recommendation was right about what it checked and silent about two criteria it never considered."

The delta, each line a criterion that applies to a static document and is not guaranteed by the flag:

Three softer residuals belong in the same list at lower weight, because they are genuine but judgment-bound rather than measurable: 1.1.1 adequacy (the flag checks that alt exists, never that it is correct; Matterhorn 13-005, 13-007, 01-001 and 01-002 are all human), 1.3.2 Meaningful Sequence (reading-order correctness, Matterhorn 09-001, human), and 2.4.6 Headings and Labels (descriptiveness). And three candidates the dispatch asked me to test should be retired, not tracked: 1.4.4 Resize Text is the reader application's responsibility per WCAG2ICT; 1.4.12 Text Spacing is scoped out for fixed-layout documents by WCAG2ICT Note 1; 1.4.10 Reflow is enabled by the tagging the flag already guarantees and carries a WCAG2ICT Note 4 escape. Naming them as retired is worth as much as naming the gaps, because the parent brief's pattern of tracking category errors (focus order, keyboard navigation, landmarks) is the failure mode to avoid repeating.

Does the three-part recommendation under-cover? Yes, in two places, and it over-specifies in a third. It under-covers by omitting 2.4.4 and 3.1.2 entirely: no tier touches either, and 3.1.2 cannot be fixed by a check at all. It over-specifies "check 13," which proposed asserting /StructTreeRoot present, /Lang set, /Figure count equal to /Alt count, no heading-level skip, and tables carrying /Table and /TH. Four of those five are now build-time hard errors, so putting them in a large-language-model critic is duplicated spend - exactly the argument the parent brief itself made against adding critic surface. Only /TH survives. The token-pair contrast validator in part two is necessary and is still the single best item in the recommendation, because it fixes the defects at contract level where they propagate to web, video and application surfaces. But it is not sufficient for 1.4.3: my failing fixture used an ad-hoc rgb("#bbbbbb") that appears in no DESIGN.md token table and would pass a token validator while rendering at 1.92:1. Palette validation and artifact validation are different claims.

The minimal addition is four mechanical checks, one source lint, and one honest expansion of the human gate. Mechanically automatable, zero large-language-model cost, belongs in the code-eval tier: (13) table semantics - every /Table subtree contains at least one /TH, and every /TH carries a /Scope (Matterhorn 15-003, the one machine-checkable table condition); (14) heading presence - heading-tag count greater than zero, and the document outline is non-empty; (15) link text - no link's visible text matches a deny-list (click here, read more, here, learn more, this link), and no two links with identical visible text resolve to different target URIs, both pure string comparisons over the extracted text and annotation dictionaries; (16) language-of-parts source lint - fail the build on any text(lang:) in the .typ source that differs from the document language, with the message that 0.14.2 cannot represent it. Check 16 is a blocker, not a validator: it stops us shipping a 3.1.2 failure we cannot fix, and it should carry a pointer to the Typst issue tracker rather than a remediation step. Semi-mechanical and worth building but worth distrusting: a rendered contrast sampler over the page raster, reliable for flat backgrounds and unreliable over photographs and gradients, which is why it cannot replace the token validator. And the human gate must stay, and must grow by three items: alt-text adequacy, decorative-versus-real classification, reading-order correctness, heading descriptiveness, use-of-colour, images-of-text judgment, and whether a declared language is appropriate (Matterhorn 11-007). Every one of those maps to a Matterhorn condition tagged human-only. That is not a scoping failure on our part; it is the protocol's own assessment of its automation ceiling.

Calibration. The mapping is mostly sound with a real but bounded delta, and I want to be clear that this research did not overturn the parent brief. Confidence is high on layer (c): I ran the compiler on the installed 0.14.2 in both passing and failing configurations for every claim above, and cross-checked against the validator's error-string table lifted from the binary. High on layer (b): the 87/47/2 split and every per-condition machine/human tag are verbatim from the Matterhorn Protocol 1.1 document. High on the WCAG2ICT applicability reads, which came from the W3C note directly. Medium-high, and this is the one real weakness, on layer (a): I read ISO 14289-1's scope clause verbatim from the International Organization for Standardization's free abstract page, but I did not read the normative body of the standard. Every claim about what clause 7 requires is the PDF Association's restatement - the Matterhorn conditions' UA1:7.x citations plus PDF/UA in a Nutshell - not the standard's own words. That is a cheap gap to close and I did not close it: the PDF Association distributes ISO 14289-1 at no cost, sponsored, 17 pages, but behind a cart and checkout flow that a research sub-agent should not complete on the founder's behalf. Worth one minute of his time if any of this becomes a published claim rather than an internal build spec.

Why this is in the vault

This closes the first and most load-bearing open follow-up in [[2026-09-03-wcag-gaps-verify-pdf-output]], whose part-one recommendation (flip --pdf-standard ua-1) is still unbuilt and was resting on an unverified standards mapping. It converts that recommendation from "flag plus validator equals AA" into a specific, costed amendment: keep part one, keep part two, replace the proposed check 13 with four narrow mechanical checks plus a source lint, and accept that 3.1.2 is blocked at the Typst layer rather than closable by a check. It also corrects two inversions that would have mis-tiered the work (1.4.1 is a flag gap not a standard gap; 3.1.2 is a toolchain gap not a standard gap).

Open follow-ups

Related

Sources

Research trail - what was reachable and what was not. iso.org and all of pdfa.org return HTTP 403 to both curl and WebFetch (Cloudflare bot block); everything from those two hosts was retrieved through headless Chrome instead, consistent with the standing note that Playwright passes Cloudflare blocks. The two PDF Association documents were read via in-page pdf.js rather than downloaded. One paywall was encountered and skipped, not circumvented: ISO's own CHF 100 store listing for ISO 14289-1. A zero-cost sponsored channel for the same document exists at the PDF Association but routes through a cart and checkout, which was deliberately not completed; this is the only unreachable primary source and it is the cause of the single medium-high confidence label in the Synthesis. Two URLs in the original dispatch were stale: pdfa.org/resource/matterhorn-protocol/ 404s (the slug is now the-matterhorn-protocol), and the PDF Association's browsable Matterhorn database at taggedpdfschool.pdfa.org/matterhorn has confirmed fidelity defects against the PDF (checkpoint labels 26/27/28 are shifted, condition 13-008 is missing, and the two no-test conditions are mislabelled as human), so the PDF was treated as authoritative for every quoted condition.

Local primary - measured, not read:

Web primary:

Paywalled, skipped: ISO's own CHF 100 listing for ISO 14289-1:2014. Unreachable: the normative body text of ISO 14289-1:2014 (clause 7 and Introduction), free but behind a checkout flow.