Contrast was never checked because no corpus ever included the stylesheet

Built and measured 2026-08-21. Re-runnable, offline, no credentials:

cd portfolio/research/a11y-contrast-decidability
python3 -m unittest discover                                   # 11 tests
python3 measure.py --recompute-from results-2026-08-21.json    # offline
# to regenerate: npm ci && npm run build in each venture in VENTURES, then
python3 measure.py --out results.json

Why this was built

Both of this portfolio's WCAG engines record colour contrast (SC 1.4.3) as manual, with the same one-line justification:

Needs computed colour, which needs CSS and rendering.

Five measurement directories repeat it as settled — a11y-own-sites and a11y-generated-sites both close with "every criterion the engines mark manual — contrast, keyboard operability, focus order and visibility — is untouched … nothing here shrinks it."

Nobody had tested it. And the reason nobody could is structural rather than an oversight: every accessibility corpus this portfolio has ever measured was HTML only. a11y-evidence-record reads theme templates, a11y-own-sites walks .next/server/**/*.html, a11y-generated-sites renders component trees to markup. A stylesheet appears in none of them. Handed markup with no CSS, an engine cannot decide contrast — so the claim reproduced itself for five passes as a property of the corpus that read as a property of the criterion.

The delivered artifact is not markup. It is markup and the stylesheet it links, and that stylesheet names its colours outright.

What is here

The resolver itself lives with the probe it belongs to, in portfolio/pipeline/probes/eaa-wcag-triage/ — colors.py (colour syntax, alpha compositing, the WCAG ratio), css.py (a deliberately small stylesheet reader), contrast.py (per-element resolution). 112 tests.

Finding 1 — 98.76% of it is decidable, and the residue has names

85 built pages across four ventures, 10,657 text-bearing elements:

The 132 undecidable elements are not a shrug. Each carries a named reason, and the three that occur are the honest residue:

That list is the product, not the ratio. The candidate (eaa-remediation-pipeline) sells an evidence record whose coverage limits are stated in the record. "1.4.3 needs a human" is a limit stated at the wrong resolution: it prices a manual review of every element on the page when the artifact settles all but 1.24% of them. A per-element tally with reasons is the same honesty at a resolution the buyer can act on.

Finding 2 — a single-scheme scan misses 39% of the real failures

prefers-color-scheme is not like a viewport breakpoint. It has exactly two states, and the stylesheet spells out both, so each is decidable — the page must satisfy 1.4.3 in whichever one the visitor's system asks for. Scoring both and taking the worse verdict is therefore not conservatism; it is the criterion.

Over the corpus as main built it when this section was measured:

A checker that renders or resolves one colour scheme reports those 384 elements as clean. They are not clean; they are unreadable for anyone whose system is set to dark, which is most people on a phone at night.

The mechanism is mundane and, once named, obvious: a component writes text-gray-500 with no dark: counterpart, inside a container that does have dark:bg-gray-800. In light mode the pair is fine. In dark mode the text stays mid-grey and the background goes dark, and the ratio collapses. print-atelier did this 384 times; the same file showed the correct pattern (text-gray-500 dark:text-gray-400) one component above, so this was a lapse that repeated, not a decision — and, being internal-only venture code, it was fixed the same day (#542, Finding 4 below): main now has 0 dark-only failures, all in this class.

Finding 3 — gradients are decidable, and the bound is exact

A gradient background is the canonical reason a static checker gives up. It does not have to be. CSS interpolates each sRGB channel linearly between adjacent stops and WCAG relative luminance is monotonic in each channel, so the luminance attainable anywhere on the gradient is exactly the interval between its stops' luminances — no interior extremum exists to be missed. For a fixed text luminance the contrast ratio then has its minimum at the point of that interval nearest the text and its maximum at the farther end.

So a gradient yields a bound, and a bound wholly on one side of the threshold settles the criterion. Only a bound that straddles it is undecidable, and that gets its own reason (gradient-straddles-threshold). Alpha compositing preserves the property, because compositing is affine per channel: a translucent panel over a gradient is still bracketed by its composited stops.

This is why local-site-studio's four generated themes — every one of which paints a radial-gradient shell — decide cleanly rather than falling into the undecidable bucket.

Finding 4 — three ventures were shipping the defect they plan to bill for finding

The scan found 965 decided AA failures on this portfolio's own built pages, as originally measured against main by this section (see Why the committed file measures main below). Not marginal ones: text-zinc-400 on white is 2.62:1 against a required 4.5:1.

Since then, print-atelier's 611 have been fixed and merged (#542, 2026-08-21, same day) — internal-only remediation, agent-decidable because that venture has no deploy workflow. local-site-studio's 21 and ai-first-markets's 333 remain open on main, still held for the account owner:

local-site-studio's and ai-first-markets's fixes are written and verified by rebuild: 62 lines of Tailwind class-name substitution across 11 files, each substitution's resulting ratio computed before it was applied, the lowest 4.63:1 against a required 4.5:1. They wait on the account owner, not on any further measurement. The committed results-2026-08-21.json and test_measure.py already measure the post-#542 tree (354 total failures, print-atelier at 0) — this section's prose was the one place still describing the pre-#542 965/611 state; corrected here rather than left as a citation that no longer matched its own committed artifact, the same gap a11y-own-sites/README.md hit and fixed the same day.

Why the committed file measures main

This directory, the resolver behind it and the two venture colour fixes were built as one pass and opened as #523. Bundling the fixes put the whole 4,542-line diff in the human-gated class — engine, tests and measurement included, none of which touch a live site — so #523 offered to split, and #524 took the split and landed the internal half.

What #524 carried across unchanged was #523's results file, and that file is a measurement of #523's tree, where the two ventures are already fixed and read 0. On main they are not fixed. So the prose here said 965 open failures while the artifact beside it totalled 611, and this directory's CI workflow cannot catch that — its own header comment says so: "a colour change in venture source makes the committed results stale — CI cannot detect that, only a re-run with builds can." Nothing was wrong with the engine; the artifact simply described a different tree than the one it was committed to.

This pass re-ran it with builds. All four ventures rebuilt from main's source with npm ci && npm run build, rescanned, results file replaced. The numbers it produced are #523's own pre-fix column, re-derived from a second checkout: 85 pages, 10,657 elements, 98.76% decidable, 965 failures against 581 light-only, 384 dark-only, and the same 106 / 24 / 2 split across the three undecidable reasons. Two sessions, two checkouts, same arithmetic — which is the strongest evidence this measurement has that it is reproducible at all.

test_measure.py now pins the per-venture counts instead of asserting the two ventures are clean. Whenever #523 or the print-atelier work lands, this suite fails until the corpus is rebuilt and the file regenerated. That is the same coupling the original guard enforced, pointed at the tree that actually exists. The print-atelier trigger has since fired: #542 landed the fix, a later pass rebuilt the corpus and replaced the results file again (354 total, print-atelier at 0), and test_measure.py was updated to pin that state — see Finding 4 above. #523/#531's two ventures have not landed, so the file still measures them open.

What could not be measured here

• Only what the four Next.js ventures build. garmin-world-traveler-watchface, world-clock-widget and the CLI ventures emit no HTML, and local-site-studio's delivered customer sites are rendered per request — a11y-generated-sites holds that corpus, and its 16 pages resolve 100% decidable and 0 failures against this venture's stylesheet.
• No render, no interaction, no assistive technology. 1.4.11 (non-text contrast), keyboard operability, focus visibility and target size are untouched and remain manual. This pass moves exactly one criterion.
• Text over images is still undecidable, and so is text under an opacity. Both are reported, not silently passed.
• The engines have diverged. eaa-wcag-triage now checks 1.4.3 and the sibling a11y-evidence-record/audit.py does not. a11y-engine-agreement established that silent asymmetry is the danger; this one is declared here and in both engines' coverage catalogs, and reconciling it is the follow-up.
• The corpus is this portfolio's own code. Four ventures built with the same framework and two with the same palette is not a sample of the web. The 98.76% is a measurement of what a Tailwind-built Next.js artifact settles; a page whose colours come from a stylesheet the scan was not given decides nothing at all, and the tally says so rather than reporting it clean.