WCAG Color Contrast Check: AA and AAA Ratios: test results

Verdicts

Date

2026-11-12

Strong · claude-sonnet-5-5 (Claude Code alias "sonnet")

Right on all 31 pairs in the 5 cases, read by hand against the formula. On the light CSS it put the fix for grey text on the tinted panel against the panel itself and gave the white-on-blue button a dark text colour; on the dark tokens it blended the transparent text and the opacity with the surface before taking the ratio; on the form it treated the placeholder as text, 18px bold as normal text, the border and the icon as 3 to 1 components and the disabled button as exempt; on the brand palette it expanded #abc-style short hex and the name gray, called the logo exempt and passed 24px regular text at 3.78. On the clean theme it answered that all pairs pass, including the two pairs just above 4.5. The ratio on the line was within 0.1 of the true one in all 31 pairs, and the closing Verdict line was exact in four of the five answers (in the fifth, the dark tokens, it added a remark after the line).

Weak · claude-haiku-5-5 (Claude Code alias "haiku")

Right on all 31 pairs, with the true ratio within 0.1 on every line and, for every pair marked failing AA, a fix that passes on the real background; the logo and the disabled button were written as exempt. Its closing line was less tidy than Sonnet's: in four of five answers it added a remark after the Verdict line.

With and without the skill

Tested 2026-11-12.

Read by hand. Without the file Sonnet gave a right answer on the two cases where it answered (the light stylesheet and the dark tokens, all 12 pairs right), but on the other three, with no tools in the test, it wrote out calls to a shell to compute the ratios and never produced an answer, in both of two runs on the form and the brand palette and in the final run on the clean theme as well; the run on the form cost about $1.30 in tokens and the clean theme $1.66. Those pairs count as not answered. So what the file gives Sonnet is mostly the table that lets it finish the arithmetic itself, not new rules. Haiku without the file answered every case with the right verdicts and good fixes, but called the logo wordmark and the disabled button failures (2 pairs); it also computed a few ratios to 4.62 where the true one is 4.54 on a tinted footer, which stayed inside the pass verdict.

Same cases and the same checks with and without the skill. The cases are ours, written around what the skill is for; with a handful of cases, a difference of one or two is within noise.

Note

Five cases written by us on 12 November 2026 (four planted, one clean): a light stylesheet, dark-mode design tokens with rgba and opacity, a form stylesheet and a short brand palette, with 31 pairs in all; every borderline pair is at least 0.04 from its threshold. Correctness was scored by a script that applies the WCAG 2.2 formula (not by patterns): the AA verdict must match, a failing pair needs a fix colour that passes on the real background, and an exempt pair must not be called a failure; a second count also needs the ratio within 0.1. The script was first proved on its own: the anchors 4.542, 4.478, 2.849 and 21, the ideal answers pass, and every realistic wrong answer (about 4.5 passes, a fix that works only on white, rounding 4.499 up, 18px bold as large text, a placeholder at 3 to 1, transparency without blending) fails. The request names the pairs, asks for a ratio like 4.52:1, AA pass or AA fail and a fix colour; it does not give the formula or the thresholds. One run per model, case and side on 12 November 2026 (the side with the file was run twice, because the worked example in the file was reworded after the first run; both runs gave 31 of 31 for both models and the second is the one counted), and one earlier run of Sonnet alone without the file, before the file was written, which gave the same picture. The two public themes this idea came from have no rule for contrast; only with and without the file was compared. The skill's table of 256 values was generated by a script and checked against the four anchors; it is exact to 0.007 on 300000 random pairs.

What was not measured

• Models other than the two named above were not run.
• Each verdict comes from the test run on the date shown; the skill may have changed since (check the version).
• Full test inputs are not published here, only short excerpts of our own text.
• Results on your own texts, languages and domains can differ.