Accessibility

Contrast, colour-vision and "not by colour alone" for palettes, cards and kits. Read when the audit reports a gate or the user asks. Tags: [strong] [limited] [practice].

The gate: WCAG 2.2 contrast [strong]

• No rounding: 4.49:1 fails. Measure rendered colours (composite alpha first). Both modes are separate gates.
• Our heuristic: weight ≤ 300 or a hairline serif gets 4.5:1 even at large sizes (thin faces render fainter).
• WCAG 3 is a draft; WCAG 2 is the only gate here. APCA is not computed: its reference code is all rights reserved (apca-w3); if asked, say so and point to its own tools.
• Logos are exempt from 1.4.3 and 1.4.11; the logo audit still asks ≥ 3:1 or a tile as practice (logo.md). Button fills with a text label need no 3:1 against the page; their focus ring does.

Pairs the palette audit tests (both modes)

text, textMuted on background/surface/surfaceAlt (4.5); onPrimary/primary, onAccent/accent (4.5); link on surfaces (4.5) and against body text when not underlined (3, G183); focus against background and filled controls (3); status colours as text (4.5) or icons (3); input borders when they are the only cue (3); extended colours with their on (4.5). The card probe adds every text node on the card and kit pages.

Colour-vision deficiency

• Simulation: Machado 2009 for protan/deutan in linear RGB, Brettel 1997 for tritan (review). Models describe an average dichromat; treat as a screen. [limited]
• About 8% of men of European descent have red–green deficiency (Birch 2012). [strong]
• Read simulated ΔE2000 between meaning-bearing pairs: ≥ 10 distinct, 6–10 risky, < 6 merges (tool defaults, [practice]). Fix by lightness difference plus icon or label.
• Brand colours meant to be told apart, including logo parts: ΔL* ≥ ~15–20 (also protects one-colour print). [practice]

Fix order

1. Move the failing role along the same scale (darker in light mode, lighter in dark).
2. Flip the partner (best_text_on); if neither white nor black passes (mid-lightness brand colours), map the UI role to a darker step and keep the brand hex for the logo and large areas.
3. Never edit a locked brand colour to pass a UI check.
4. Semantic colours may shift hue to clear a brand or CVD clash.
5. Isoluminant or greyscale warnings: change lightness, not hue.
6. Re-run the audit after every change; fixes interact. A new hue is a last resort and needs the user.

Dark mode

Every gate applies again. Leave margin above 4.5:1 for body text; avoid pure #000 with long pure #fff text; a white onPrimary copied from light mode is the classic failure. Light mode reads better for long text (Buchner et al. 2007) [strong], so dark is an option, not the only theme.

Reporting

Gates first with measured ratio and threshold ("text-muted on surface-alt 4.21:1, needs 4.5:1 → step 600, 5.02:1"). Quote audit numbers verbatim. Passing contrast is a floor, not proof the palette is good.