Contrast, measured

Every ratio below was computed from the resolved token values (WCAG 2.1 relative luminance). Thresholds: 4.5:1 for normal text, 3:1 for large text (≥18.66px bold or ≥24px) and for the visual boundary of a control.

Light

Dark

Two known gaps. Do not "fix" them silently

1. White on the accent fill fails. The primary button paints --primary-foreground: #FFFFFF on the brand accent:

The accents are mid-lightness, so white loses on all three and ink wins on all three. Switching --primary-foreground to the ink token fixes every brand at once and changes how every primary button looks — that is a brand-owner decision, not a developer one. It is recorded here with the numbers so the decision can be made rather than inherited. The kit ships the current values so a new app matches the live product.

Where tertiary is allowed, and where it is a bug. #8E8E93 is 3.08:1 on cream — the disabled tier, which WCAG exempts. It is correct for a glyph (a breadcrumb separator, a chevron), for placeholder text, and for a badge on a row that is itself disabled. It is WRONG for anything a reader has to read: a specimen's label, a finding like "Non-metro", a count, a field name. Four of those were caught and lifted on 2026-08-20 — each had been 3.08:1 and is now 5.65:1.

The test is one question: if this text went missing, would the reader lose something? If yes, it is secondary.

2. Tertiary ink is below 4.5:1 in light. #8E8E93 on card is 2.94:1. It is correct for genuinely disabled text and placeholder hints (which WCAG exempts) and wrong for anything a reader has to read. If a label matters, it is secondary, not tertiary.

The one thing the kit did change

--accent-ink — the accent as TEXT — was hard-coded to emerald in the source app for both themes. Correct for iq, and it painted green ink on a Roof or REI app, which breaks the system's central promise. In this kit it is per-brand, darkened in light mode until it clears 4.5:1 against the zebra row (the worst of the three light surfaces): iq #036B4E, rei #036EBC, roof #B84A19. Dark mode keeps the raw accent, which already clears it on titanium.

The guide's own display value (--brand-accent-text) is untouched and still used for large type, where 3:1 applies.

Rules that follow from the numbers

• Never use --brand-accent as text. It fails in all three brands. Use --accent-ink.
• Any new surface must keep body ink ≥4.5:1 and hold the row ladder's floor (≥8.9:1 was the measured worst case across the four row states).
• A control's boundary needs 3:1 against what is behind it — that is why the select chevron is a full-opacity muted token and not opacity-50 on top of one.
• Colour is never the only cue: selection also has an inset bar, an error also has a title and an icon, a soon row also has a badge.
• Recompute rather than eyeball. The script that produced these numbers is twenty lines of luminance arithmetic; the ratios in this file were not estimated.