Verification

The catalog above records what themes claim. This section tests it.

How the test works

For every catalogued theme we take its signal colors — the small set that carries meaning (success/error/warning, diff add/delete, syntax roles) — and run two computations:

1. Simulate each color under each dichromacy (deuteranopia, protanopia, tritanopia) using the Machado, Oliveira & Fernandes (2009) physiologically-based model at 100% severity — the same model behind Coblis and most modern simulators. This produces what each color actually looks like to a viewer with that deficiency.
2. Measure the closest pair among those simulated colors using CIEDE2000 (ΔE00), the standard perceptual color-distance metric. The minimum pairwise distance is the theme’s worst case: the two signals most likely to be confused.

How to read the numbers

ΔE00 is roughly: ~1 = a just-noticeable difference under ideal lab conditions; <3 = two colors most people will struggle to tell apart on a real screen; >8 = clearly different at a glance. A theme is only as strong as its weakest pair.

Verdict scale (a heuristic, not pass/fail):

• STRONG — closest pair ≥ 8 ΔE00 under all three dichromacies. Robust for any viewer.
• ok — closest pair ≥ 3. Usable, but the closest pair is a known tight spot.
• WEAK — closest pair < 3 under at least one deficiency. Some pair collapses on hue+lightness.

Reproduce every number below: python3 scripts/verify_colorblind_themes.py (stdlib only, deterministic — it reads the same hex tables that generate the examples).

See the difference, don’t just read it. Each plate below shows a STRONG theme against a WEAK theme under one simulated dichromacy — Normal row, then the CVD view. The WEAK pair’s two chips collapse to near-identical tones; the STRONG pair stays apart.

Under deuteranopia, GitHub’s keyword #ec8e2c and warning #d29922 become #C1AD2D and #BDA928 — 7.5 RGB units apart, effectively one color. GitButler’s addition #99d9ff and deletion #ffbf66 stay 156 units apart.

Read this table honestly — three caveats:

1. A WEAK verdict is not “broken.” Most of these themes deliberately add a luminance gap between hues that share an axis (GitHub’s keyword orange #ec8e2c vs. warning amber #d29922 collapse to ΔE≈1.1 in hue under deuteranopia, but they’re never used as a binary good/bad signal). The metric measures hue+lightness collapse; it does not model font weight, icons, or the redundant text cues these UIs also provide.
2. Single-deficiency themes collapse by design under the others. modus-deuteranopia scoring WEAK under deuteranopia looks alarming until you see the weakest pair is red #ff5f59 vs green #44bc44 — two colors the deuteranopia variant deliberately stops using as a semantic pair (errors route to a different channel). Judge a theme against the deficiency it targets.
3. The strongest independent signal is the engineering gate, not the hue choice. Every STRONG entry (Kiro, Wireshark, Cindervale, Opo, Blinds) ships a test that measures exactly this and fails the build on regression — which is why they top the table. Themes that only assert CVD safety cluster at the bottom.

The full per-theme numbers (all three deficiencies, weakest pair identified) are the script’s stdout and are reproducible.

The one-line takeaway: the themes that measure survive; the themes that claim often don’t. If you’re choosing a theme for a CVD user, prefer one with a simulation/contrast build gate.

App/platform-level themes verify far better than editor themes

Running the same metric over the app-level signal palettes (the small good/bad/status sets each product shows) gives a markedly different picture: 11 STRONG, 3 ok, 5 WEAK — versus the editor/terminal tier’s 8/17/17. Apps only need to separate 2–4 status colors, so they can space them far apart; editors must cram 6–8 syntax roles into one palette.

Two findings worth naming:

• Per-deficiency palettes confirmed empirically. Talishar ships one palette per dichromacy, and the numbers show exactly why that’s the right call: each palette is strong under its target deficiency (deutan palette ΔE 17.3 under deuteranopia; protan ΔE 20.2 under protanopia) but collapses under the others (tritan palette: 2.6 under deuteranopia). Per-deficiency variants beat one “universal” palette — matching what Grafana (4 themes) and modus/ef do.
• The GitHub orange/amber weakness propagates downstream. Gitea and SimpleCov inherit GitHub’s colorblind pair (#f0883e orange vs #d29922/#9a6700 amber), and both score WEAK for the same reason the flagship does. SimpleCov’s dark mode dodges it (its missed-orange is far from its skipped-yellow); its light mode doesn’t. When you port a palette, re-measure it.

GitButler is the strongest CVD palette in this entire report — its diff addition/deletion text pair (#99d9ff/#ffbf66 dark, #0d59bf/#bf590d light) sits ΔE 42–61 apart under every dichromacy, roughly 5× the “clearly different” threshold. That’s what designing the two-color case from scratch, rather than remapping an existing 8-color palette, buys you.