Colour

What to decide for a set's palette; palette_build.py and palette_audit.py do the arithmetic. Tags: [strong] [limited] [practice] [popular-unsupported].

Strategy

• Differ first, fit second. Colour is the weakest distinctive asset: 12% fame, 39% uniqueness on average vs 40% / 71% for shapes (Phua et al. 2026) and the most shared (Ward et al. 2020). [strong] Ownership = shape + colour + years.
• A decision is hue + chroma + lightness + ground. Lightness and saturation carry their own effects (Labrecque et al. 2024: saturation reads as power, backfires for gentle products) [strong]; a dark ground pulls a hue toward its negative meanings (Celhay et al. 2024) [strong], so choose the ground as strategy.
• Associations are vocabulary, not rationale: reported, not shown to cause feelings (Jonauskaite et al. 2025). [strong] No behaviour claims ("red converts", "colour lifts recognition 80%"). [popular-unsupported]
• Conform or differ: strong brands gain from looking like themselves, weak ones from the segment (Heitmann et al. 2020) [strong]; atypical wins with exposure (Landwehr et al. 2013) [strong]; in high-risk categories (health, money, law, children) stay in the family and differ on lightness/chroma or a small accent (Garaus & Halkias 2019). [limited] Moderate deviation beats extreme.
• Competitor map: search_library.py --industry X --summary (measured public sites), plus brand.py site competitors for named rivals (≤ 8). Plot primaries in OKLCH; empty hue sector or lightness band = room.
• Near colour = same colour: reject a primary within ΔE2000 < 10 of a direct competitor's, or same hue sector (±20°) and ±0.08 L while both are chromatic; 10–20 is borderline. Courts treat close shades as one (BGH, Milka lilac, 2004); thresholds [practice].
• Local meaning: search each market for party, flag, team or mourning readings of the main pair: a warning; a national flag as emblem ground is out (Paris 6ter).
• Refresh: equity is shown, not declared; marketers overrate their own assets (Brus et al. 2025). [limited] Recognised and unshared → keep the hue, tune; shared or unknown → free to move. Spread sets tune → evolve → shift.

Architecture (palette.json)

• Two layers: brand[] (identity colours, locked once approved, never edited to pass a UI check) and modes.light|dark roles (16 keys) pointing at 11-step OKLCH scales; the anchor step holds the brand hex exactly.
• 1 signature, 0–1 secondary, 1 accent, a tinted neutral, 4 semantic. 2–3 hues total: one hue and more than three both rate lower (O'Donovan et al. 2011). [strong]
• Neutrals tinted toward the brand hue (C ≤ 0.045); C 0.05–0.09 reads as neither grey nor colour. [practice]
• The light ground is a decision (--light-bg): white reads clinical, cream/stone warm (see ai-defaults.md). Dark mode is derived: tinted near-black, brand fills lightened along their scale, kept small. [practice]
• Extended colours (≤ 5, --extra) for variants and charts, never core roles. 60-30-10 has no basis. [popular-unsupported] The loudest colour takes the least area. [practice]

Pairing

• Harmony is lightness structure, chroma balance and hue count, not wheel templates. One very light, one very dark, an orderly ladder; surfaces from one hue family separated by lightness; the accent differs in hue and stays small (O'Donovan 2011; Schloss & Palmer 2011). [strong] Complementary and triadic schemes are not rated harmonious; never justify a pick by "it's the complement".
• Well-rated directions: warm + cyan, orange + navy, green + blue/yellow, light warm figure on dark cool ground. [strong]
• Hue physics: yellow, lime and teal are vivid only when light (need a dark partner); blue and violet only when dark; red and magenta peak mid-lightness and carry neither white nor black text well. [strong] (arithmetic)
• Avoid isoluminant pairs (both saturated, ΔL* < 12: they shimmer; Wolfe & Owens 1981) and red–green as the only signal (Birch 2012). [strong] Fix with lightness, not hue.
• Judge on the rendered card or site. The audit finds what is broken; taste decides what is good.