Every claim on this page traces back to a real API response. Where we made a judgment call instead of measuring, it's named below.
Eight people, measured — Fitzpatrick I, III, IV and VI. Fitzpatrick II and V are not measurable with this panel — not covered, not claimed as covered, just not in this eight. We prompted for all six bands when generating the panel; the real measurement landed on four. That gap is disclosed here, not smoothed over.
Eight people are a sample, not a population. This product measures whether a colorway separates from these eight measured skin tones, not what any population of customers looks like.
All eight reference images are AI-generated (Gemini 3.1 Flash Image) — fictional people, not real individuals, not photographs of anyone. We chose this over sourcing real photos to avoid a likeness/consent problem entirely rather than manage one.
The lead metric is ΔL* — the CIELAB luminance difference between the garment color and a panel member's measured skin tone. CIEDE2000 (ΔE) is shown alongside as a secondary, supporting number; the "loses contrast" claim is driven by ΔL*, not ΔE, because luminance separation is what the eye reads as an edge between garment and skin — hue or chroma difference alone doesn't guarantee that.
We call anything below ΔL* 15 low contrast. We call a pairing "low luminance separation" below ΔL* 15 — a threshold we set for this product, not a cited external standard. Below it, in our own viewing of the rendered images, the garment silhouette starts reading as a value-match to the skin rather than a distinct shape. We are not citing an external standard (WCAG, ISO, or otherwise) — this is our own operational cutoff for this product, and we're not asserting it as anything else.
Sampled by averaging a center crop of the clean, unworn product photo — not the composite try-on images, where background and each person's own lighting would bias the reading.