RECONSTRUCTING ONE RULE
How does a color
become a category?
Three colors. Three lightness values. One contrast decision.
This model only classifies lightness contrast. It does not identify a season or recommend what looks good on someone.
01 Represent the inputs
Choose colors or enter HEX values. The labels are assigned by you; the model does not detect a face.
Example A loaded. Change a color to experiment.
02 Make the rule visible
The category boundaries
ΔL* < 2525 ≤ ΔL* < 50ΔL* ≥ 50These are designer-chosen experimental thresholds, not validated personal color analysis standards. Classification uses unrounded values.
CONTROLLED EXPERIMENT
Three inputs, one changing variable.
Only the value labeled “skin” changes. Hair and eye inputs remain fixed. These are synthetic color tests, not photographs of people or simulations of particular lighting conditions.
| Test | Skin / hair / eyes | Skin L* | Hair L* | Eyes L* | ΔL* | Result | Replay |
|---|
WHAT THE RECONSTRUCTION REVEALS
Five observations
- The system accepts values, not people. Grayscale inputs labeled “skin” still produce a category. It never verifies that they represent skin.
- A single input can change the judgment. With hair and eyes fixed, changing the skin input moves the result from Low to Medium to High.
- The middle value disappears from the decision. Eyes are between the darkest and lightest values in all three runs. Only the extremes determine the range.
- Categories are a choice added to measurement. The numerical difference is continuous; the chosen boundaries at 25 and 50 divide it into three names.
- The result loses context. The calculation cannot tell whether an input changed because of lighting, selection, or appearance. Two decimal places do not establish reliable knowledge about a person.
Method, limitations & credits
Each six-digit HEX value is treated as sRGB. RGB channels are linearized, then converted to relative luminance Y using D65 coefficients. CIELAB L* is calculated from Y. Contrast is defined here as max(L*) − min(L*), not WCAG text contrast. Hue, chroma, facial area, personal taste, cultural context, and camera calibration are excluded.
Concept: the author’s personal color analysis system map. Implementation and synthetic test execution: OpenAI Codex. Built with HTML, CSS, JavaScript, and browser color inputs; no external analysis library. Hosting: OpenAI Sites. Classification thresholds were chosen for this experiment.