experiment EX-VE-COL-006
CVD Minimax Palette and Semantic Survivability
EX-VE-COL-006: CVD Minimax Palette and Semantic Survivability
Question
Can a palette optimized for worst-case category separation across multiple observer models reduce categorical confusion while semantic redundancy preserves meaning when color fails?
Phase A: computational minimax search
Define candidate categorical colors in a bounded gamut.
For each candidate palette:
- evaluate normal-vision perceptual separation;
- transform through multiple severity levels of protan, deutan, and tritan simulation;
- calculate pairwise separability in a versioned perceptual space;
- enforce WCAG contrast requirements for the actual foreground/background use;
- penalize semantic-token collisions;
- maximize the worst-case separation rather than the average.
Illustrative objective:
maximize
min(observer_model, severity, color_pair)
discriminability(color_pair)
subject to
conformance constraints
gamut constraints
semantic-role constraints
No simulated score is labeled accessible.
Phase B: semantic channel dropout
For states such as success, warning, failure, selected, disabled, and unknown, create conditions that selectively remove or degrade channels:
- hue removed / grayscale;
- protan simulation;
- deutan simulation;
- tritan simulation;
- forced colors;
- icon removed;
- text label removed;
- border or pattern removed.
Measure whether participants can still identify the state.
Redundant channels
Candidate channels:
- text label;
- icon geometry;
- border/shape;
- position;
- pattern or line style;
- color;
- programmatic semantics.
The aim is not maximum cue count. The aim is the smallest independent set that maintains performance without excessive clutter.
Human validation
Recruit affected participants across CVD types and severity where feasible. Compare:
- current palette;
- minimax palette;
- minimax palette plus redundant semantics.
Measure accuracy, time, confidence, and confusion matrix.
Falsification
HY-VE-PE-007 is weakened if the optimized palette does not improve held-out category discrimination or if gains exist only in simulation.
HY-VE-PE-004 is weakened if semantic channel survivability scores do not predict actual state-identification robustness.
Required output
- palette-generation algorithm and version;
- all simulation model versions;
- category-confusion matrices;
- state-recognition results;
- explicit list of conditions where color remains insufficient;
- no universal "CVD-safe palette" claim unless affected-user evidence supports it.