EX-ITTEN-002 — Contextual color-token robustness

Question and falsifier

Which metric best predicts color discrimination in real component surrounds? Any claim that one isolated metric is sufficient fails if a context-aware or spatial model materially improves held-out human judgments.

Work completed

Twelve interface colors were crossed with six backgrounds (72 pairs). The pipeline computed WCAG luminance contrast, Euclidean CIELAB difference, Euclidean Oklab difference, and an explicitly labeled toy induction stress value (8% displacement away from the surround in Lab).

Pilot results

Rank correlation was .6357 for WCAG versus Lab, .9398 for WCAG versus Oklab, and .6791 for Lab versus Oklab. Of 72 pairs, 31 passed 4.5:1 and 47 passed 3:1. The toy induction shift had median 6.1736 and maximum 9.2009 ΔE76.

The result establishes metric disagreement in this sample, not perceptual superiority. WCAG contrast, Lab distance, and Oklab distance describe different constructs. The induction magnitude is generated, not observed.

Human experiment ready to run

• Render complete controls in six surrounds at fixed angular sizes, including thin strokes, text, fills, disabled states, and alerts.
• Use a calibrated wide-gamut monitor with recorded primaries, transfer function, white, black, and ambient illumination.
• Tasks: same/different discrimination, state identification, visual search, and appearance matching to an adjustable isolated patch.
• Compare ΔE00, CAM16-UCS, Oklab, WCAG contrast, and at least one spatial appearance model against held-out accuracy and matching error.
• Stratify by normal color vision, protan/deutan class, and older adults. Record visual acuity and display viewing angle.
• Primary analysis: cross-validated mixed-effects logistic model; report calibration, not only rank correlation.

Current position / continuation

Stage: code and metric baseline complete; no appearance judgments collected. Next action: add CAM16-UCS/ΔE00 reference implementations and generate calibrated rendered stimuli before preregistration. Data: ../experiments/data/ex-itten-002.json.