EX-ITTEN-008 — Color-difference metric validation

Question and falsifier

Which metric predicts task-specific discrimination across modern wide-gamut and HDR displays? A universal-metric claim fails if rankings or calibration vary materially by regime, task, observer, or device.

Work completed

Three thousand fixed-seed random sRGB pairs were compared with Euclidean Lab and Oklab before and after a simple gamut-compression transform. Rank stability and relation to luminance difference were recorded.

Pilot results

Rank stability after the mild compression was .9990 for Lab and .9987 for Oklab. Cross-metric rank correlation was only .7824. Correlation with luminance difference was .3675 for Lab and .7417 for Oklab in this generated corpus.

The stress transform was too mild to separate within-metric robustness, but the cross-metric disagreement justifies empirical comparison. There is no perceptual ground truth here, and ΔE00, CAM16-UCS, Jzazbz, and HDR viewing conditions remain to be implemented.

Validation experiment ready to run

• Build a stratified pair corpus spanning hue, lightness, chroma, near-threshold and suprathreshold distances, gamut edges, skin/medical/status colors, and SDR/HDR.
• Measure actual display spectra and luminance; record adaptation and surround.
• Tasks: forced-choice discrimination, suprathreshold difference scaling, appearance matching, and real interface state identification.
• Metrics: ΔE76 baseline, ΔE00, CAM16-UCS, Oklab, Jzazbz and current refinements, cone-contrast, and luminance contrast.
• Evaluate held-out predictive likelihood, calibration error, rank correlation, and worst-group error rather than selecting by pooled correlation alone.
• Publish raw trials, transformations, calibration files, exclusions, and code.

Current position / continuation

Stage: reproducible baseline stress test complete; metric validation not achieved. Next action: add reference ΔE00/CAM16-UCS/Jzazbz implementations with unit tests, then construct a physically measurable stimulus corpus and preregister validation. Data: ../experiments/data/ex-itten-008.json.