Chapter 6 — Practice-on-your-own solutions

The failure. Under deuteranopia (the most common red-green CVD), red and green cells collapse to similar olive-brown tones and become almost indistinguishable; yellow reads as a slightly lighter olive against the same background. For roughly 8% of male viewers (and 0.5% of female viewers), a green-vs-red cell distinction disappears entirely and the dashboard reports one undifferentiated status across all 40 categories. Because the analyst’s own vision usually is not affected, this failure ships silently — the analyst inspects the chart, sees red-vs-green clearly, and cannot see what the audience will not see.

The rewrite. Add redundant encoding so the meaning is carried by at least two channels, one of which is not colour:

• Position or shape as the primary carrier: sort the 40 categories by status (all failing categories in one contiguous block, all warning in the next, all passing in the third), separated by a thin divider row; the reader scans the divider positions to see the counts without needing to decode any colour.
• Direct glyphs in each cell as a redundant carrier: a white check for pass, a hollow circle for warning, a cross for fail, sized to be legible at dashboard resolution.
• Colour retained as the tertiary carrier: keep the red/yellow/green (or better, red/amber/blue — blue survives deuteranopia intact), but treat it as reinforcement, not as the load-bearing signal.

The resulting dashboard passes an at-a-glance colour scan (the discipline of the «40 categories at once» view), passes the CVD simulator, and passes a black-and-white print. All three surfaces matter for a support-ops dashboard shipped to a mixed audience on unpredictable rendering hardware.

Meta-observation. The rewrite is not «stop using red and green.» The rewrite is «stop making red-vs-green the only difference between good and bad.» Chapter 6’s accessibility rule is about redundant encoding, not colour avoidance — the fix is additive.

The palette-type mismatch. The data is divergent: it has a meaningful midpoint at zero (on-plan) with values that mean qualitatively opposite things on either side (positive variance = beating plan; negative variance = missing plan). A sequential palette from light-to-dark blue treats the range as a single ordered scale, so the eye reads +5% and −5% as adjacent shades of pale blue and cannot distinguish «just above plan» from «just below plan.» The pale end of the palette maps to whichever extreme happens to be smallest in absolute value, obscuring the sign entirely. This is the «wrong palette type for the data’s shape» failure Chapter 6 warns against.

The one-sentence Refine direction.

«Switch this heatmap to a diverging palette anchored at zero — use ColorBrewer RdBu reversed so that below-plan cells render red, on-plan cells render neutral white or light gray, and above-plan cells render blue; the palette must be perceptually symmetric around the midpoint even though the data range is asymmetric (−30% to +40%).»

Colour at zero: neutral white or the lightest gray in the palette. The midpoint must be visually distinct from both hues at low intensity so the reader can locate the on-plan cells at a glance. Red at zero would imply mildly negative; blue at zero would imply mildly positive; neutral is the only colour that carries the meaning «on plan.» Also: because the data range is asymmetric (extends further above zero than below), the analyst should lock the palette midpoint at zero rather than at the midpoint of the data range; otherwise a modest below-plan cell will render darker red than a much larger above-plan cell renders blue, and the visual weight will misrepresent the sign.

Meta-observation. The palette-type decision is upstream of the specific palette. «Which shade of blue?» is the wrong question when the data has a midpoint; «which two-hue diverging palette anchored at zero?» is the right question. Palette-type-to-data-shape is the mapping analysts most consistently get wrong when they accept Copilot’s default without naming the data’s structure in the prompt.