Two more tools, and why the palette is not our taste

Correlation explorer and regression lab are live.

The correlation one makes you click through to the scatter, deliberately. A coefficient cannot tell you about a curve, a cluster, or one influential point, and Anscombe's quartet is sixty years old and still the best argument against reporting r alone. Pearson and Spearman sit side by side so a monotone-but-curved relationship is visible without opening anything.

The regression lab puts the residual plot directly under the fit, at the same width, on the same x-scale. Most tools bury residuals behind a tab. That placement is the entire opinion of the tool: you look at the residuals before you look at r², because a high r² with structured residuals is a warning rather than a result.

On the palette. It was not chosen aesthetically. We ran a validator that simulates the common colour-vision deficiencies and measures perceptual distance between every pair of colours that might appear together, against both the light and dark backgrounds. Candidate orderings were enumerated and only ones clearing every gate were kept.

Two things fell out of that which I would not have guessed:

Slot order matters and is not cosmetic. In a bar or line chart only adjacent colours must separate, so the sequence of hues is a safety mechanism.

Scatter is much stricter than bars. Every pair of colours ends up adjacent somewhere in a point cloud, so all pairs must clear the bar — and no eight-colour palette does. Hence the three-category cap on scatter, which looks arbitrary in the interface and is not.

I have decent eyes for colour and I would have shipped something worse. The lesson generalises: if a property is computable, compute it.