Box plot

Median, quartiles, and whiskers.

import { BoxPlot } from "diagrammatic";

Four models, twenty-eight runs each. The box is the summary; the dots are the argument. Small multiples would be a second job. This figure is the call count that decided the cost conversation.

Illustrative data. FrontierBench tool-use split. 28 runs per model.

fig. 29 · Tool calls per eval run, every point kept

swift's median is 298 and sits above the 200-call cost line; its tail clears 500. atlas-1 and nova have medians of 90 and 100, both under the line. quill sits in between and still has a 596 run. The dots make the tails countable runs.

Use it for: Comparing distributions across groups on one scale.
Watch out: It hides multimodality; a two-humped group looks like a calm box.