Plot

Measured temperatures with their standard deviations against a modelled decay, which is the commonest figure in an engineering report and the one this platform could not previously draw without a browser. The intervals name themselves as SD rather than leaving a bare ±, so a reader can tell them from a confidence interval by reading rather than by asking.

A four-parameter logistic drawn over the observed responses on a logarithmic concentration axis, with the observations carrying 95% confidence intervals that say so. Shows the two things a dose–response figure is usually wrong about: an unlabelled uncertainty, and a log axis whose base is never stated.

Two years of regional revenue as grouped bars on a baseline of zero. The zero baseline is not a stylistic choice: a bar encodes quantity by length, so a cut baseline turns a three per cent difference into what looks like a threefold one, and this is the figure people cut most often.

The same three quarters drawn twice is impossible in one panel, so this is the cut version alone — and the engine's own warning about it is part of the figure. A deliberate teaching artefact: open it to see what a truncated baseline does to a nine per cent spread, with the sentence explaining it printed underneath rather than left for a reviewer to say.

Two distributions of request latency as Tukey boxes with their outliers kept and their means marked. The mean is drawn separately because a box shows the MEDIAN, and a reader who assumes it is the mean misreads every skewed group — which for a latency distribution is every group.

A size distribution binned by Freedman–Diaconis, with the rule stated in the figure. Bin width is the single choice that decides what a histogram appears to say — the same data at two widths shows one mode or three — so a histogram that does not name its rule is a figure whose main claim cannot be checked.