A Span Chart (also Range Bar, Floating Bar, High-Low Graph) draws a bar that floats between a minimum and maximum value for each category. Unlike a bar chart, there is no zero baseline — the left edge of each bar is the minimum, and the right edge is the maximum. Both values, their difference (the span), and the bar's position on the axis carry information simultaneously. The viewer can read: where does this range sit? (bar position) and how wide is the range? (bar length). A reference line cuts through all bars, allowing the reader to see which ranges straddle a threshold and which sit entirely above or below it.
The data describes performance variation across AI models — a story that requires showing both extremes and their relationship to a deployment threshold. A standard bar chart can only show one value per category (the max, or the mean) and would conceal the worst-case performance entirely. A dot plot with two dots per category would work but loses the immediate visual reading of span width. The span chart makes both the level of performance and the consistency of performance readable simultaneously — which is exactly the analytical question a deployment decision requires.
Span charts show only the two extreme values. They give no information about the distribution between min and max: whether most models cluster near the top, near the bottom, or scatter evenly. A wide span could mean all models are mediocre except one outlier at each end, or it could mean models genuinely vary across the full range. Without that context, the span chart can mislead — a single outlier at the minimum drags the bar left without representing typical performance. When distribution matters, a Box & Whisker Plot is the correct upgrade: it adds the median and quartiles while preserving the min-max span.
FT Visual Vocabulary: Distribution — Range. Abela quadrant: Comparison — comparing ranges across categories. Tufte: the floating bar is honest about its baseline absence — it would be wrong to extend bars to zero, which would imply that zero is the lower bound when the data's minimum is 33.
The one decision worth knowing: bars animate from the midpoint outward, not from the left edge rightward. This is not cosmetic. Animating from zero or the left would imply a start-to-end directionality the data does not have. Expanding symmetrically from the centre makes the visual argument: uncertainty (range) grows around a central estimate — which is the correct mental model for a min-max range.