The MeasureIssue 14Subscribe

Essay · Tools

The number nobody measured

Design tools have learned to sound certain. The harder discipline is saying which of your numbers you would actually stand behind.

Words

Ester Halvorsen

Published

28 August 2026

Reading

11 minutes

There is a particular kind of confidence that comes from a decimal point. A tool that tells you your spacing scores 0.78 has told you something exact-sounding about your work, and you will believe it, because the number has a tenths place and nothing that precise is usually wrong.1 The trouble starts when you ask what the 0.78 was computed against.

In the case of most screenshot-to-design tools, the answer is: a language model’s description of a picture. The model looks at your reference, says the sidebar is about 250 pixels wide, and the score is the distance between that and what your page rendered. One side of the subtraction is measured. The other is a guess wearing the same typeface as the measurement.2

“The sidebar was 240 pixels in every run. The model said 250 twice, and did not mention it twice.”
From the appendix: four runs of one vision model against a single 1440×900 capture.
An open book, pages fanned, on a desk
The other half of the argument: a number you cannot check is no better than a page you cannot read.

Two ways to be wrong

The failure is not that the estimate is bad. For what it is asked to do, it is remarkably good: it identifies a layout, a metric grid, a chart and a table correctly, every time, and it reads the labels off them verbatim. Semantics are reliable. It is geometry that is not, and geometry is the half a score is built from.

So the useful move is not a better model. It is refusing to score the dimensions whose target side is an estimate, and saying so on the page, by name, with the reason attached.3 An audit that reports four numbers and three refusals is more informative than one that reports seven numbers, three of which are decorated noise.

It is also harder to sell, which is probably why nobody does it.

Notes

  1. 1Halvorsen, E. “On measuring what you have already drawn.” Journal of Applied Typography 14 (2019): 22-41.
  2. 2The four runs are described in the appendix. The sidebar was 240px in every one; the model said 250px twice and did not mention it twice.
  3. 3A per-pixel score against an estimated target is not a measurement of the page. It is a measurement of the estimate.