Bands are less over-read than continuous numbers, at the cost of hiding within-band differences

What it is

Low / medium / high forces the decision to stop at three thresholds. It is hard to invent a story between 61 and 67. A continuous bar and two decimal places invite exactly that story. Bands spend resolution to buy less over-reading. The cost is that true differences inside a band are flattened — 61 and 89 both called medium look like the same point, even if they needed different acts.

This is a choice of scale, not whether to show.

Why it happens

A continuous scale offers sortability. Sortability is a hole when the score is unreliable: people treat the order of the noise as the order of the information. Bands lock order to a few boundaries and forbid comparison beyond them. Weather and clinical communication use short grades to confiscate that false comparison.

After the confiscation, acts can only flip at a band. If the flip the real decision needs sits inside a band, bands go dull. Three equal-width cuts, with calibration error largest in the high end, will still pack the high band with points that should not be there — banding does not replace calibration; it only limits reading.

Studying it

The same items as continuous % versus three bands versus five. Watch: false comparisons between neighbours (choice still changes when the gap is smaller than calibration error), and misses where two points inside a band should have different acts and do not. Independent variables: how bands are cut (equal width / by calibration error / by action threshold), band names (numeric ranges vs. words). Dependent variables: over-resolution, missed resolution, time to decide.

Cuts must follow acts. “Banding failed” on aesthetically equal slices is often a bad cut, not a bad idea.

Where it stops holding

An expert monitoring wall that needs to see drift wants a continuous scale as an instrument, not as decoration — that is a professional meter, not a caption on one generation. Two bands (pass / fail) are often right on a high-stakes gate and dull on an exploratory task. Names like “87–100” smuggle the percent sign in through the back. This entry does not treat a whole batch tagged high and collapsing as a signal; that is the next card.

Applying it

• Default to three bands, cut where the act changes: must be reviewed by a person / spot-check / usable as-is, not at 33 and 66.
• Inside a band, forbid a finer bar or a second decimal. The gain of banding is that further comparison is not allowed.
• Say that the inside of a band is flattened: medium does not mean equally sure, only that the act is the same.
• Check: take two same-band samples that, on calibration, should not share an act, and see whether the UI forces different acts. If it does, the cut is wrong. Then take two continuous values three points apart and see whether people invent different stories — if they do, go back to bands.

Related

• Same group: L1.08.1 Confidence is the model’s self-report; the user has no independent way to verify it · L1.08.2 Percents are read as frequency promises, and most model numbers are uncalibrated · L1.08.4 Showing uniformly high confidence on a whole batch provides no discrimination · L1.08.5 Confidence is worth showing only when the user can change the next action because of it
• Nearby: L1.03 Visualizing uncertainty · L1.04 Presenting confidence · L5.05 The moderation principle of transparency
• Search terms: confidence bands · over-reading continuous scores · within-band collapse