The expression must match the granularity of the user’s decision

What it is

A meteorologist wants the whole density. A driver wants “should I detour.” Dumping the density on the driver, or “might rain” on someone dispatching a reservoir, are both grain mismatches. Uncertainty at decision granularity asks that the resolution of the display just support the next action — not more, not less.

This entry assumes uncertainty is going to be shown. It is about the shape of that showing.

Why it happens

A decision is a threshold operation: above a line, act; otherwise don’t, or take the other branch. Information finer than the threshold cannot change the action; it only spends reading. Information coarser than the threshold leaves people stranded on both sides of the line, filling resolution by guesswork. Studies of uncertainty visualization keep meeting this: the data resolution of a graphic and the decision resolution of a task are not the same thing. Alignment is to the task, not to the data.

Generate surfaces commonly miss both ways. Token-level entropy is drawn for someone who only needs “can this mail go out”; a whole page is tagged “fairly sure” when the user needs to know whether one number is citable. The first is too fine, the second too coarse. Both failures look like “users ignore uncertainty.” They looked; they could not use it.

Studying it

Elicit a real decision first: review or not, cite or not, detour or not. Then offer encodings: one band for the whole text, per-sentence bands, a density, a numeric interval. Dependent variables: whether the action flips with true uncertainty, surplus reading time, a subjective “did this help.” Independent variables must include where the decision threshold sits — move the threshold and the best grain moves.

“Can they read this chart” is not the primary endpoint. Read and still not change the decision is still a mismatch.

Where it stops holding

When one screen has two kinds of decision maker (the person on shift and the later auditor), there is no single grain. Layer it: action-level by default, distribution-level on expand. Exploratory work with no threshold yet is where fine grain is right. A phone can hold one question at a time; protect the action level first. This entry does not discuss whether to show, nor the false precision of extra decimal places.

Applying it

• Write down the user’s next action and the threshold that flips it, then pick an encoding. If the action is “should a human glance,” three bands suffice; if it is “which number may I cite,” mark that number.
• The default view answers the action question. Put the full distribution behind an explicit expand for those who need it, off the default path.
• When only a local span is uncertain, mark the local span. Do not paint the whole page “low confidence” because one figure is shaky.
• Check: move the model’s uncertainty across the threshold you wrote down, and see whether the encoding forces the action to flip. If it will not flip, the grain is not aligned.

Related

• Same group: L1.03.1 Uncertainty must be expressed, not hidden · L1.03.3 Over-precise numbers manufacture false certainty
• Nearby: L1.04 Presenting confidence · L1.08 Confidence displays and their misreadings · L5.05 The moderation principle of transparency
• Search terms: uncertainty at decision granularity · actionable uncertainty · decision threshold