Confidence belongs next to the number

There is a particular failure mode in civic analytics that is easy to miss because it looks like success. A team builds a map of a district, colours every ward by satisfaction score, and puts it on a wall. The map is beautiful. It is also, in three of those wards, based on eleven responses.

Nothing on the map says so. The eleven-response ward is the same shade of confident green as the ward with four thousand. A district officer looks at it, sees a problem area and a healthy area, and allocates attention accordingly. The allocation is wrong, and the map is the reason.

Resolution implies certainty, whether you meant it or not

Visual precision reads as epistemic precision. A choropleth at ward level makes an implicit claim: we know enough about each of these units to distinguish them from one another. When that claim is false for some units, the interface is lying without anyone deciding to lie.

The usual defence is a footnote — sample sizes available on request, methodology in appendix B. This does not work. The footnote is read by the analyst who wrote it and by nobody else. By the time the map reaches a decision, the caveat has been stripped by every retelling.

What we do instead

Every score the Civiq.plus twin exposes carries a confidence value and its provenance. Confidence is not a single arbitrary number; it is computed from seven weighted contributors:

• Freshness — how recently the underlying signals arrived
• Survey coverage — whether representative sampling exists for this geography
• Field reporting — the volume and distribution of field worker reports
• Community coverage — the breadth of registered community participation
• Media coverage — the density of media signal for this topic and place
• Source diversity — how many independent signal families agree
• Model confidence — the model's own certainty for this classification

The interface treats low confidence as information rather than as an embarrassment. A ward with thin coverage does not get a confident colour and a small asterisk. It gets a visibly different treatment, because the honest reading of that ward is we do not know yet, and that reading should be as easy to see as any other.

This has a second-order benefit that surprised us. When coverage gaps are visible on the main dashboard rather than buried in a data quality report, they get fixed. A ward that has been grey for three weeks becomes a field deployment decision. Confidence scoring turned out to be a coverage-improvement mechanism as much as an honesty mechanism.

The uncomfortable version

There is a real cost to this. A dashboard that admits uncertainty is less satisfying to present than one that does not. Someone will eventually ask why the system cannot just give a number, and the answer — because the number would be made up — is not what they wanted.

We think that trade is obviously correct, but it is worth naming rather than pretending it does not exist. A platform that tells you what it does not know is harder to sell and better to rely on.