How should a dashboard show the uncertainty of a metric so that operators react to signal rather than noise?

Open question

The SRE book asks that dashboards answer basic questions about a service and that effective alerting systems have good signal and very low noise. Dashboards, however, draw every number with the same visual certainty. An error rate of 33% computed from three requests in the last minute looks like one computed from three million. The Prometheus documentation has a section on the errors of quantile estimation: a quantile read from a histogram is interpolated within a bucket, so with a bucket from 200 ms to 300 ms and most requests near 220 ms the estimate can land at 295 ms, and the error depends on the bucket boundaries around the true value. A p99 line drawn from such buckets carries an uncertainty that the chart never shows. The general statistical tools (interval bands, standard errors, sample counts) exist; what is unclear is which presentation changes operator behaviour for the better.

Sub-questions:

• Does showing the sample count next to a ratio (as a label, as line thickness, or by greying out low-count points) reduce reactions to low-traffic spikes, and at what count threshold have teams set the cut-off?
• Do interval bands around a percentile line help, or do operators read the band's upper edge as the value and page more often?
• For quantiles estimated from histograms, has anyone displayed the bucket-width error (the interval the true value is guaranteed to be in) and what did that change?
• Which smoothing windows and rate intervals have teams settled on for low-traffic services, and how do they mark the smoothing on the chart?
• Have teams measured false-alarm and missed-incident rates before and after such a change, rather than reporting preference?

What a useful answer contains

The service's traffic level, the metric and how it is computed (raw counter ratio, histogram quantile, summary), the presentation used before and after, how long each ran, the counts of pages or investigations that turned out to be noise in each period, and whether the operators changed in between. Comparisons on the same service over comparable periods are more useful than descriptions of a single dashboard; vendor defaults restated as findings should say so.