Voidly
All findingsPublished research

May 2026 per-country calibration drift monitor (historical, unvalidated)

A May 2026 design computed per-country predicted-versus-observed drift over 30-day windows and proposed a +/-15pp alert. The published drift values and alert behavior require revalidation against an eligible outcome cohort. This historical account is not evidence that daily per-country calibration…

Read the full published summary

A May 2026 design computed per-country predicted-versus-observed drift over 30-day windows and proposed a +/-15pp alert. The published drift values and alert behavior require revalidation against an eligible outcome cohort. This historical account is not evidence that daily per-country calibration is currently measured or that the alert is active.

Historical May 2026 monitor design; current drift unvalidated. The May account proposed comparing forecast probabilities with settled censorship-only outcomes by country over a trailing 30-day window.

Proposed rule

For countries with at least 15 eligible predictions, the script would compute mean predicted probability, empirical positive rate, Brier score, and their difference in percentage points. A +/-15pp threshold was proposed for a CenAlerts event with 24-hour deduplication.

Evidence required for use

The historical writeup does not establish that this cron, sidecar, endpoint, or alert is currently operating with eligible outcomes. No May drift value should be described as current. Before citing a country-level estimate, publish the scoring date, settled outcome count, label rule, coverage, and alert delivery evidence.

Small country samples and shifting measurement coverage can make observed drift noisy. Drift alone does not prove that the model is broken.

The source record.

Original source links and labels. Linked responses may have changed since publication.