Voidly
All findingsPublished research

May 2026 Sentinel calibration refit report (historical, unvalidated)

A May 2026 refit writeup reported 810 forecast-outcome pairs and an in-sample Brier change from 0.59 to 0.22. The cohort and calibration effect have not been revalidated under the current eligibility rules. These numbers are historical, not evidence of present forecast accuracy.

Historical May 2026 refit report; performance unvalidated. This writeup described a calibration problem and an isotonic refit. The archived numbers are not a current accuracy or coverage claim.

What the May audit reported

The May draft reported low forecast probabilities in bins with much higher observed positive rates. It then described an isotonic mapping fit to 810 forecast-outcome pairs and reported an in-sample Brier change from 0.59 to 0.22. Those pairs and their outcome eligibility have not been revalidated under the current rules. The in-sample improvement cannot establish forward performance.

Method and limitation

Isotonic regression fits a monotonic probability mapping after a base model. The reported May refit applied that mapping to a watched-country set. An outcome join that includes disruption-only incidents would distort both the fit and its evaluation. A defensible new result needs settled censorship-only labels, a held-out forward cohort, and clear country coverage before any calibration number is republished.

Current public 7-day forecasts withhold interval_90 while interval coverage remains unverified. The May refit figures do not justify restoring an interval or a live calibration time series.

The source record.

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