Historical May 2026 calibration audit; reported gains unvalidated. This account describes a label and calibrator repair attempted in May. Its reported figures are not present forecast accuracy.
The two problems identified
The May audit identified a stale isotonic mapping fitted to labels that counted IODA connectivity disruptions as censorship, and an outcome join that could also count disruption-only incidents as positives. Both can make a forecast appear better or worse than it is for the censorship target.
The May refit
A script rebuilt censorship-only outcomes from the historical archive, fit a new isotonic mapping, and compared Brier and expected calibration error on the same 30-day sample. The writeup reported 720 pairs and an in-sample Brier change from 0.380 to 0.120. It also reported drift moving from +56.45pp to 0.00pp across different label views. Those values are historical, not a forward test, and the cohort has not been revalidated under current eligibility rules.
What would validate a current claim
Rebuild settled, censorship-only outcomes from a fresh eligible cohort, compare the old and new mappings on a forward holdout, and report country coverage, base rate, Brier, ECE, and missingness. An in-sample zero drift is expected from fitting and is not proof that the live forecast is calibrated. Current public intervals remain withheld pending their own coverage check.