Historical May 2026 prototype; interval publication held. This article records an ACI implementation and replay. It does not claim that current forecasts ship conformal intervals or meet 90% coverage.
Why ACI was explored
The May design used Adaptive Conformal Inference to adjust an interval
threshold as observed outcomes arrived. The proposed update was
alpha_next = alpha_now + gamma * (target_miscoverage - miss).
A miss lowers alpha and widens the next interval; a covered outcome
raises alpha and can narrow it. This is a method description, not a
validation result for Voidly. See
Gibbs and Candes (2021).
May replay
The original writeup reported a replay of 840 alert-outcome pairs, alpha 0.21, raw quantile 0.96, and 91.3% empirical coverage. Those historical figures have not been revalidated under the current eligible forecast-outcome cohort. The May implementation also capped the applied half-width at 0.35, so its own published intervals could be narrower than the raw ACI interval. The replay is not proof of present coverage.
Publication gate
Current public 7-day forecast responses withhold interval_90.
ACI update and health fields must not be presented as an operating
calibration guarantee. Restore interval publication only after a
settled, eligible outcome cohort shows the requested coverage and the
cap's effect is disclosed.