How Codex Reset Radar works
The radar separates public messages, completed events and forecasts: a forecast is never presented as a completed reset, and a delayed source is never presented as “no reset today.”
Public states stay simple
The UI only uses understandable states such as reset announced, reset expected, no new reset found, or updates delayed. Editorial and verification fields stay in the data layer.
Forecasts must be explainable
Production probabilities should come from a backtested model using historical reset behavior and public signals, with forecast history retained. When evidence is insufficient, the site should not invent a percentage.
V1 probabilities and release conditions
V1 uses the conditional empirical distribution of historical reset intervals to estimate at least one Global Reset within 24, 48 and 72 hours after the model cutoff. Banked and unknown types are excluded; Tibo source estimates do not add probability. Numerical release requires temporal backtesting and review.
At least 20 qualified completed intervals and 10 comparable intervals beyond the current elapsed time are required. No numbers are shown within one hour of a reset, with insufficient evidence, or after expiry. These are conservative engineering gates, not accuracy guarantees.
How do we avoid future-data leakage?
Daily replay uses only revisions known at each cutoff, reserves the last 30% of time for evaluation, and separates it with a 72-hour embargo. Release checks Brier scores, calibration and a comparison baseline, with at least 30 evaluation cases and 10 distinct cycles, plus real-data auditing and human review. A current database export or passing synthetic tests does not prove those requirements are met.