When is the next reset?

Over the past 90 days, automatic reset announcements on the same UTC date count as one event day. We need at least 30 days of observation and 5 event days to show a probability. The third-party historical record may be incomplete.

The Gamma–Poisson model uses α = 1 + event days and β = 7 + observed days. The baseline probability of at least one announcement in the next h days is 1 − (β / (β + h))^α.

Reset previews, model releases, usage incidents, and community signals have manually chosen weights of 2.0, 0.35, 0.25, and 0.10. Each is scaled by source weight and decays with age. Only the past 72 hours count, and only the strongest signal in each group is used.

These weights are experimental assumptions, not trained or calibrated values. Banked resets and ordinary client updates do not increase the automatic-reset estimate. Without valid new signals, the adjusted estimate equals the baseline. The 7-day result is a longer-range extrapolation.

We save each forecast and its source cutoff times, then wait for the outcome. Missing collection periods do not count as “no reset.” We do not present another project’s results as our own.

Historical replay and prospective scoring are kept separate. We need timestamped forecasts and comparisons against the baseline using Brier score and Log Loss before claiming better accuracy.

tibo-reset-watch informed public-post monitoring; tibo-reset-lab provides an auditable forecasting experiment. This site is independently implemented.

Automatic resets, banked resets that must be redeemed, and regular account usage renewals are different events. Read OpenAI’s explanation ↗