How the forecast works
What is forecast
ResetChance forecasts a public event, not your account: the first qualifying public confirmation that a Codex extra reset has completed, within 24 or 48 hours of the forecast start. Target: RC-TARGET-1.
It cannot see your usage, your plan's limits or when your own access returns. A reset announcement is not a confirmation, and a confirmation does not tell you that your own account can be used again.
Not counted: your own usage window, banked credit, general limit increases, product launches, recovery from an outage, promises of a future reset, or the same confirmation repeated.
Target versions
- RC-TARGET-1, from 2026-09-26 (current): The first qualifying public confirmation that a Codex extra reset has completed, within 24 or 48 hours of the forecast start.
A changed target is always a new version; past events are never re-labelled to make one series. Sources: codex-reset.com (Data: codex-reset.com). If an explanation or link here is wrong, report it for revision.
Models and evaluation
No production model approved.
Until a model passes its frozen evaluation, the site shows facts and a reason instead of a number. The research candidate below is not in use.
| Candidate | RC-ABC-AGE-STATE-1 (variants C0, C1, C2) |
|---|---|
| Forecast origins | every day at 00:00, 06:00, 12:00, 18:00 UTC |
| History used | 365 days, at most 1460 frames |
| Frame weight | 0.25, halving every 60 days |
| Starting count per outcome | 0.5 |
| Age kernel width (log days) | 0.75 |
| Age shrinkage | 14 days |
| Promise kernel width (log hours) | 1 |
| State shrinkage | 28 days |
| Outcome known after | 48 hours |
| Training minimum | 60 complete days and 20 independent reset events |
| Test minimum | 60 complete days and 8 independent reset events |
| Prospective minimum | 60 complete days, 6 independent events, 8 seven-day blocks |
| Score comparison | Brier score against the strongest baseline on the same origins; 2000 resamples of 7-day blocks |
| Numeric answers required | at least 90% of healthy origins |
Baselines: the older site's rate formula (rebuilt from its documented maths, not copied), the recent overall rate, and the age-only candidate. Each forecast uses only evidence known at its start; evidence found later that the reset had already happened voids that start instead of counting as a hit. Resampling blocks are not new independent events.
Limits and versions
Initial version (methodology-1). No earlier versions exist.
- 2026-09-28, methodology-1 — Technical change: Initial version: target RC-TARGET-1 and the research candidate RC-ABC-AGE-STATE-1 with its fixed parameters. No model is approved.
Limits: few independent resets exist, source coverage before this project started is not proven, and forecast frames close in time are related. Where there is not enough data the page says so instead of showing a number. The forecast is a statistical estimate from public history, not inside knowledge.