HELP & METHODOLOGY
Announcements, odds and calendar guide
Scope and sources
This independent tool tracks Codex resets, product improvements and Tibo’s 28-day promise. ModelYard and codex-reset.com indexes are read in parallel; an optional server-side X API can be added. Categories can lag or omit posts; disagreements are excluded from statistics. Every record links to X. This tool does not verify your personal quota.
Cross-checking and response time
Mirrors of the same post count once. Recent explicit plans remain pending until completion. OpenAI Status supplies Codex component and incident context, never a reset probability multiplier. The attributed external 24/48h estimate remains separate from this site’s rest-of-today model.
Caching and source failures
Pages read the server snapshot immediately; refreshes run in parallel in the background with a 6-second cap per request. Healthy announcement sources remain usable if others fail. If none are live, history stays visible and odds pause. Upstreams are checked at most once per minute; rate limits honor Retry-After. Manual checks do not bypass the cooldown.
Data: codex-reset.com · OpenAI Status
What do today’s odds mean?
The estimate is the chance that the next verified full-reset announcement arrives before today ends. Chinese defaults to Beijing time; English defaults to US Pacific time, including daylight saving time. Overrides are saved per language. Post times, date filters and CSV use the selected timezone. A log-normal model is fitted to up to 12 independent reset intervals, requiring at least 6. The formula is [S(t)−S(t+hours left today)]/S(t). The log-scale standard deviation is at least 0.35 to avoid excess certainty from tiny samples.
Limitations and pause rules
The odds are not backtested or calibrated and are not an official forecast. Percentages are hidden for stale data, too few samples, elapsed times beyond 1.5 times the longest interval, or a tiny survival tail. An observed reset is presented as a fact rather than a 100% prediction. Planned and banked resets are not executed full resets.
How are the campaign dates chosen?
The calendar uses the verified promise to ship an improvement or a full reset each day for 28 days. It starts the day after publication in Pacific time; this is our interpretation. Verified resets and updates are grouped by Pacific publication date, while post timestamps use the selected display timezone. Unconfirmed days are not failures. Future dates remain unannounced; forecasts do not invent an official schedule.
Banked resets and daily progress
Resets, improvements, both outcomes, and banked reset notices are shown separately. A day counts as recorded when a verified reset or improvement is indexed. A banked reset is not an automatic full reset and does not count as a completed reset day. Click a date to inspect the evidence.
Historical reference windows
Follow-up reset notices within 12 hours count as one episode. Up to 13 episodes form up to 12 intervals. At least 5 are required for a median and 20th–80th percentile reference window. This window is not a probability interval and does not roll forward when expired.
Languages and notifications
Chinese and English use separate URLs with a language switch. Titles and summaries use the index’s language fields; missing English summaries are disclosed with a link to the original post. Browser alerts need HTTPS, permission and an open page. The server keeps checking every minute. Initial historical loads do not trigger alerts. Favorites, timezone and preferences stay in your browser.
RSS · Scheduled reset calendar (no forecasts)
History and service health
The server persists up to 2,000 related records and fetches up to 100 index entries per check. Each source has a health indicator. All announcement sources failing or checks older than 90 minutes are flagged; old records remain available. Preserve the data volume when updating.