INDEPENDENT CODEX RESET TRACKER
Codex reset forecast methodology and evidence policy
The live estimate is this project’s own heuristic, not the original website’s algorithm and not an official reset schedule. Evidence, forecasts and confirmed events have different meanings.
Sources and freshness
The collector reads Tibo’s X account (@thsottiaux), the OpenAI status feed, openai/codex GitHub releases, OpenAI news and relevant Hacker News discussions. General collection defaults to five minutes; X defaults to about fifteen minutes. Source failures retain earlier records and use backoff. A healthy page does not guarantee that every source is current.
Default rule calculation
The score starts at 3. Eligible signals from the preceding seven days contribute impact × exp(−ageInHours / 24) × 0.35. Each source category is capped at 50 before applying its configured weight. A recent confirmed reset can subtract up to 40 points, decaying over 24 hours. The San Francisco weekday 09:00–18:00 window adds 4; other times subtract 4. The 48-hour result is rounded and limited to 0–99; the 24-hour result is rounded from 65% of the 48-hour reading. Administrator changes to weights affect results.
When the reading is provisional or unavailable
The collector must have a fresh completed X sync and relevant recent evidence. Missing, stale or insufficient reset evidence can make the reading unavailable. Recent unreviewed reset reports cause a provisional estimate and pause automatic probability alerts. A high score is not a reason to confirm an event.
What the language model does
The optional language model reviews batches of new or changed public evidence, with deduplication, a cooldown and a daily call budget. Reloading the page does not itself invoke it. Provider failures can pause automatic review. Model output is advisory: it does not confirm resets, send event alerts or replace the displayed rule calculation.
Human confirmation and delivery
An administrator checks the original source, event time and reset type before choosing “Confirm & notify”. Confirmed events and pending reports are labeled separately. SMTP acceptance is recorded as a send result; it does not prove that a message reached the inbox or was read.
Validation and limits
Predictions are recorded prospectively and matched with administrator-confirmed outcomes after their windows mature. Missing outcome records stay unknown, rather than becoming “no reset”. No accuracy claim is made. Subscriber counts describe this site’s confirmed email subscriptions and do not influence the reset estimate.