A small source-linked record, not a complete archive
This starter dataset uses official OpenAI documentation and original reset posts by Tibo (@thsottiaux). Each record includes a concrete source, a short quote, the stated eligibility, and a day-level date. Confirmed means the source supports the recorded announcement or action; it does not mean we inspected every affected account.
The history is deliberately incomplete. We do not import competitor totals, fill gaps with assumed resets, or treat a missing event as proof that nothing happened. The checked-at timestamp records this editorial source review, not continuous monitoring, a service availability guarantee, or access to a personal usage dashboard.
Dates and event types stay distinct
A post’s publication time is not necessarily the time a reset took effect. These records therefore leave the exact occurrence timestamp empty. Announcement and completion posts describing one reset should not become two events. OpenAI explicitly dates the documented September global reset to September 7; a possible September 8 announcement/date discrepancy remains unresolved, so it is not counted separately.
Global records describe directly applied resets or announcements of them, within the source’s eligibility scope. Banked records describe saved reset grants that users redeem themselves. A banked grant must not advance a global-reset clock. Future limit changes or other updates should remain separately classified rather than being silently counted as resets.
Historical estimates are not a reset schedule
The homepage offers a low-confidence historical estimate, not an announced date or a calibrated probability. It uses the median of the latest 12 completed intervals between up to 13 distinct confirmed public-reset dates, excluding banked grants and future records. At least six intervals are required. The median is added to the latest reviewed date and rounded to the nearest calendar day. The displayed range is the 25th–75th percentile interval, rounded outward to whole days, not a 50% chance of a reset.
If the whole projected range passes without a newer reviewed reset, the point and range advance together by the rounded median gap. This creates the next rolling cadence window without adding an event to the archive or assuming that an undocumented reset occurred. A newer reported completion still pauses the estimate until archive review. Calculations use UTC calendar dates, never an exact reset time, personal account allowance, or banked grant availability.
Live Polymarket Yes prices are a separate, date-specific signal, not cumulative odds or a Codex Limits forecast. Its rules can count banked grants and qualifying same-day announcements, not only completed automatic resets, but only qualifying actions after that market was created. Thin trading can make prices unreliable. Market prices never alter the historical calculation or become verified reset records; the display requires an active contract, an unexpired New York date, a recent snapshot, and a clear separation from OpenAI commitments.