FORECAST METHOD
How the Codex reset forecast works
Codex Reset separates public wording from historical cadence, publishes the evidence behind each percentage, and keeps the limits close to the claim.
What the percentage means
Alert v3 has two Watch scores only: 93% when Tibo states a concrete date, range or clock, and 83% for an explicit future reset without concrete timing. Elevated context may remain visible but never notifies. In quiet periods the page returns to the separate cadence model.
| public signal | score | what it establishes |
|---|---|---|
| Strong Watch with concrete timing | 93% | Tibo explicitly points to a future global reset and states a date, range or clock: 93%. |
| Strong Watch without concrete timing | 83% | Tibo explicitly promises a future global reset without a specific time: 83%. |
| Current Tibo words with concrete timing | 93% | The same 93% rule applies when Tibo's current words make the claim; reply topology does not subtract points. |
| Elevated Watch promoted by a valid fact | 83% | Exact, fresh corroboration may promote the same local Watch to strong; it never supplies public prose, probability or a second Event. |
| Elevated context | 83% | Hedged or suggestive wording may remain visible as context, but it does not create an Alert Event. |
| Third-party or timing-only context | 83% | Third-party text and bare timing continuations cannot manufacture or repeat an Alert. |
Only a strong Watch notifies subscribers: 83% without concrete timing or 93% with a stated date, range or clock. Elevated context never rings; topology is trust context, not score arithmetic.
The cadence model, honestly
When no scored wording is live, rate-v3 estimates a recent reset rate from the latest eight completed reset intervals, weighting each interval equally. It treats arrivals as memoryless: waiting longer does not raise the odds.
The model rounds its 24-hour and 48-hour outputs to whole percentage points. What says how thin the evidence is behind them is the range printed beside each one, not the last digit. It describes irregular goodwill resets in the verified record; it does not turn them into a schedule.
The wait comparison uses the same recent window: up to eight completed reset intervals, with the actual sample count shown; fewer than three intervals hides the comparison. The percentage counts intervals strictly shorter than the current wait, excluding ties. The typical wait, average and longest are the median, mean and maximum within that window. The current unfinished wait is not a sample. The range beside each probability describes how thin the evidence is given the exposure behind the rate, not a promise that the truth lands inside it. Recent history is not a schedule. Sample through names the end of the last completed interval actually used, always in UTC. Calculated at is the time the forecast was computed (updated_at); it does not mean a new sample was added.
Will Stage 0 (2026-09-10): history is the base. The forecast also publishes the state of Tibo's reset-related posts over the last seven days of the feed (quiet since the reset, a hint, or an unclear post; a live promise belongs to the Alert above) with the post that decided it. That state is shown and explained but not counted in the percentage, until the record shows it has predictive value; the 2026-09-09 test found that weighting the percentage by it made the forecast worse.
Offline single-factor evaluation (2026-09-10, shared cutoff 2026-09-09 11:00 UTC): adding the observed portion of the unfinished wait to the rate denominator returned DATA. Neither event definition establishes continuous observation; pre-window event counts are 18/20 for the production definition and 17/20 for the old audit. In the reconstructed 90-day diagnostic, 24h Brier was slightly lower but 48h Brier was worse, with both block intervals crossing zero. This does not establish an improvement. The historical baseline remains rate-v3; ordinary posts and hints are not additional probability factors.
inference In the live API snapshot observed 2026-08-28, the walk-forward test covered 295 predictions. rate-v3 recorded a Brier score of 0.104, versus 0.106 for the naive baseline; lower is better. It tied the previous rate-v2 score at 0.104, so this is evidence against the naive baseline, not proof that rate-v3 beats every alternative.
Track record
This is a historical score-v1 snapshot dated 2026-08-27; it does not define Alert v3. It reports the pre-v3 samples exactly as observed, with no smoothing, extrapolation, or hidden misses.
| band | resolved n | hits | hit rate | open |
|---|---|---|---|---|
| dated commitment | 2 | 1 | 50% | none |
| plain promise | 0 | — | insufficient history | none |
| legacy tease (historical) | 0 | — | insufficient history | The historical snapshot recorded 1 open sample at score 50; this is not a current v3 score claim. |
| legacy additional-evidence watch (historical) | 0 | — | insufficient history | none |
Retained as an append-only pre-v3 snapshot; it is not rewritten into the current decision contract.
The curated timeline is the best public landing-time proxy, not direct per-account telemetry. A hit means a verified reset landed inside the declared window plus the published 60-minute grace.
When resets happen
The UTC histogram below groups the verified announcement record by hour. The pattern is a rhythm, not a schedule: historical clustering cannot name the next arrival.
Every recorded announcement is shown in the same hourly bins used on the homepage.
Boundaries
These limits travel with the forecast so an excerpt does not overstate what Codex Reset knows.
- Codex Reset cannot see anyone's OpenAI account, quota, or usage. Personal countdowns use only times the visitor enters, and the /codex-usage paste analyzer keeps supplied text in the browser.
- A reset forecast is a probability derived from verified public signals and cadence. It is not an OpenAI schedule, leak, or commitment, and it does not predict outages.
- A reset is called verified only when a public announcement says so. Timing that merely sits near an incident is labeled temporal, never causal.
- Codex limit numbers are quoted from OpenAI's published pages with their checked date. Where OpenAI publishes no number, including the weekly quota size, Codex Reset says so instead of estimating.
- The range printed beside a percentage is a sampling band on the rate estimate. It is not a calibrated coverage guarantee, and it never changes which number the alert chain publishes.