What is the probability actually about?
The homepage always reserves a place for the probability percentage. It estimates the chance of another automatic Codex allowance reset in the next 24, 48 or 72 hours. These are cumulative windows from the latest data refresh. A banked reset credit is a separate event, and your own weekly timer is separate too.
Our historical labels are public announcement times, used as a proxy for reset timing. We cannot measure when a reset reaches an individual account. An announced reset stays “awaiting rollout” until a newer completed entry appears in the public feed; a deadline passing is never treated as completion.
A percentage and a Pacific countdown
NR-1.7 retains the continuous waiting-age curve and adds a numerical rule for an untimed reset promise. The main estimate covers the next 24 hours. A stated delivery window is shown in Pacific Time (America/Los_Angeles); it is not an exact rollout appointment.
The historical baseline weights past intervals with a 30-day half-life. Six age bands (0–2, 2–5, 5–10, 10–15, 15–20 and 20+ days) estimate reset-announcement hazards. Each band is smoothed toward the overall rate using 14 days of prior exposure, and rates are joined between band centers. This removes the ten-day plateau. A longer wait does not automatically mean a higher 24-hour chance: recorded intervals cluster at some ages and thin out at others.
For an explicit, source-linked promise whose deadline is still ahead and falls inside the forecast horizon, we assign a 95% announcement prior, combining it as P = 0.95 + 0.05 × P(history), capped at 98%. This is a declared assumption, not a measured fulfillment rate. If a promise gives no delivery time, we use a declared 95% eventual-fulfillment assumption and a 14-hour delivery scale. For a rolling window, we condition on the promise still being unfulfilled, then combine that conditional prior with the history baseline. This raises the near-term estimate substantially but is not calibrated on observed promise-to-completion delays. If the promise remains unfulfilled, the conditional near-term estimate falls; it does not silently become a completed reset. The percentage remains visible; the state stays awaiting confirmation while the live feed lists the promise as scheduled. A missed timed deadline also stays awaiting confirmation until a completed entry appears. When a post says “next week,” we interpret that as the following Monday-to-Monday Pacific calendar week. For a declared, uncalibrated illustration, we spread the 95% fulfillment assumption uniformly over that week and condition on no completion yet. Horizons ending before that week receive no next-week boost; the post does not get the 14-hour untimed rule. These calendar boundaries and the uniform timing are our assumptions, not a date promised by Tibo.
Signal weights you can inspect
| Signal | Odds multiplier | Rule |
|---|---|---|
| Explicit timed reset promise | 95% prior | Only when a future deadline falls inside the horizon. |
| Untimed reset promise | 95% eventual; 14-hour timing scale | Conditional rolling-window prior while still unfulfilled. Editorial assumptions, not backtested. |
| “Next week” reset promise | 95% over next Pacific calendar week | Uniform timing within an inferred Monday-to-Monday range; no boost before it starts. Editorial, not backtested. |
| Vague hint | Up to 2× | Must mention a reset. Unrelated product posts do not qualify. |
| Reviewed official release hint | Up to 2× | One underlying release event. Quotes and retellings do not multiply weight. |
| Confirmed Codex model launch | Up to 2.5× | Requires an official, Codex-specific source; a launch is still not a reset promise. |
| Independently verified rollout report | Up to 1.2× | Requires reviewed supporting evidence. |
| Unverified rumor / unrelated post | 1× | No numerical adjustment. |
| Explicit denial / cancellation | Down to 0.4× | The newest qualifying statement replaces an older promise. |
| Recent relevant recovery | Up to 1.15× | One resolved incident in the last 24 hours; fades to neutral and does not stack with a promise. |
| Banked credit / unrelated post | 1× | No effect on the automatic reset forecast. |
What the historical backtest says
The NR-1.6 history curve is a modest change, not a proven prediction breakthrough. These results do not validate the timed or untimed announcement assumptions or release-event weights.
We replayed 166 daily forecast origins, 2026-03-28 through 2026-09-09, fitting only announcements dated before each origin. Each forecast has a full 72-hour outcome window. At least 12 completed intervals are required before evaluation starts.
| Window | NR-1 Brier score | Constant-rate baseline | Recorded positives |
|---|---|---|---|
| 24 hours | 0.1610 | 0.1599 | 32 / 166 |
| 48 hours | 0.2469 | 0.2453 | 60 / 166 |
| 72 hours | 0.2743 | 0.2749 | 82 / 166 |
Brier score measures squared probability error; lower is better, zero is perfect. The baseline uses the same recency-weighted history at a single constant rate. The 24-hour model does not beat that baseline in this replay. This curve was revised after examining the archive, so these scores are not independent validation. Historical predictions also tend to understate the observed event frequency.
This is a retrospective, incomplete catalogue. We do not have historical discovery timestamps, so the replay cannot account for reporting delays or entries that were missed. Overlapping 48/72-hour outcomes are correlated. This is a history-only backtest; it does not validate today’s tweet-adjusted probability.
Download every backtest prediction ↗
Where the live evidence comes from
The forecast service checks Codex Resets’ public API for newer reset entries, scheduled announcements and active watches, retaining Tibo’s original X links. We use the source text, not that site’s probability. This is a curated reset-signal feed, not complete access to Tibo’s X timeline. A missing signal does not prove he has posted nothing.
We also review original X posts in a browser and store the source URL, author, publication time, first observation, review time, event grouping and expiry. Tibo and Sam’s release-week posts count as one event because Sam quotes Tibo. Community routing claims stay at neutral weight until their underlying evidence is independently checked. A missing post, inaccessible account or failed collection is not evidence against a reset.
Service context comes from the official OpenAI status page. General API or ChatGPT incidents without Codex / ChatGPT Work scope do not change the estimate. Community likes, repost totals and anonymous rumors do not receive numerical weight.
A Cloudflare background collector checks public feeds every five minutes and saves the accepted archive and forecast. The initial HTML, live browser, history and RSS use that common record. This refreshes existing reviewed-event weights as well. It does not continuously read X accounts: browser-reviewed additions require another review and publication. The evidence panel shows the actual browser review time, separately from each five-minute calculation. A review older than 24 hours is labelled due; existing evidence still follows its stated expiry rather than treating a collection gap as a denial. An open homepage refreshes every five minutes. Every response includes source checks and an expiry time. If the history feed fails, the live percentage is unavailable. If another source fails, the estimate uses the remaining accepted evidence and discloses the gap; a failure is never assigned a negative-signal multiplier. Expired snapshots keep the percentage slot but replace its value with “—%.” If an announced delivery window passes without confirmation, we show “awaiting confirmation” instead of inventing a reset.
We checked Tibo’s original announcement: “a reset is also landing by midnight today.” The source does not specify a timezone; the public feed interprets the deadline as September 12, 00:00 PDT (Pacific Time). That interpretation is not an official exact delivery time.
What improves this model next?
We now save prospective forecast snapshots, source changes and community round outcomes. A sufficient sample of completed forecast windows is still needed before fitting tweet weights or claiming calibration. NR-1 is a first transparent model, not a proven advantage. Model versions and the current retrospective backtest are published here so the claims can be checked.