Method

Official X accounts, vendor status pages, release notes and docs are marked official. Hacker News and other public threads are marked community. Each source has its own polling interval, and its last check is shown on the provider page.

A language model reads each new item as untrusted text and returns a structured verdict: reset, banked reset, announced reset, boost, limit change or incident, with a confidence from 0 to 100. Text from the web is treated as data and can never change how the model behaves or what gets published.

1. 01Official source, confidence 75 or higher: verified automatically.
2. 02Community source, confidence 60 or higher: shown as unconfirmed. Unconfirmed events never move a forecast.
3. 03Promotion: an official report, or three independent community reports of the same event, upgrades unconfirmed to verified.
4. 04Everything else, including any keyword-only match when no model is available, waits in a human review queue and is not shown.

With at least three completed reset intervals, we estimate the reset rate from the most recent eight and report the chance of one or more resets in the next 24 and 48 hours, with an 80% sampling band. The model is memoryless: goodwill resets are irregular, so waiting longer does not raise the odds.

An active official announcement sets a floor under that number: 93% when it states a time or range, 83% by 48h (and 50% by 24h) when it does not. It lapses 60 hours after posting, or when a reset lands.

Every day we backtest the model: would it have predicted a reset in the next 24 hours using only what was known then? Brier score, lower is better, against a baseline that uses all history.

We cannot see anyone’s account or usage. A forecast is a probability, not a promise or a leak. Personal windows differ per account, which is why your own clock is entered by you. Vendors rarely publish how large their weekly limits are, and we do not guess.