Chance of a discretionary quota reset · next 48 hours

51%Codexlast reset 6 days ago · 7 in the past fortnight

27%Claude Codefrom a stated baseline — not measured

95%Claude Code scheduled recyclecounted separately — do not add

Uncalibrated. These are hand-set priors from the public event record, not fitted parameters, and none has been checked against an outcome. There is no accuracy score behind them yet. The Codex figure is derived from this ledger's own record of when resets actually happened and updates as time passes; the Claude Code figure comes from a stated baseline, because no supported channel exposes its usage state. How they are built · why they may be wrong.

Forecast

Two vendors, two methodologies, because they expose different amounts of state. These are the less trustworthy half of this project — the ledger is the asset.

Probability of at least one discretionary reset

At the July 2026 baseline — Codex 38h since its last reset with 6 resets in the trailing 14 days; Claude Code 25h since recycle at 62% weekly utilisation — with no live evidence signals.

regime	Codex				Claude Code
regime	12h	24h	48h	72h	12h	24h	48h	72h
quiet	10%	17%	27%	35%	3%	5%	10%	14%
normal	14%	24%	39%	50%	4%	9%	16%	24%
launch	20%	34%	55%	68%	7%	14%	27%	37%

Claude Code's scheduled track

Reported separately from the discretionary number, because the vendor gives two knowable things and they answer different questions. Do not add them. The scheduled recycle rides an observed ~72h cadence; at 25h elapsed the next one is due in ~47h.

window	12h	24h	48h	72h
scheduled recycle	0%	0%	95%	100%

Instantaneous rates, launch regime

vendor	λ (events/hour)	mean wait
codex	0.0192	52h
claude-code	0.0064	156h

What these numbers cannot do

Three limits, stated because a percentage sign invites more confidence than this deserves:

The clock is pinned to 2026-07-07T12:00:00.000Z (ordinary weekday hour). Claude Code's figures move with the day of week — the calendar term is the only clock-dependent part of either model — so an unpinned page would show numbers nobody could reproduce.
The evidence terms are inert. Both models carry a weight for a vendor-employee social post. Scraping X is out of scope by decision, so those terms can never rise above zero. Treat both models as having one fewer input than the formulas suggest.
Claude Code's utilisation is hand-entered, not measured. The sentinel that would have measured it was closed — no supported channel exposes subscription quota state, and none distinguishes a vendor-wide grant from an ordinary rollover. See Methodology.

Full formulas, every weight and half-life, and the calibration path are on Methodology.