Banked resets changed the meaning of “waiting” — what I learned logging Codex limits for five months

Part 3 in my notes on reading public Codex reset signals. This one is about how my own habits changed, not about a product feature.

The thing I thought I understood

When I started logging public Codex reset signals in late April, I had a simple mental model. Quota runs down, time passes, quota comes back. Waiting was one long continuous stretch that you endured until it ended.

That model was fine for about five weeks. Then a banked reset showed up in the log — a one-off credit that doesn’t follow the normal 5-hour or weekly cycle at all — and quietly broke the assumption.

Banked resets don’t wait, they stack

A banked reset is a credit that loads into an account. You don’t wait for it. It arrives when it arrives, possibly because of something entirely unrelated to your usage (a milestone, a product rollout, an incident, a “we’re giving one for every day you didn’t have access” promotion). It isn’t tied to your clock.

Once I understood that, “waiting” split into two very different experiences:

Waiting for a window to refill. Fixed, predictable-ish, boring. You know roughly when it ends.
Waiting for a banked credit you didn’t earn by waiting. Unpredictable, and frequently not about you at all.

The second one is the interesting one, because it inverts the old logic. Under the old model, usage bought waiting. Under the new one, something outside your control could hand you capacity regardless of how carefully you rationed — and it could just as easily skip your plan entirely.

The five-month arc, in three acts

The full log runs April 28 → October 10, 2026: 44 public signals across 36 unique reset days, a 4.7-day average gap, and one 37-day silence in the middle.

Act one — the quiet months (late April to June). Three signals, and the gaps between them were 37, 8, and 27 days. An average of 24 days. In this period, “waiting” literally meant waiting, because there was almost nothing else happening. This is the stretch my mental model was built on, and it’s the least representative part of the log.

Act two — the regime change (July onward). The average gap dropped to 2.8 days and it stayed there. Signals started arriving in clusters: 12 unique reset days in August, 10 in September. The log stops reading like a metronome and starts reading like a calendar with events on it.

Act three — the banked-credit era (late September to now). The recent entries are dominated by banked reset and reset-card activity — Tibo posting about banked credits loading for paid accounts, reset cards tied to launches, and one banked-reset rollout that landed with enough of a kerfuffle that it was logged twice on the same day. The 28-day “improvement or full reset” window announced on October 5 is logged as an upcoming watch signal, not a landed reset, precisely because of this distinction.

What actually changed for me

Three real shifts, none of them about prediction:

I stopped treating a banked reset as the thing I was waiting for. It’s not part of a cycle; it’s an event that may or may not include me. Assuming it lands is how people talk themselves into a bad week.
I stopped reading long gaps as “nothing is happening.” The 37-day stretch wasn’t a broken tracker; it was the real shape of the early log. And it could happen again. A tracker that only knows how to describe busy periods is a tracker that will confidently mislead you the next time it’s quiet.
I got more disciplined about “for whom.” A reset card for paid accounts is a rollout claim. Some of my logged signals are specifically paid-plan. Whether a specific credit reached a specific account is, by construction, not something a public log can know.

None of this made me better at predicting resets. It made me better at not being surprised by them — which, for planning around a tool, turns out to be the thing I actually needed.

The honest caveat: I only observe public signals. I have no visibility into anyone’s actual quota, plan, or credit balance. Everything above describes the pattern of reports, not the behavior of any account.

If you’re reading this thread

The reasoning behind the confidence labels is on the methodology page. If you want to check my work rather than take my word for it, the raw log is public — every event with its date, source, and confidence is on the history page, and the same data as JSON is at reset-events.json. The running summary I publish is reset-status.json.

If my read of the pattern is wrong, you can find the exact entry that shows it.