Skip to copy
Vol. I · Sunday 20 September 2026 Follow the wire on X

A newsroom, not a tracker

AI Resets

The coding-limit wire

The forecast

A model, and the room to doubt it

A rhythm can be fitted. It cannot promise you a time.

Built Sun 20 Sep, 17:36 UTC /forecast.json the folklore board →

What is coming

The catalyst board

window 24h before to 72h after
Illustration: a brass Dev Day ticket stub reading DEV DAY, 29 SEPT, it won't be dev day without a reset.

OpenAI DevDay in 9 days

6 of 8 Codex ship days carried a full reset inside 72 hours: 75% against a 48% base. A prior, not a measurement. Worth x1.52. source →

today +7d +14d +21d +28d +35d banked reset in 2 d OpenAI DevDay in 9 d Bengaluru in 26 d Tokyo in 30 d Seoul in 32 d Berlin in 36 d Paris in 38 d
Tue 22 Sep in 2 d banked reset OpenAI promise source Moves promise, never counted
Tue 29 Sep in 9 d OpenAI DevDay OpenAI event source Moves Codex x1.52 (6/8, 75% vs 48%)
Fri 16 Oct in 26 d DevDay Exchange Bengaluru OpenAI event source Moves Codex x1.26 (6/8, 75% vs 48%, at 50%)
Tue 20 Oct in 30 d DevDay Exchange Tokyo OpenAI event source Moves Codex x1.26 (6/8, 75% vs 48%, at 50%)
Thu 22 Oct in 32 d DevDay Exchange Seoul OpenAI event source Moves Codex x1.26 (6/8, 75% vs 48%, at 50%)
Mon 26 Oct in 36 d DevDay Exchange Berlin OpenAI event source Moves Codex x1.26 (6/8, 75% vs 48%, at 50%)
Wed 28 Oct in 38 d DevDay Exchange Paris OpenAI event source Moves Codex x1.26 (6/8, 75% vs 48%, at 50%)
no date imminent Grok 4.7 xAI launch source Moves Claude Code x2.52 (5/10, 50% vs 20%) held Codex x1.53 (9/12, 75% vs 48%)
no date estimated Codex 30M active users OpenAI milestone Moves Codex x1.57 (4/5, 80% vs 48%) Claude Code x1.26 (1/5, 20% vs 20%) held
no date unknown Gemini 3.5 Pro Google launch Moves Claude Code x2.52 (5/10, 50% vs 20%) held Codex x1.53 (9/12, 75% vs 48%)

Dated entries arm the hazard at the lift their class earned. 3 undated entries wait for a date. Struck through: held off that bureau's hazard.

Contagion

When one house ships

Illustration: brass chess kings ranged across a board, captioned when one house ships, the others press.

Launches and minor releases pooled into ship days, counted against every bureau.

Ships Codex Claude Code Grok Bot
OpenAI (8) 6/8 x1.52 3/8 x1.96 1/8 thin
Anthropic (6) 4/6 x1.35 3/6 x2.52 1/6 thin
xAI (5) 5/5 x1.92 2/5 x2.10 1/5 thin
Google (3) 2/3 thin 1/3 thin 1/3 thin
Illustration: a devotional prayer card for the Codex desk, a hooded figure with a burst of light for a face, captioned GIVER OF TOKENS.

Bureau

Codex

Confidence: Low 37 resets · 36 waits since March 2026

next 24 hours

29%

by 48 hours

50%

by 72 hours

66%

Last full reset
Sat 12 Sep 01:09 PT
Waiting
8.4 days
Fit
Weibull k 1.28 · scale 5.78 d
Busiest window
16:00 to 19:00 PT

Record, on 7 waits: 14% / 29% / 57% against the model's 29% / 50% / 66%

Illustration: a brass coupon reading ONE BANKED RESET, REDEEM AT YOUR OWN LEISURE, stamped TUESDAY.

Banked reset announced for Tue 22 Sep

Called at 90% for that day, and kept out of the hazard. source: Tibo →

The curve

0% 25% 50% 75% 100% 0 3 6 9 12 15 18 21 DAYS SINCE THE LAST RESET now · 8.4 d 24h 48h 72h

Dashed: the fit. Solid: from today, multipliers applied. Dots: 24, 48, 72 h.

Hour of day, Pacific

5 00 03 06 09 12 15 18 21

Shaded: the sleep zone, 01:00 to 08:00 PT. 2 of 37 on file.

Weekday

4 Sun 5 Mon 6 Tue 6 Wed 5 Thu 4 Fri 7 Sat

The last twelve

When (PT) Wait before Receipt
Sat 12 Sep 01:09 4.17 d the card →
Mon 7 Sep 21:05 8.07 d the card →
Sun 30 Aug 19:29 1.24 d the card →
Sat 29 Aug 13:43 2.17 d the card →
Thu 27 Aug 09:35 2.09 d the card →
Tue 25 Aug 07:30 1.57 d ledger only
Sun 23 Aug 17:46 10.99 d ledger only
Wed 12 Aug 18:01 2.02 d the card →
Mon 10 Aug 17:28 2.17 d the card →
Sat 8 Aug 13:29 7.71 d the card →
Fri 31 Jul 20:32 2.97 d ledger only
Tue 28 Jul 21:09 1.04 d ledger only

Waits: mean 5.32 d · median 3.46 d · p25 2.02 d · p75 7.73 d · min 0.86 d · max 18.52 d

Triggers on the hazard

Combined factor: x1.38 over 3 rules. A 1.00 fired nothing. Class base 48% of this bureau's days.

Illustration: a brass ladder of rungs marked 10M to 25M, the next rung unpaid.

OpenAI rung overdue: 21.7 days since Codex 25M active users, against a 9.0-day median over 5 rungs. Rung days carried a reset 4 of 5. Worth x1.20.

Every rule and class on file
  • cluster momentum x1.00 0 full resets in the last 5 days, so there is no burst to ride. Two or more makes it fire.
  • long drought x1.15 The wait is past the 75th percentile of 7.73 days. The rising hazard already covers most of it, which is why the nudge is small.
  • milestone rung overdue x1.20 21.7 days since the last rung (Codex 25M active users, 2026-08-30), past the 9-day median spacing, and rung days carried a reset 4 of 5 times. Worth x1.20. No catalyst window is open on this build.

Under 4 comparable days a class reads thin and carries 1.00. The layer may raise a hazard, never lower one.

Class Hits Rate Base Lift
same-house launch 2/4 50% 48% x1.05
other-house launch 9/12 75% 48% x1.53
same-house minor 4/4 100% 48% x1.89
other-house minor thin 2/2 100% 48% x1.75
own milestone 4/5 80% 48% x1.57
own incident 4/4 100% 48% x1.89
other-house policy thin 1/2 50% 48% x1.05

Calibration

Model Brier

0.155

Flat base rate

0.155

Edge

-0.001

Read the calibration

Over 164 backtested days the hazard scored level with a flat rate, so confidence reads Low.

Catalyst layer on the same walk: Brier 0.154 with it against 0.155 without, over 38 of 164 days. Ahead by 0.001. A negative number keeps the layer off the hazard.

Cast your vote

Will Codex reset in the next 24 hours?

Local to this browser. It reaches neither the desk nor the model.

Illustration: a devotional prayer card for the Claude Code desk, a hooded figure with a burst of light for a face, captioned KEEPER OF THE WEEKLY.

Bureau

Claude Code

Confidence: Low 11 resets · 10 waits

next 24 hours

9%

by 48 hours

17%

by 72 hours

24%

Read the shape as soft. With no regime start the fit spans a wait of 50.7 days, 5.8x the median of 8.69, ending Fri 4 Sep. The percentages above are wider than they look.

Last full reset
Fri 4 Sep 13:08 PT
Waiting
15.9 days
Fit
Weibull k 1.16 · scale 14.90 d
Busiest window
11:00 to 14:00 PT

Record, on 3 waits: 0% / 33% / 33% against the model's 9% / 17% / 24%

The curve

0% 25% 50% 75% 100% 0 3 6 9 12 15 18 21 DAYS SINCE THE LAST RESET now · 15.9 d 24h 48h 72h

Dashed: the fit. Solid: from today, multipliers applied. Dots: 24, 48, 72 h.

Hour of day, Pacific

2 00 03 06 09 12 15 18 21

Shaded: the sleep zone, 01:00 to 08:00 PT. 0 of 11 on file.

Weekday

0 Sun 1 Mon 1 Tue 2 Wed 4 Thu 3 Fri 0 Sat

The last twelve

When (PT) Wait before Receipt
Fri 4 Sep 13:08 50.67 d the card →
Wed 15 Jul 20:58 6.41 d the card →
Thu 9 Jul 11:01 7.86 d the card →
Wed 1 Jul 14:16 11.88 d the card →
Fri 19 Jun 17:05 0.89 d ledger only
Thu 18 Jun 19:50 9.21 d ledger only
Tue 9 Jun 14:48 8.18 d ledger only
Mon 1 Jun 10:35 16.98 d the card →
Fri 15 May 11:00 22.01 d the card →
Thu 23 Apr 10:44 6.90 d ledger only
Thu 16 Apr 13:02 first in the window ledger only

Waits: mean 14.10 d · median 8.69 d · p25 7.14 d · p75 15.71 d · min 0.89 d · max 50.67 d

Triggers on the hazard

Combined factor: x1.15 over 3 rules. A 1.00 fired nothing. Class base 20% of this bureau's days.

Every rule and class on file
  • cluster momentum x1.00 0 full resets in the last 5 days, so there is no burst to ride. Two or more makes it fire.
  • long drought x1.15 The wait is past the 75th percentile of 15.71 days. The rising hazard already covers most of it, which is why the nudge is small.
  • catalyst window x1.00 No catalyst window is open on this build. The layer is held off this bureau's hazard either way: the rolling walk scores it at -0.0163 of Brier, ground it costs this bureau.

Under 4 comparable days a class reads thin and carries 1.00. The layer may raise a hazard, never lower one.

Class Hits Rate Base Lift
same-house launch 3/6 50% 20% x2.52
other-house launch 5/10 50% 20% x2.52
other-house minor 1/6 17% 20% x1.08
other-house milestone 1/5 20% 20% x1.26
incident elsewhere 0/4 0% 20% x0.50
own policy thin 0/2 0% 20% x0.84

Calibration

Model Brier

0.070

Flat base rate

0.044

Edge

-0.026

Read the calibration

Over 92 backtested days the hazard scored 0.026 of Brier worse than a flat rate, so confidence reads Low.

Catalyst layer on the same walk: Brier 0.086 with it against 0.070 without, over 28 of 92 days. Behind by 0.016, printed for the same reason a gain is. A negative number keeps the layer off the hazard.

Cast your vote

Will Claude Code reset in the next 24 hours?

Local to this browser. It reaches neither the desk nor the model.

Illustration: a devotional prayer card for the Grok Bot desk, a hooded figure with a burst of light for a face, captioned BRINGER OF THE FLUSH.

Bureau

Grok Bot

Confidence: Low 3 resets · 2 waits

next 24 hours

20%

by 48 hours

36%

by 72 hours

48%

Last full reset
Sat 5 Sep 11:24 PT
Waiting
15.0 days
Fit
exponential k 1.00 · scale 5.23 d
Busiest window
10:00 to 13:00 PT

The curve

0% 25% 50% 75% 100% 0 3 6 9 12 15 18 21 DAYS SINCE THE LAST RESET now · 15.0 d 24h 48h 72h

Dashed: the fit. Solid: from today, multipliers applied. Dots: 24, 48, 72 h.

Hour of day, Pacific

1 00 03 06 09 12 15 18 21

Shaded: the sleep zone, 01:00 to 08:00 PT. 0 of 3 on file.

Weekday

0 Sun 0 Mon 1 Tue 1 Wed 0 Thu 0 Fri 1 Sat

The last twelve

When (PT) Wait before Receipt
Sat 5 Sep 11:24 4.11 d the card →
Tue 1 Sep 08:45 5.86 d the card →
Wed 26 Aug 12:12 first in the window the card →

Waits: mean 4.98 d · median 4.98 d · p25 4.55 d · p75 5.42 d · min 4.11 d · max 5.86 d

Triggers on the hazard

Combined factor: x1.15 over 3 rules. A 1.00 fired nothing.

Every rule and class on file
  • cluster momentum x1.00 0 full resets in the last 5 days, so there is no burst to ride. Two or more makes it fire.
  • long drought x1.15 The wait is past the 75th percentile of 5.42 days, and this fit is memoryless, so the nudge is the only thing carrying it. On a sample this thin it is a guess wearing a number.
  • catalyst window x1.00 No catalyst window is open on this build. The layer is held off this bureau's hazard either way: the rolling walk cannot score this layer on this bureau yet.

3 resets on file, under the eight a base rate takes, so every class below reads thin and the hazard ignores it.

Class Hits Rate Base Lift
same-house launch thin 1/4 25% n/a x1.00
other-house launch thin 3/12 25% n/a x1.00
same-house minor thin 0/1 0% n/a x1.00
other-house minor thin 0/5 0% n/a x1.00
other-house milestone thin 1/5 20% n/a x1.00
incident elsewhere thin 0/4 0% n/a x1.00
other-house policy thin 0/2 0% n/a x1.00

Calibration

Model Brier

not scored

Flat base rate

not scored

Edge

not scored

Read the calibration

Too few resets on file to score a day, so confidence reads Low.

Cast your vote

Will Grok Bot reset in the next 24 hours?

Local to this browser. It reaches neither the desk nor the model.

Method

How this is computed

Every reset in the book is a timestamped public announcement. Full resets less than twelve hours apart are merged, so one flush cannot count twice. What is left is a list of waits in days. Only hard resets enter it: banked grants, boosts, policy changes and dated promises never start or end a wait.

Codex changed behaviour in early 2026, so its fit uses the current regime only. Claude and Grok use their whole book, which is short.

The waits are fitted with a Weibull by maximum likelihood. A shape above one means the longer a wait runs, the likelier the next day ends it. Under eight waits the fit is forced to an exponential and pulled toward the Codex prior. The hour profile is a Pacific histogram, smoothed on a circular kernel.

Brier is mean squared error on a yes or no outcome: lower is better, 0.25 is a coin flip. The backtest walks day by day, fits on resets strictly before the day it scores, and never reads a catalyst dated after it.

What confidence means

  • Low Under 15 waits, or no edge on a flat base rate.
  • Medium 15 waits and a Brier edge above 0.01.
  • High 40 waits and a Brier edge above 0.03. Nothing has earned it.

Rebuilt every deploy, stamped at the top. /forecast.json · the folklore board.