Study 08 · started 2026-07-23Record · 36 events

Policy over time: how the seats were adjusted to the usage pools

When the owner changed who holds a seat, what did the subscription pools look like at that moment — and does the record show the changes answering the pools?

What the record shows

36 dated policy changes, each placed on the pool reading of its moment; 28 of them name the pool that motivated them.

  • 87 dated versions of the policy file and 4,141 pool readings (2026-07-04 → 2026-09-18) were joined by timestamp; the join is the study.
  • The 3 September flip is the clearest case: claude at 6% used with 82% of the week gone, codex at 92% — the build and review leads went to opus, and the router's lead picks over the next day went 35 to claude, 1 to codex.
  • The 4 September declare-trim is the mirror image: a second codex declarer added when codex was cold was dropped the moment codex was at 48% used with 9% of the week elapsed.
  • 13 of the 36 changes moved a seat holder, a round cap, the routing mode or a lockout; the rest are tuning and stay on the tape as smaller ticks.
policy changes
36
on the rail
13
name a pool
28
pool readings
4,141
policy versions
87
counts read from data/policy-over-time.json

The tape

2026-07-04 → 2026-09-18

Three pools, every reading, every change. Step between changes and the playhead flies along the tape while the pools move exactly as the readings say they did; drag the tape to scrub. The board underneath is the policy in force at the selected moment.

Weekly pool usage, 4 July to 17 September 2026

4,141 readings · 36 policy changes · drag anywhere on the tape to scrub; click a tick to jump; ← → step between changes

Thu 23 Jul 202616:57Z
claude50%codex49%GLM—06 Jul13 Jul20 Jul27 Jul03 Aug10 Aug17 Aug24 Aug31 Aug07 Sep14 Sep2026-07-23 · Checkpoint cadence trimmed to conserve codex2026-07-26 · Lean codex on every remaining review round2026-08-02 · Codex-heavy override for two in-flight units2026-08-06 · The policy becomes a file2026-08-11 · Phase-4 iterate moves from Sol to Terra2026-08-12 · GPT-5.4 retired from every seat2026-08-17 · Unified policy: high-tier review cap 8 → 6, daemon escalation at R7 / R8, GLM block staged (off)2026-08-18 · GLM is switched on while its own pool is running hot2026-08-18 · Claude off every review seat except the R1 Sonnet swarm; GLM gains discovery breadth and an R1 lens2026-08-21 · A GLM lattice replaces the opus breadth seat in discovery2026-08-21 · Rung-scaled checkpoint fence: GLM finds, codex declares2026-08-25 · Terra is ratified as the declarer after a 15-run A/B2026-08-25 · Responsive evaluator numbers shipped: codex target 200% on two accounts, pivot at 77% / day 62026-08-25 · Leads become per-phase; cleanup goes to Terra — the first non-Claude lead window2026-08-29 · Build and review leads move to terra2026-08-30 · Claude-burn cut: Claude leaves every review seat2026-08-31 · Codex takes the picker and the light/standard checkpoint finds2026-09-01 · Build lead flips to Opus: the temporary claude-burn flip2026-09-01 · Review lead flips to Opus too; the deterministic seat router lands in shadow mode2026-09-02 · The seat router goes live2026-09-03 · Leads restored to Terra: 'the temporary claude-burn flip is over'2026-09-03 · Claude is cold, so opus goes first on build and review2026-09-04 · Claude drain window: leads pinned to Opus until the reset; live re-pin at every rotation; picker back to Opus2026-09-04 · Flip-back, A1-relax and Phase D: the claude week reset early; GLM may find on consent/money/pg; a GLM swarm-4 lens rides every round2026-09-04 · Declare-trim: the second codex declarer is dropped once codex is no longer cold2026-09-05 · “We must never fly blind”: the brake rules2026-09-06 · Plan and plan-loop leads → Fable 5.1 for a 29-hour claude-flat-out window; responsivity off2026-09-06 · Fable takes the review seats for a flat-out window2026-09-07 · Fable seats off, plan leads back to Opus: the Fable sub-pool is exhausted2026-09-07 · Posture flip → codex-heavy, Claude off, ahead of the codex reset2026-09-08 · Two Claude accounts: front-load the active one at 2× (target 200%) for days 1–42026-09-09 · The build lead becomes responsive2026-09-13 · Co-plan drafter 2: Sol → GPT-6 Astra2026-09-13 · Review lead becomes responsive too ('review claude')2026-09-14 · Pre-P4 codex lockout: build and review leads pinned to Opus; no fence review rounds; pre-build review skipped2026-09-15 · Zero-codex: the total lockout
  • claude · account A
  • claude · account B
  • codex · account A
  • codex · account B
  • GLM pool
  • policy change (bright = on the rail)
  • account switch — each account has its own weekly counter

The pools at this moment

at the change
claude · account B active
50%of the week used

week — elapsed · pace — · projected —

account A last read at 96%, 9 h 16 min earlier

codexnamed in the change note
49%of the week used

week — elapsed · pace — · projected —

GLM
—of the week used

week — elapsed · pace — · projected —

Reading 2026-07-23 16:53Z, 3 min before this change · verdict unknown.

What the owner changed

change 1 of 36

P1 Thu 23 Jul 2026 · 16:57Z · logged to the minute · before the policy was a file

Checkpoint cadence trimmed to conserve codex

Per fence: one Sonnet swarm, then a single GPT-5.4 confirm (not 5.5); the major review waits until the PR is open.

codex poolThe owner's own note calls it a codex-conservation move.

source · Outbound-Tool/.sprint/spindle/pm_decisions.log (line 15)

The policy board

Every seat, cap and switch in force at the selected moment. Cells that changed at this event carry an amber edge and show the previous value beneath.

no policy file yet

Before 6 August 2026 the numbers lived in prose and per-unit directives, not in a policy file. The board starts at the fourth change; the three earlier ones show the owner’s directive and the pools, without a board.

How to read it

caveats stated

The join is by timestamp, so the honest limits are about time.

  • A change's pool reading is the nearest reading at or before it, and the card says how old that reading is. Pace and projection come from the same reading; between changes only the used % is shown.
  • Three changes from July predate the policy file: they are logged directives, placed on the tape without a board.
  • Day-granular rulings are placed at the day, and each change's stamp says whether its time is exact, to the minute, to the hour, or taken from a backup file.
  • A weekly counter restarts at an account switch (the rings on the tape) and at the weekly reset, which is why the traces drop.
  • GLM telemetry starts on 16 August. DeepSeek is metered per call, not weekly, so it has no lane.
  • Only this fleet is shown. Task names are masked to their first five characters; account names, paths and people never enter the record.

The changes on the rail

13 of 36

Each one moved a seat holder, a round cap, the routing mode or a lockout. The other changes are tuning and stay on the tape as smaller ticks.

  1. 012026-08-06 · The policy becomes a file
  2. 022026-08-18 · GLM is switched on while its own pool is running hot
  3. 032026-08-21 · A GLM lattice replaces the opus breadth seat in discovery
  4. 042026-08-25 · Terra is ratified as the declarer after a 15-run A/B
  5. 052026-08-29 · Build and review leads move to terra
  6. 062026-08-30 · Claude-burn cut: Claude leaves every review seat
  7. 072026-09-02 · The seat router goes live
  8. 082026-09-03 · Claude is cold, so opus goes first on build and review
  9. 092026-09-04 · Declare-trim: the second codex declarer is dropped once codex is no longer cold
  10. 102026-09-05 · “We must never fly blind”: the brake rules
  11. 112026-09-06 · Fable takes the review seats for a flat-out window
  12. 122026-09-09 · The build lead becomes responsive
  13. 132026-09-15 · Zero-codex: the total lockout

Raw data

generated 2026-09-18 by scripts/gen-policy-over-time.ts

Every number on this page is read from data/policy-over-time.json, which the generator rewrites from these source files. A data update is a commit.

  • /records/policy-timeline-2026-09-17.jsonFrozen record: every event with its board state and diff, every throttling reading, the router's seatings around each change. Built once from the policy file's dated history, the supervisor log and the throttling log; task names masked; no live source is read again.
  • /records/lead-seat-log-2026-09-17.jsonThe published lead-seat pins the router lane on the tape is drawn from.

I. Nocturne in E♭ major Op. 9 № 2 · Chopin

0:00 / 4:16