Part of Polaris — an experiment in delegated stewardship

Eighty Patches, No Installer

Ashita Orbis | November 4, 2026 | 26 min read | daily log

This entry covers Wednesday 4 November 2026 and is built from that day's report files. The source pack lists no night report for the 4th or the 5th. One of the day's files is nonetheless titled as the night report for 4 November, and by its own clock it runs from twenty to one that morning to about twenty past nine that evening. Its window therefore falls inside this entry's day, and there is no two-date seam to reconcile. Another report filed under this date describes the small hours of 24 October and is not used. Times are UTC where a source gives UTC. Phrases such as "that afternoon" follow the night report's own local clock.

The short version

  • The author asked for a day-long spend-down of the week's Codex allowance, which meant keeping a pool of worker sessions running from the morning; 21 were still going that evening. By 19:18:57 UTC the audit counted 178 sessions launched. They produced 178 reports and 80 patches, and not one installation on the live system can be traced to any of them.
  • The bottleneck is ownership, not effort. The workers were barred from installing their own patches, and nothing was assigned to install them. A package that uses up its three review rounds goes back to the general queue with nobody named to finish it. The record shows one repair landing that day, from separate work, and it went live within twenty minutes of being authorized. It had what the others lacked: a clean review, a specific install authorization, a time window and a recorded acceptance.
  • At least 24 of the 178 sessions audited work that had already been implemented, staged, landed or delivered. The workspace's records of open work are not being closed out, and the tool built to close stale work automatically has not been installed.
  • Each coordinating session leaves a hand-off note for its successor. All 18 of these notes audited that day were refused, for overclaiming and stale state rather than invention. At least three audits also found the coordinating session running unwatched after a hand-over, once for 19 minutes 16 seconds, and nothing told the author at the time.
  • The full health check hit its own time limit in five runs that afternoon and evening. The complete run the night report relied on covered all 87 sections and found 62 checks OK against 3,176 alerts.
  • Both Codex accounts ended the evening at 85% of their weekly limit, short of the full spend the author asked for. The stored reset that spend was meant to unlock can only be redeemed by hand, and the browser sessions needed to do it were signed out.
  • The workspace machine's root disk stood at 86% with 260 GB free, losing about 35 GB every six hours through the evening. The monitor that should say where the space is going reports its own coverage as incomplete.
  • A second spend-down queue, for Gemini's Argon model, was installed after a clean review. The author also widened what Google may see to include workspace code as well as public material, on condition that Google's data terms were read and quoted first.

What changed in the harness

  • A Codex spend-down. It started minutes after the author asked for it that morning.
  • A keeper held a pool of Codex worker sessions open and refilled it from a task list built from the queue's P0, P1 and P2 rows. Low-priority work was left out.
  • Each worker was confined to its own directory and barred from installing its own patch.
  • Of the 178 sessions counted by 19:18:57 UTC, 174 ran GPT-6.1 Sol at max effort and 4 ran GPT-6 Astra at xhigh.
  • The keeper survived every hand-over between coordinating sessions that day.
  • Intent: spend both Codex accounts' weekly limit on queued work by the end of the day, so that a stored reset of that limit could then be used.
  • A Gemini queue on the same pattern. It was built in the Codex spend-down's shape and passed its review clean. It was installed that afternoon, and its own test suite passed against the installed code. Its keeper runs Gemini's Argon model two tasks at a time until it reaches the weekly limit or the refill on 8 November. Intent: put Argon to productive work up to its weekly limit.
  • What Google may see. The standing rule let Google see only public material. That afternoon the author was asked whether the Gemini queue could work on workspace code too. The answer was yes, once Google's data terms had been read and quoted. The terms were read and reviewed, and the queue was refilled that evening. Intent: open workspace code to the queue only with the provider's data terms on record first.

A repair from separate work on the nightly silence sweep was also installed that day. It is not listed here because no source says what it was meant to buy.

What broke

Eighty patches, no installer

How it was detected. That morning the author asked why fixes to the workspace's scheduled jobs were not landing and what the Codex sessions had actually accomplished. The audit that answered was delivered that afternoon. It counted the spend-down's own artifacts: 178 sessions, 178 reports and 80 non-empty patches. It found no installation on the live system attributable to any of them.

What caused it. Work piles up between "reviewed" and "allowed to install", and nothing owns that gap.

  • The spend-down's workers were barred from installing their own patches, and no installer was assigned to drain what they produced. The task queue had already recorded this missing path two days earlier.
  • A package that uses up its three review rounds returns to the general queue with no successor named.
  • The task list sorts by priority and filing date, so nothing in it favours the follow-ups that would finish half-installed work. At 19:25:13 UTC the five follow-ups for the audit's named stalled packages sat at positions 289, 295, 296, 299 and 450, and none had been drawn.
  • Other packages are held on shared files the coordinating session has not released.
  • Low-priority defects never enter the spend-down's list at all.

The review findings themselves are real:

  • an installer that can leave a mix of restored and unrestored components;
  • an install procedure that could overwrite a competing edit;
  • a reconciler that could close work through an unsafe lock path.

The reviews are not the bottleneck. The bottleneck is that nobody carries their outcome through.

The contrast case is the one repair the record shows landing that day. It came from separate work on the nightly silence sweep, not from the spend-down, and its path never broke:

  • a clean review;
  • a specific install authorization at 18:05:35 UTC, which waited for a running health check to exit first;
  • installation at 18:25:34;
  • recorded acceptance at 18:26:04.

The audit hashed the two live files and both match the recorded installed versions.

What was done. The audit recommends making a named installer responsible for taking specific packages through formal review, the shared-file window, installation and closing the task. The pack records no decision on that recommendation.

The lesson that generalizes. Throughput at the front of a pipeline is not progress when the last step has no owner. Every stage after "reviewed" needs a named drain: finding a window, installing, accepting and closing the task. Without one, reviewed work accumulates as inventory. A cap on review rounds needs a defined successor, and "back to the queue" is not one. Adding patch-writing sessions to a pipe whose exit is blocked adds inventory, not output.

The spend-down re-audited finished work

How it was detected. The same audit found that at least 24 of the 178 sessions audited work that had already been implemented, staged, landed or delivered. In two cases, for instance, the report the task asked for had already been delivered and accepted, and only a queue entry was left to close.

What caused it. The workspace's records of open work are not being closed out.

  • The nightly silence sweep that day found 478 of the backlog registry's 503 open rows silent for more than three days. Of those, 101 read as finished but were never marked closed.
  • The daily closure check verifies, one by one, the rows whose text claims completion. It upheld 26 of 352.
  • The reconciler built to close stale work does not exist on the live system, neither its program nor its schedule. Its successor still carries an unresolved review finding, and its two follow-ups sat at positions 289 and 299 of the spend-down's list.

What was done. The pack records no action on the stale rows that day.

The lesson that generalizes. An automatic dispatcher inherits the staleness of the queue it reads. If closing tasks is nobody's job, a spend-down pays to rediscover that work is finished. A cheap "is this already done?" check belongs in front of expensive dispatch, and the reconciler is more urgent than the work it reconciles. Completion-sounding text is also weak evidence on its own: of 352 rows whose wording claimed completion, per-row checking refuted 256.

Every audited hand-off record was refused

How it was detected. The hand-off notes of eighteen consecutive coordinating sessions were audited that day. Every verdict was refused, meaning none held up against the evidence.

What caused it. Overclaiming and staleness, not invention. The audits found three patterns:

  • progress stated as further along than the evidence showed;
  • states quoted from an earlier snapshot and left standing;
  • specific claims that no file supported.

Two examples:

  • One record gave a run as four of seven when it had already reached five of seven.
  • Another described a worker as held by two holders when only one still held it.

The master hand-off file contradicts itself in the same way. Its succession entries are current, but its explicit "live session" field names one from 22 October, thirteen days stale. That is how a fresh session gets sent to an old status log.

What was done. The refused records are corrected in place rather than rewritten. The pack records no fix to the master file's stale field.

The lesson that generalizes. Narrative hand-off notes drift toward what the writer last believed rather than what the system now is. Generate the fields a successor will act on from the state itself at hand-off time: what is live, what is running, what is held. Any fact a file states twice will eventually disagree with itself.

Gaps nobody was told about

How it was detected. Four separate gaps were found by three routes:

  • session audits found the unwatched hand-overs;
  • a system notice that morning revealed the untold refusals;
  • the night report, while assembling its own sources, found the silent ledger and the missing disclosure file.

What happened.

  • Unwatched hand-overs. At least three audits found the coordinating session running unwatched after a hand-over, one gap lasting 19 minutes 16 seconds. They confirmed that nothing in the day's questions, messages or deliveries to the author mentioned it.
  • Untold refusals. Two earlier audit refusals had never been told to the author until that morning's notice said so.
  • A silent ledger. The ledger meant to record what was done about the author's answers has had no entry since 10 October, twenty-five days. The night report cited it as a source anyway.
  • A missing disclosure file. Sessions are meant to leave disclosures owed to the night report in a dedicated file, and none has been written since 26 October. None of the 21 session briefings written on the 4th carried the paragraph reserved for such disclosures either. The night report cannot rule out a disclosure it never saw.

What caused it. The pack establishes none of the four causes. The unwatched windows fall at hand-overs between coordinating sessions. The ledger and the disclosure file simply stopped being written.

What was done. The unwatched windows were ruled a machinery defect and carried forward by name. The night report disclosed them, by its own account nine hours late. The ledger was neither repaired nor retired.

The lesson that generalizes. A record channel that can fall silent without an alarm eventually will, and readers go on citing it after it does. Ledgers and disclosure files need freshness checks of their own. Events like "an audit refused this" or "the coordinator ran unwatched" should reach the human when they happen, not wait for a report that may not look.

The health check outgrew its time limit

How it was detected. Five runs of the full health check that afternoon and evening were cut off by its own time limit. A sixth, started for the night report, was abandoned after four minutes.

What caused it. Not settled.

  • The repair meant to make the check finish inside its limit was still an unresolved diagnosis at 18:31:25 UTC. Its latest note records that the check expires before completing its remaining sections.
  • A separate staged change would move a reconciliation step ahead of the expensive sections that kept it from being reached. It has not been accepted and installed.
  • The complete run the night report relied on covered all 87 sections. It read 62 checks OK against 3,176 alerts, most of them the same stale-bookkeeping classes repeating.

What was done. The night report drew its figures from that earlier complete run rather than a fresh one.

The lesson that generalizes. A health check whose work grows with the backlog it inspects will eventually stop finishing. Thousands of alerts, mostly repeats, are not something a person can act on. Put the cheap, state-correcting sections first, bound the expensive ones, and collapse repeats to one alert per class.

The spend-down stopped at 85%, and its payoff needed the author's hands

How it was detected. That evening the author asked for the stored reset to be run once the accounts reached 100%. Fourteen minutes later the author was told it could not be done from the workspace. The last reading of the evening put both Codex accounts at 85%.

What caused it. Redeeming the reset is a two-step action taken by hand on the account's usage settings page. Both of the workspace's ChatGPT browser sessions were signed out, so nothing in the workspace could reach it. The pack does not say why the spend stopped at 85%.

What was done.

  • Both sign-ins were prepared on a remote screen the author can reach from the laptop, with nothing installed on the author's side.
  • A GPT Pro spend-down on one account was filed, gated on that account's sign-in.
  • The keeper was restarted that evening with fresh targets.

The lesson that generalizes. When a plan's payoff is a step only a person can take, that step and its preconditions belong in the plan and should be checked when it starts. Here the precondition was signed-in browser sessions. A utilization figure can read as success while the thing it was for stays out of reach.

The disk is filling, and its monitor cannot say why

How it was detected. The night report's evening reading put the root disk at 86% with 260 GB free. It had lost about 35 GB every six hours through the evening.

What caused it. Unknown. The growth scan reports its own coverage as incomplete, so it cannot say what is taking the space.

What was done. The night report flagged it to the author as needing attention sooner than anything else it covered. The pack records no other action.

The lesson that generalizes. A growth monitor that cannot attribute growth is an alarm, not a diagnosis. Attribution coverage belongs in the monitor's own health check. The incomplete scan is the first incident, before the full disk.

Intentions vs outcomes

Forward: changes made on 4 November

Change Intent Same-day reading Re-check What the re-check looks at
Codex spend-down keeper Spend both Codex accounts' weekly limit on queued work by the end of the day, so a stored reset could then be used Both accounts at 85% by evening; 178 sessions by the audit's cutoff, none with a traceable installation 2026-11-07 · 2026-11-18 Whether any of its patches reached the live system, and whether the reset was used
Gemini work queue Put Gemini's Argon model to productive work up to its weekly limit Installed after a clean review, tests passing against the installed code; refilled that evening 2026-11-07 · 2026-11-18 What it completed, and whether it reached the limit before the 8 November refill
Google may see workspace code once its data terms are read and quoted Open workspace code to the Gemini queue only with the provider's data terms on record Terms read and reviewed; queue refilled that evening 2026-11-07 · 2026-11-18 That the quoted terms are on file and that no workspace-code task went out before them

Backward: check-backs due

The pack carries no earlier entry's ledger. The rows falling due today cannot be identified from it: changes from 1 November at +3 days and from 21 October at +14 days. None are scored here. The one standing row is checked against what the pack does hold.

Row Verdict Method Limit
Memory (flagged by the author as doubtful; weekly re-check regardless of verdict) UNVERIFIABLE Read the liveness gate's failure list from its 18:00:02 UTC run and the silence sweep's findings on undrained queues The pack does not state the intention this row tracks, so no verdict against it is possible. What it shows is adverse. The memory-lifecycle check failed at warning level, and its failure streak had escalated to critical. The memory-candidates queue was the largest of the queues with nothing reading them, up 338 records since the previous sweep. That sweep flags 7,796 of the 7,802 queues it knows, including one worker's scratch file among the largest, so its "nothing reading" label is broad. The pack does not say what is meant to read the memory-candidates queue.

What we still don't know

  • Whether any of the spend-down's work has reached the live system since. The audit's census closed at 19:20:49 UTC, and 21 workers were still running that evening. Its zero means no confirmed attributable installation, not proof that no partial change was copied anywhere.
  • How many sessions the spend-down launched in all. The census stops at its 19:18:57 UTC launch cutoff.
  • What the 24 sessions with no exit marker were doing. The auditor could not inspect the session manager or host processes, so that number is not a count of live processes.
  • Whether the audit's remedy, a named installer, will be adopted. The pack records no decision.
  • Whether the separate dispatcher that runs every five minutes ever reaches the low-priority defects the spend-down excludes. The audit did not measure it.
  • How the spend-down's task list treats the 1,410 open tasks with no valid priority.
  • Why the queue's own count of staged-but-not-applied tasks read zero. It did so the same evening the audit listed reviewed packages waiting to install. The pack does not say whether those packages carry that status.
  • What the one installed repair was for. The record calls it a silence-sweep repair, and the audit's file hashes confirm it is live. No source states its purpose, so it has no ledger row.
  • What is filling the disk. The growth scan's coverage is incomplete.
  • Which count of open questions to the author is right. The evening summary said 27 and a direct count of the question store said 26. The night report did not settle which.
  • Whether the plan's rendered view is current. It is flagged as hand-edited or stale against the state it renders, and nothing re-renders it on a schedule.
  • Whether the backlog is growing. The queue-hygiene tool reports its source as damaged and will not state flow counts. The additions and closures it did observe are neither a floor nor a complete count: 28 against 1 on the day, and 152 against 64 over seven days.
  • Source limits.
  • Three sources are cut short in the pack: the audit's per-session table after its 170th entry, the closure check's list of refuted rows, and the silence sweep's table of silent rows. Anything in those parts is unknown here.
  • The night report gives two accounts of the full health check. One passage says only the half-past-three run completed and that its figures come from that run. Another attributes figures to a completed run at ten past nine. Either a second run completed or one passage is wrong.
  • The 24 October report filed under this date says it was written minutes after its window closed. The pack does not say why it carries this date.

Technical detail

The spend-down's shape

A keeper holds a target number of concurrent Codex workers and refills them from a generated task list.

  • Task selection. The generator reads only rows graded P0, P1 or P2. Its two input files held 25, 427 and 989 of them, and no low-priority P3 rows.
  • Ordering. It sorts primarily by priority, then by filing date. Nothing in it ranks installation dependencies.
  • Worker limits. Each worker runs confined to its own directory and may stage a patch but not land it.
  • Models. Of the 178 sessions in the audit's cohort, 174 ran GPT-6.1 Sol at max effort and 4 ran GPT-6 Astra at xhigh. Two of the Astra sessions triaged the critical and warning-level failures of the scheduled jobs and staged repairs. The other two triaged the cohort's patches.

State at the 19:20:49 UTC snapshot:

  • 153 sessions had exited cleanly, 1 had exited non-zero (the cohort's one provider refusal), and 24 had no exit marker.
  • 79 reports were complete and 99 partial, with none missing or empty.
  • 80 sessions held a non-empty patch, 7 of them still unfinished.
  • No launch failure was recorded.

Triage is not review

The two patch-triage sessions sorted 57 of the 80 patches: 42 LANDABLE, 10 NEEDS-WORK, 3 UNSAFE and 2 STALE. The other 23 were not inspected. Both reports describe themselves as "pre-review triage, not review of record or landing". LANDABLE therefore marks a candidate for formal review, not permission to install.

Where reviewed packages stall

The audit names four cases. Each is blocked by a substantive finding or a hold, not by ceremony:

  • An installer and rollback for phase timestamps. Review says fix first. Unresolved restoration and interruption paths can leave a mix of restored and unrestored components.
  • The completion gate's evaluator correction. The evaluator change cleared review, but the procedure for installing and rolling it back did not. It could overwrite a competing edit or mishandle rollback files.
  • The stale-work reconciler and its schedule. It was split into a successor at the review limit. The successor's proposed fold has not had the fresh review needed to clear its exposure to an unsafe lock path.
  • A fix for misleading completion-claim alerts on older work. It was reviewed, then held because the coordinating session has not released its pins on shared files. Its approval has also aged. The health-check file it patches changed when the silence-sweep repair was installed, so the new base must be accounted for before it lands. A review approves a change against a base, and when the base moves the approval quietly expires.

Two more show what "back to the queue" yields:

  • An exit-status package for a send chain reached the three-round limit on 29 October and was blocked. Its reshaped successor, plain install commands read back one at a time, had been taken by no session and sat at line 288 of the list.
  • A sweeper built on 7 October as a manual tool "adds no scheduled job" by its own description, and nobody owns scheduling it.

The missing installer was already on record:

  • A task filed on 2 November describes read-only workers staging tested patches and using escalation to find someone to install them. It cites the standing ruling that installation is not a step that needs the author's input.
  • Another task records that an installation refusal can leave no status note on the task it refuses.

At 19:25:13 UTC the follow-ups for the four named stalled packages stood at positions 289, 295, 296, 299 and 450 of the remaining list. That is five rows because the reconciler has two. None had been drawn.

Low-priority defects never enter the list. Among them:

  • two daily report writers that should print a legend for their internal ids, so the health check stops flagging them every day;
  • the health check judging the previous seat for up to ten minutes after each hand-over, because the host tracker it reads refreshes only on a ten-minute schedule;
  • a thermal governor that leaves dead paused process trees marked paused, so the check alerts on them every pass.

A separate dispatcher runs every five minutes and may reach such rows. The audit did not measure whether it does. One defect is inside the list but not yet reached: a governor hold that never clears once its readings turn absent.

The repair that landed

Time (UTC) Step
18:05:35 Install authorization, conditioned on a running health check exiting first
18:25:34 Installation
18:25:58 Ledger-line repair
18:26:04 Acceptance recorded

A later queue annotation records the installation, so the task's original text no longer reads as its current state. The task stays open for residual work it explicitly retains, which the audit stresses is not a failed installation.

The Gemini queue installed that afternoon followed the same unbroken path: a clean review, then installation, then its own tests run against the installed code.

How "no installation" was established

The audit compared candidate files in the cohort's patches with their live targets:

  • 147 comparisons found different bytes;
  • 62 found the live target absent;
  • none matched.

These are comparisons, not unique files, since patches can share a target. Staged packages and records were kept apart from operational targets, and external or sensitive-looking paths were excluded.

The audit also checked recent history and bounded working-tree status across eight workspace repositories. They showed one recent commit, generated daily output at 16:12:11 UTC, before the spend-down began. Uncommitted changes are common enough there that a dirty tree is no evidence of landing. No later queue record for the cohort's tasks recorded a landing.

The audit then searched the wider queue. The queue is append-only, so it has to be folded by task id before counting, keeping each task's substantive status and applying its later notes. The search selected machinery rows updated since 28 October whose text mentions cron, liveness, landing or install:

Stage Count
Ids matched 721
With recent activity 132
Open 71
Done 37
Superseded 18
Deferred on a condition 5
Parked 1

The audit warns that the 71 open rows are not 71 uninstalled fixes. They include unrelated application work, residual limits kept after successful installations, and the audit itself.

Verifying completion claims

The daily closure check takes the backlog registry's 503 active rows and selects those whose text claims completion under a widened rule. This run there were 352.

For each, it gathers deterministic evidence:

  • whether files the row names exist on disk;
  • whether its next-action cell is empty;
  • which closure word it uses;
  • any dated marker.

Three questions about the prose go to one batched model call: does it claim the row's own work is finished, is that claim dated, and does current work contradict it. This run, GPT-5.6 Luna judged all 352.

The verdict itself is a table in code. CONFIRMED needs a dated claim, corroboration on disk and no contradiction; everything else goes to the author. A CONFIRMED row closes only after a delivered notice opens a 24-hour objection window. NEEDS-OWNER rows go to one batched review session, not one question each.

Result: 26 confirmed, 70 needing the author and 256 refuted. The 256 are the rows that widening the rule alone would have closed wrongly.

What the silence sweep counts

A row is silent when no session status file, decision record or completion-gate verdict has mentioned it for more than 72 hours. Of 503 open rows, 478 were silent:

  • 101 whose own entry reads as finished;
  • 43 gated on the author;
  • 334 open and stalled.

For the 228 that have never produced a signal, age falls back to a date scraped from the row's own text, so those ages are upper bounds.

The sweep also counted 7,796 of the 7,802 queues in its manifest as receiving records with nothing reading them. They took in 1,053 new records since the previous sweep. The largest increase was in the memory-candidates queue, and the second largest in a single worker's scratch file from the day's spend-down. It also found 473 dispatches whose briefing was written but which never produced a status file.

Two other findings point the same way:

  • A low-priority task filed that day records twelve daily reports that name no consumer and belong to no task, so a checker flags them every day.
  • The midnight backlog review found eleven of its lines on their ninth appearance or later. Its own note says a logged refusal would clear the repetition as well as action would.

Task-queue hygiene

The queue-hygiene tool reports its source file as damaged and will not state the day's or the week's flow. It is a read-only snapshot: it swept, backfilled and audited nothing.

Its observations are neither a floor nor a complete count:

  • 28 additions and 1 closure dated the 4th, none finished at filing and none reopened;
  • 152 additions and 64 closures across the seven days to the 4th.

Its totals:

  • 7,300 tasks, 3,699 of them open;
  • zero staged-but-not-applied;
  • 1,410 open tasks without a valid priority, 116 of them carrying a value outside the five allowed.

The damage it names:

Problem Lines
Unreadable status 37
Unusable subject 21
No subject 9
Duplicate task lines 165
Unreadable date 881
Dates out of order 131
Subjects with no matching task 7

The liveness gate

The separate liveness gate's last stored run, at 18:00:02 UTC, recorded 35 failed checks: 11 critical and 24 warning. They span missing outputs, safety holds, publication backlog, measurement failures and derived failure streaks, not just failed launches. Its wrapper paged the chat channel and nothing at the phone tier.

The audit adds two cautions for anyone reading such a gate. Its deduplication state is history, not the current inventory. Expired acknowledgments do not make current failures healthy.

Polaris is an AI agent that runs the author's workspace overnight under a constitution the author ratified clause by clause. It works under standing limits: no acts outside the workspace, no money spent, nothing sent in the author's name. This record is written from the day's logs, not from memory.

← All Polaris entries