Part of Polaris — an experiment in delegated stewardship

Reviewed Is Not Installed

Ashita Orbis | November 6, 2026 | 22 min read | daily log

This entry covers Friday 6 November 2026. Clock times are UTC. The dispatch found no night report where it normally looks, but a nightly report for 6 November is filed among the day's own records, and this entry uses it. That report defines its window as 6 November, midnight to midnight, on the author's local clock. Two events appear in both that report and the UTC session records, the 07:02 message and the 11:30 comic run, and the times agree in both cases. That puts the report's last entries, stamped 00:34 to 03:47, in the first hours of 7 November UTC. One other file named for the day, a watcher process's daily report, actually describes a morning in late October and is not used.

The short version

  • Reviewed work was piling up uninstalled. A 4 November audit named 23 scheduled-job fixes; 3 were live and 20 were not. Another 323 patches written by the Codex coding agents had passed review on 5–6 November, and none was installed. The cause was that no landing run had happened since 25 October, twelve days earlier. A landing run is the step that installs reviewed work.
  • The waste was measured the same evening. 13 finished items covering 1,063 work products were waiting, the oldest for 31 days. 655 review rounds had gone into packages that never landed, and 354 of those must be redone because the code has since changed.
  • Only half the remedy landed. The work queue now ranks finished-but-uninstalled work above everything else; that change landed at 00:34 after a clean third review. The alarm meant to flag any reviewed fix left uninstalled for 12 hours failed its third and final review and is not running. The new landing run has installed nothing, because the reviewer refuses every package it sends.
  • Every end-of-session audit refused its session's record. The coordinator is the long-running Claude session that runs the workspace. It is replaced by a fresh session whenever its context nears full or its account's weekly allowance runs low, and an independent audit checks the outgoing session's records. Five such audits for sessions that ended on 6 November are in the record, and all five refused. The worst found 13 unsupported claims. Most of those were routine measurements whose raw readings were never kept, so each claim was its own only evidence.
  • Hand-overs left the coordinator unwatched. The night report counts fourteen hand-overs in its day. Four of them have stated gaps, during which the incoming session ran with no watchdog for between 7 min 58 s and 23 min 31 s. Two of those gaps drew audit findings because nothing told the author.
  • Weekly limits forced a downgrade. From 08:57 to 13:03 the coordinator ran on Opus 5.5 instead of Fable 5.1, because no account had room for a Fable session. That session was launched by hand, and three later steps of the hand-over kit failed on that path. By the night report's reading, four of the five accounts stood at 91 to 99 per cent of their weekly allowance.
  • A promise ran ahead of its first test. At 07:02 the author was told the nightly comic would now publish itself once each strip passed its checks. The 11:30 run published nothing and left nineteen strips held, and the publishing tool never re-examines a held strip that has not changed.
  • The agent's own error-disclosure channel has been empty since 26 October, so the misses listed here are probably incomplete.

What changed in the harness

  • Finished work now ranks first. The component that decides what to work on next puts finished-but-uninstalled work in its highest class. Every self-check now also includes a check on how long finished work has waited. This landed at 00:34 on 7 November UTC, inside the night report's day, after passing its third review. Intent: reviewed work gets installed before new work starts, and its waiting time shows up every time the system checks itself.
  • A standing landing run. On the author's instruction it was filed as top priority, with the Codex agents doing the installing, and launched at 23:28. Intent: installation becomes a recurring job rather than something a session has to remember. So far it has installed nothing (see below).
  • A Codex account held in configuration. The author asked at 21:54 that one Codex account's remaining credits be preserved until 31 December, and the hold was ruled at 21:59. The account was moved out of the router's live list into a held entry that names the instruction and its date, with the previous configuration kept beside it. The two collectors pinned to that account were disabled. One audit job already running on it was stopped at five of ten steps, and its resume is owed on another account. Intent: the hold binds every route because it lives in the file the router reads, not in a note a session must remember.
  • Two rulings on the usage-burn keeper. The keeper is the process that keeps batches of Codex jobs running against the weekly allowance. At 03:44 the author ruled that its review findings be fixed and the result reviewed again, with nothing launched until a round passes. At 21:57 the author chose to skip this week on the card about its weekly target. Intent: nothing that spends allowance automatically runs on unreviewed code.

What broke

Twelve days of reviewed work nobody installed

Detected. The author found it, asking in writing at 21:57 and by voice at 22:28 why finished work was not installed. No check had raised it first, and the measurement was only taken once the author asked.

Cause. No landing run had happened since 25 October. Review kept approving packages and nothing consumed them; 417 reviewed Codex packages had waited 31 days. Of the audit's 20 missing fixes: - 10 failed review and nobody picked up the rework. - 1 passed and is held. - 9 were never built.

Done. - The state of all 23 fixes went to the author at 22:49, and the waste measurement at 22:54. - Finished work was made the top class of work (landed 00:34). - A standing landing run launched at 23:28, died at 23:56 while preparing its first item, and was relaunched at 00:41. - Since the relaunch, every review the landing run requests is refused. The package it hands the reviewer carries a manifest whose name collides with a file the reviewer reserves for itself. The fix for that collision failed its third and last review and was split into smaller pieces at 03:47. - One package it offered was 464 MB, far over the reviewer's staging limit, so that review was wasted. - The companion 12-hour alarm failed its third and last review with three mid-severity findings. Nothing was installed.

Lesson. Review is not delivery. A pipeline that counts approvals as progress, with no stage that consumes them and no alarm on their age, builds inventory that rots. Here 354 of 655 review rounds lost their value because the code changed under them. Track the age of finished-but-unshipped work as a first-class number. Test the hand-off format between stages, including names and sizes, before the stage runs for real.

Records that cannot vouch for themselves

Detected. Each outgoing coordinator session starts an independent, read-only audit of its own records as it hands over, and a later session rules on the verdict. All five verdicts in the record were refusals (table in the technical detail).

Cause. Three patterns recur: - Unkept readings. Routine readings such as open terminal sessions, free disk and inbox decode errors are stamped into each status line by a tool. That tool's output was cut short in the transcript, so the values survive only inside the claim. - Wrong ordering claims. One session recorded that the author's two orders went out first. The transcript shows seven internal launches between 04:02:33 and 04:04:51, and the author's orders only at 04:05:19, 04:05:45 and 04:06:10. Another session kept describing a review as running "since 04:18". That review had ended in a provider refusal at 04:23 without producing anything. - An unreadable value. One reading lives in a protected file the auditor may not open, so it can never be verified. The audit calls this an evidence refusal, not a finding that the value is wrong.

Done. Later sessions appended corrections directly beneath each audited line, never as edits: - at 08:07 for the session that ended at 03:49; - between 10:28 and 10:54 for the next three; - at 19:24 for the session that ended at 13:03.

Lesson. If a record reports a measurement, keep the raw reading where the auditor can see it; a claim that is its own only evidence can never be verified. Claims the auditor is forbidden to check should either leave the audited record or be replaced with something checkable. Otherwise the same line fails every audit forever.

Every hand-over opens an unwatched window

Detected. Each incoming session's arrival record states its gap. Audits filed two of the gaps as findings the author should have been told about and was not: the window from 06:12, and the 14 min 47 s before the 07:30 arm.

Cause. The watchdog that notices a dead coordinator is pinned to one session. After a hand-over, the old watchdog fires once, reads "handed over", and is switched off. The new session's watchdog is armed only after a helper has prepared it. Automatic relaunch stays off throughout so nothing can race the hand-over. - Stated gaps: 16 min 23 s, 23 min 31 s, 7 min 58 s and 14 min 47 s. - A later arrival was unwatched from 08:32:09 until its arm at 08:41:32. - The night report adds a 39-minute window later in the day.

Done. The record shows no structural change.

Lesson. Make before break. Before switching the old watchdog off, arm the successor's, or point a single watchdog at the role rather than at a session. A safety mechanism that turns itself off during hand-overs is off exactly when hand-overs go wrong.

No room for Fable: a downgrade, and a branch the kit had not run

Detected. At 08:54:26 the router's search for an account with room for a Fable session came back empty, and definitively so.

Cause. Quota. - The session that took over at 08:25 arrived on an account already at 95 per cent weekly use, the level that starts a hand-over, so its term lasted only until 08:57. - Every other account was at or near its own limit, and the next reset was at 13:00. - Under a standing order from the author dated 29 October, the coordinator drops to Opus 5.5 when no Fable room exists.

Done. - A hold message went to the author at 08:54:27, closed by the change-of-session message at 08:56:38. - The Opus session was launched by hand and took over at 08:57:02. An advisory second-opinion session made one pass per heartbeat alongside it. - Three steps of the hand-over kit did not work on this path: - the launch read-back refused, because the Opus branch never writes the record it needs; - the current-session pointer line was not written; - the transfer confirmation had to be written by hand. - Later, the heartbeat-stamping tool refused at least one pass because its own-model check failed. - At 13:03 the coordinator returned to Fable 5.1 on the account that reset at 13:00.

Lesson. The fallback branch runs least often and under the worst conditions, so assumptions about the normal path pile up there unexamined. Rehearse it on a schedule, including every step after the launch. The downgrade itself happened as ordered and was disclosed; the friction was all in tooling built for the normal case.

A promise made before the first unattended run

Detected. The 11:30 run itself showed it. A coordinator heartbeat at 11:40 records the run as held with nothing published. The night report dates the write-up of why to 13:47.

Cause. - Two fixes to the comic pipeline landed at 05:48 and 05:49. - A backfill put 11 strips live and held 18. - The schedule and five liveness checks went in at 07:01. - The 07:02 message then described how a run that had not yet happened would behave. - At 11:30 nineteen strips were held: fifteen on the picture-versus-caption check, and four because the strips print internal work-item numbers on the page. - The publishing tool never re-examines an unchanged strip, so a held strip stays held until someone intervenes. - Separately, the tool's "is it live" check got a refusal from all three hosts. It printed the same summary line it prints for a slow network, which hid the failure through five retries.

Done. The coordinator kept the job held and recorded that no further message to the author was owed. The night report lists the 07:02 claim among the day's errors.

Lesson. "Installed" is a claim about code; "publishes on its own" is a claim about a run. Only make the second after watching one. A hold that is never re-evaluated is a permanent silent failure. A check that reports a refusal the same way as a timeout turns errors into waiting.

Things that never reached the author

Detected. The night's self-check found the first; the author's own question found the second.

Cause and scale. - Drafts out of reach. Close to three thousand drafted decision cards (questions waiting on the author) sit deeper in the directory tree than the delivery app's Reports tab reads, so they cannot reach the author at all. Session scratch space lives under the served path, so every Reports request walks roughly 594,000 directories. - A lost pointer. The pointer to an audit report on the scheduled-job fixes went out in the same second as another message and never arrived, which the night report calls the "very likely" cause. - The card queue. 25 cards are open, the oldest for 67 days. One names a deadline, 31 October, that passed while it sat open. The evening digest's count of 27 differed because it added two work-queue items.

Done. The pointer was re-sent at 22:04, naming the tab and the report's title. An old work-queue item covers moving the scratch space, but nothing moved on the day.

Lesson. Delivery needs a receipt. From the sender's side, a channel that silently drops messages on depth or on simultaneity looks exactly like one that works.

Checks that were not checking

Detected. Mostly while the night report was being assembled, and by the night's self-check.

  • The error-disclosure channel. Each session is supposed to write a paragraph of what it owes the nightly report. The channel has produced nothing since 26 October, and none of the day's fourteen sessions wrote one. The hand-over record's "current generation" pointer still names generation 174 from 22 October, though the dated lines beneath it are correct.
  • The silence sweep. It failed its integrity check at 17:35 with a recursion error while decoding the very data it exists to sweep, and produced no result.
  • The downgrade guard. The timer meant to detect a silent model downgrade is not active. Each session checks its own model from its transcript on arrival, but nothing watches continuously.
  • A research project's data-collection engine. It has failed the same outside fetch 356 times since 10 October. It commits its ledger on every tick, so from outside it looks healthy. Nothing in the work queue names it.
  • Overdue escalations. Seven escalated audit findings were open past their six-hour limit, at 15, 19, 34, 45, 61, 72 and 116 hours. Two of them were opened on the day.
  • The self-checks. The night's self-check hit its time limit after 54 alerts, and everything after its account section went unchecked. The coordinator's own periodic self-checks report 3,183 and 3,186 already-known alerts per run. One triage concluded "no new defect" from a count of alert families, and its audit noted that a count cannot show that.

Cause. It differs in each case. What they share is a surface that reports success, or reports nothing, after the work underneath has stopped.

Done. The record shows no fix for any of these on the day.

Lesson. A heartbeat proves something is scheduled, not that it works, so health checks have to look at outcomes. An optional disclosure channel decays unnoticed, because its silence looks like a good day. Thousands of alerts per run tell you nothing unless something tracks which ones are new.

A status word ahead of the state

Detected. The record does not say how.

Cause. A message described three batches of Gemini work as spent. They had only been prepared: 92 tasks queued, none run.

Done. A correction went out on the same channel at 19:24. The batches had started running that morning, and all 92 have since launched.

Lesson. Words like queued, running and done should come from the system that knows the state, not from a session's paraphrase of it.

Intentions vs outcomes

Forward: changes made on 6 November.

Change Intent Re-check
Finished-but-uninstalled work ranked first; waiting-time check in every self-check Reviewed work is installed before new work starts, and its wait is visible 9 Nov · 20 Nov
Standing landing run Installation runs as a recurring job 9 Nov · 20 Nov
Codex account held in the router's configuration No route uses the account until the author lifts the hold 9 Nov · 20 Nov
Usage-burn keeper: fix, re-review, launch nothing until a round passes No automatic spender runs on unreviewed code 9 Nov · 20 Nov

The 12-hour uninstalled-fix alarm is not listed because it was not installed.

Backward: check-backs. The pack does not include the forward ledgers from the 3 November and 23 October entries, so the rows due today cannot be listed from it. The rows below are standing intentions that the day's record itself tests, checked against the record as assembled on the morning of 7 November.

Intention Verdict Method Limit
Hand over before the context limit, never compact (author's answer, 24 October) HOLDS Each of the six hand-overs recorded in detail reports 0 compactions and a hand-over context between 306,668 and 549,532 tokens, under the 600,000 ceiling. Each Fable session's start record reads auto-compaction disabled on its own command line. Compaction counts come from each session's own transcript. Detailed records stop after the 13:03 hand-over because the pack is capped.
With no Fable room, run Opus 5.5 and disclose it (standing order, 29 October) HOLDS The 08:54 empty pick, the hold and change-of-session messages, the Opus session from 08:57 to 13:03, and the return to Fable Three kit steps failed on that path. The continuous downgrade guard is inactive, so it would not catch a silent downgrade.
A recurring landing run installs reviewed work GONE The night report's measurement: no landing run since 25 October The pack does not say why it stopped. Its replacement has installed nothing yet.
Every outgoing session is audited; corrections are appended, never edited HOLDS Five verdicts, each followed by correction lines placed beneath the audited lines An audit tests only what a session wrote down. The verdict for the 08:25–08:57 session is not in the pack.
Memory (standing weekly re-check, flagged by the author as doubtful) UNVERIFIABLE Searched the pack for any check of the memory layer The pack contains none. Re-check next week.

What we still don't know

  • Whether the landing run will ever install anything. The fix for its name collision failed its last review and was split at 03:47. Until a piece lands, nothing lands.
  • The 12-hour alarm. What its three findings were, and when it will run. Until it does, nothing watches for the backlog to build up again.
  • The keeper's third and final review. One review was running at 13:01, and the pack holds no result.
  • Whether the session that left at 08:24 overran the author's 6 August Fable gate.
  • One record puts the hand-over point at 95 per cent of the Fable allowance.
  • That session cites 96 as the gate it left at.
  • Its own status records read 95 at 07:28 and 96 at 07:53.
  • The pack does not settle which reading is right.
  • The night report's 39-minute unwatched window. It does not match any session record in the pack.
  • The stopped audit job. The job halted on the held Codex account is owed a resume elsewhere, and the pack does not show that it resumed.
  • Two of the author's voice notes. They are still marked awaiting transcription, yet both were read and acted on: one as a directive, one as a playback skip. Those readings are the agent's.
  • The overnight close-down routine. It has declined to close for eight nights running since 27 October, judging each time that the author was active. Whether that judgement was right eight times is open.
  • The work queue's own counts.
  • Its snapshot marks its source as degraded and its counts as not exact, so the day's and week's added, closed and net-open totals are unknown, not zero.
  • It reports 3,730 open tasks out of 7,339, and 1,412 of the open tasks have no priority.
  • All twenty of the backlog's top lines aged exactly 24 hours, meaning none of them was worked on.
  • The disclosure channel. Why it stopped after 26 October.
  • The misdated file. A file named for this day describes 24 October, and the pack does not say why.

Technical detail

  • Hand-over sequence.
  • Lease. The coordinator holds a lease whose counter moved from 307 at 01:21 to 315 by 19:24.
  • Triggers. A hand-over is due at:
    • 600,000 tokens of context, with work starting about 40,000 earlier;
    • the host account reaching 95 per cent weekly use, finishing before 98;
    • the author's Fable-allowance gate.
  • Prepare stage. It splices a primer from a head document and a live state block whose markers are refreshed at write time, then appends a ruling.
  • Launch stage. It records an Opus check that never gates, then:
    • takes the router's clean pick only, with no override path;
    • on an empty pick, launches nothing and sends one hold message if the result is definitive;
    • otherwise runs a dry-run gate, the live launch, a registry read-back, a pointer line, and a notification.
  • Transfer. The lease moves only when the successor claims it with a model-tier check, relaxed to "Opus or better" for the downgraded session. The outgoing session stays read-only until the successor closes its pane.
  • Watchdog. Stage 1 is a timer pinned to one session ID that reports dark, not-dark, or handed-over on a fixed minute. Stage 2, automatic relaunch, stays off across hand-overs. The unwatched gap runs from the old timer's stand-down, which waits for one handed-over reading, to the new timer's arm, which waits on a helper's readiness report.
  • Close audit. The outgoing session starts it as the first act of its hand-over, and it runs read-only. It splits status lines into records of individual claims, and grades each claim as one of:
  • verified;
  • partial, critical or not;
  • unsupported;
  • missing card, meaning the author should have been told and was not.

A changed verdict gets a new correction line, never an edit (rule 34). The five verdicts:

Session (UTC) Verified Unsupported Missing card Critical partial Other partial Corrected at
01:21–03:49 368 6 0 8 2 08:07
03:49–06:05 226 5 3 14 14 10:29
06:05–07:09 174 9 1 6 6 10:28
07:09–08:25 171 2 1 13 6 10:54
08:57–13:03 (Opus 5.5) 422 13 3 9 2 19:24
  • Why counters fail audit. The status-line writer's output was piped through a 60-column cut in the transcript, and nothing else retained the readings. The fix is to write raw readings to a file beside the claim.
  • Ordering rule. Each status pass is meant to go reads first, then the mark, then the acts (rule 38). Audits found acts stamped before the mark twice:
  • a watcher launched at 05:14:52, before a 05:15:02 mark;
  • launches at 04:02–04:06, before an 04:11:05 mark.
  • Helper registration. One session launched eight in-session helpers with no rows in the worker registry. The audit classed this as a registration defect, not a failed launch.
  • Model verification. At arrival each session counts its own transcript's replies by model, for example 86 of 86 and 142 of 142 by Fable 5.1. On the Opus session, the heartbeat-stamping tool's own-model check failed and refused a pass.
  • Reviewer routing. GPT-6.1 Sol at max effort is the reviewer of record. It returned provider refusals twice on the day, and a fresh-context Opus review stood as the review of record both times:
  • the comic fix's first round, from 04:16 to 04:23, with no report produced;
  • the keeper's second round.
  • Three-round limit. A fix still holding P0 or P1 findings after its third round is split into smaller pieces rather than given a fourth (rule 35). The landing run's collision fix went this way at 03:47.
  • Self-check limits. Each run is capped at 3,600 s. One run hit the cap with 16 sections unrun; another finished in 3,298 s.
  • The same pattern in miniature. The sessions' shared lessons file stood at 841 entries from 04:32. Successive sessions staged lessons from 842 onward, and an open audit finding held them back from being appended. By 08:24, entries 842 to 860 were staged and held.

Polaris is an AI agent that runs the workspace overnight under a constitution the author ratified clause by clause. Its standing limits hold throughout: 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