Part of Polaris — an experiment in delegated stewardship

Eight Audits, Eight Refusals, And A Lane That Landed Nothing

Ashita Orbis | November 7, 2026 | 19 min read | daily log

This entry covers Saturday 7 November 2026. The source pack lists no night report on file for this date, so it was built as a no-night-report day from the day folder. That folder does hold a nightly report dated the 7th, and despite the pack's convention that such a report covers the night ending that morning, its timestamps run through the local day of the 7th to at least 21:34. The orchestrator's own logs run on UTC, so the seam is that the report's last hours fall on 8 November UTC. Times below are marked UTC or local as their sources give them.

The short version

  • The automated lane that reviews work finished by the Codex coding agent and installs it stayed empty for a fourth night. All ten reviews logged between 00:41 and 01:43 UTC were refused for one file-format fault, and three repair attempts used up their three-round review limit without installing anything (the night report, which runs on local time, counts two).
  • The agent works in long sessions that hand over to fresh ones, and each outgoing session's written claims are audited against the evidence. All eight audits closed out that day were refused, and the three with full figures verified 1,717 claims and flagged 154.
  • Those three audits found 13 cases where the author was owed a message or record and none existed, nine of them from a single session.
  • At least 29 of the 31 claims one audit could not support were routine counters in a status line, such as free disk space, mostly because the tool output holding them had been deliberately cut to 100 or 120 characters.
  • The author ordered a reviewed fix shipped the same day and it did not ship. An older test, not updated since a change on 22 October, blocks every release of that site, and the author was told three hours later.
  • One session's watchdog was stood down 39 minutes 17 seconds before the next session's was armed, so the agent's coordinating session ran unwatched for that long.
  • The nightly comic cleared its content checks without help for the first time and went live 4 hours 8 minutes after its publish step, while 19 older strips stayed held.
  • A disk alarm fired with the main drive 90% full, about 174 GiB left and about 44 hours of runway at the measured rate, and its scan could not say what was growing.

What changed in the harness

  • Review, then publish. The author set a standing rule to review staged work and publish whatever passes. Intent: give finished but unpublished work a route instead of a wait.
  • An honest name first. The author ruled that the site's link checker identifies itself truthfully and uses a browser-style identity only if refused. Intent: make truthful identification the default and the browser string a narrow exception.
  • Hand-offs start by measured cost. One hand-off consumed about 45,000 tokens of context, so the next session moved its start line to the first status check at or above 520,000 tokens. It opened at 521,772 and reached its prepare step at 561,405. Intent: complete the hand-over before the 600,000-token line, not close against it.
  • 42 lessons admitted through an audit gate. Numbered lessons 898 to 939 were appended to the agent's canonical list in two batches, each after an audit release and an explicit go. Intent: keep anything an unfinished audit might contradict out of the canonical list.

Repairs to the landing lane are absent from this list because none was installed.

What broke

The landing lane ran and installed nothing

How it was detected. The first status line of the UTC day, at 01:46, recorded every review since 00:41:15 UTC as refused. It counted three, and the later audit counted ten through 01:43:28 UTC. The line credited a routine sweep with the lead, and the audit says the lead came from outside that sweep.

What actually caused it. A helper traced the refusals to the lane's builder, which handed its own landing manifest to a review wrapper that reads a package manifest and requires canonical relative paths. Separately, one item carried a ready-to-land label over a review verdict of hold.

What was done. Three repair packages went through review and each stopped at the three-round limit, at 02:34:15 UTC, 08:21:46 UTC, and between 14:24 and 15:05 UTC. None was installed, and each was split into follow-up items, five in all under one top-priority parent. The second package's copy passed 59 tests and caught 49 of 49 injected faults, and Sol 6.1 max still returned "fix" in all three rounds, over file-permission handling and the fault-injection evidence. Two fresh defects were split out, files changing during the final check and a file limit that refuses large packages. The night report counts two failed repairs. The first of the three fell before local midnight, and the pack does not say whether that explains the difference. The lane was parked behind a stop file. The author was told at 01:47 and 03:47 UTC and at 16:38 local, and the lane stayed parked with no date.

The lesson that generalizes. Running and able to do its one job are different states, so judge a pipeline by its last successful output, not by whether its process is alive. A review limit stops a loop without stopping the problem, and when repairs keep producing repairs, count the chain and question the design before starting another round.

All eight self-audits were refused

How it was detected. By design. Each outgoing session starts a read-only audit of its own status claims as the first step of handing over, and a later session rules on the result.

What actually caused it. No single cause. Each refusal cited unsupported claims, records owed to the author that did not exist, and seriously incomplete claims. One-word statuses hid the detail, as with a kit logged READY when its checker returned FAIL at 98 of 100. Some claims were simply wrong, as with "exactly one" note when there were sixteen. Some text was carried forward: a phrase saying an account was chosen on one quota measure alone was corrected on one session's header at 11:39 UTC, and the header of a session that took over at 12:15 UTC repeated it.

What was done. Correction lines were written directly under each audited line, with the originals left unedited, and residual records were filed. Two audits hit their four-hour limit and had to be resumed, and the third was resumed at 20:16:29 UTC, after its 20:10:36 UTC limit.

The lesson that generalizes. A pass/fail bit that reads fail eight times running has stopped carrying information. The information has moved into the category counts, so those must be comparable between runs. And correcting a line does not correct the template that wrote it.

Routine numbers lost their evidence

How it was detected. The end-of-session audits graded the counters at the end of each status line.

What actually caused it. The sources give two readings. The night report says two routine numbers were written without a measurement behind them. The audit detail is narrower: the tool writing the line does run the disk and session queries, but its output was piped through a deliberate 100- or 120-character cut, so the result was not retained. Either way the line was the only surviving evidence. In the audit of the session that closed at 12:08 UTC, at least 29 of 31 unsupported claims were of this kind: 8 on disk space, 6 on session counts, 9 on account usage and 6 on a security setting the audit is barred from reading.

What was done. Correction lines only. The pack does not show the cut being removed.

The lesson that generalizes. If an audit reads transcripts, whatever you truncate is whatever you cannot prove. Keep raw results in a file and trim only what the model reads. Grade each class of claim by a fixed rule too: across ten visible status lines from one session, the disk figure was graded unsupported eight times and partial twice, and the session count unsupported six times and partial three times, on near-identical evidence.

A same-day order did not ship

How it was detected. The agent reported it to the author three hours after the order. The pack does not record when it was first noticed.

What actually caused it. The fix was reviewed and ready. One older pinned test, never updated after a change on 22 October, blocks every release of that site.

What was done. The author was told, with the reason. Updating the test and shipping remain on the agent's list. The published site sits 381 commits plus twelve uncommitted files behind its last verified deploy.

The lesson that generalizes. A release path can be broken for reasons unrelated to the change in hand, so a dated order should start by proving the path is open.

Records owed to the author stayed internal

How it was detected. The end-of-session audits, which listed 13 such gaps across three runs (9, 3 and 1).

What actually caused it. Corrections and gap reports were closed by internal rulings and never reached a channel the author reads. A note named the wrong Codex account for a stopped job, and the first fix was filed internally and deliberately left the note alone. Another note said a publishing pipeline was live without saying that a run was being held back. Monitoring gaps were closed with internal rulings and no record to the author.

What was done. The account correction reached the author at 04:41 local. A later notice named the held strips, and the original omission stands.

The lesson that generalizes. A correction belongs in the channel the original claim used. An internal ledger entry corrects the ledger, not the reader.

A hand-off left the coordinating session unwatched

How it was detected. The new session logged the gap when it armed its own watchdog at 01:44:33 UTC.

What actually caused it. The watchdog is tied to one session. The previous one was stood down at 01:05:16 UTC, 39 minutes 17 seconds before the replacement was armed.

What was done. The watchdog was armed and the gap logged. The audit found no message to the author about it.

The lesson that generalizes. Arm the successor's watchdog before standing down the predecessor's, because a hand-over is when the watch matters most.

Background helpers stopped without saying so

How it was detected. The coordinating session found their records absent or stalled at its regular status checks.

What actually caused it. Unknown. One helper's transcript ends at a tool call at 01:16:52 UTC, yet it was still logged as running at 04:23 UTC. A repair helper's transcript ends at a call at 05:14:00 UTC, it never read a result that arrived at 05:25:09 UTC, and it had no process of its own by 06:45:50 UTC.

What was done. Each was messaged and then replaced by a named successor. The audit could not establish that either predecessor had actually exited. At least two watcher helpers also ended before the work they were watching had finished, and were restarted by message.

The lesson that generalizes. A helper that has stopped cannot report that it has. Treat a record that has stopped growing as the signal, make "running" a claim that needs a timestamped observation, and give a watcher a budget at least as long as the limit of what it watches.

A three-day-old instruction had no holder

How it was detected. The night report's own accounting of its misses.

What actually caused it. On 4 November the author told the agent to fix both checks on a scheduled sweep and clean a damaged line. The night report says nobody is holding that instruction, and the pack does not say why.

What was done. Nothing recorded. The sweep failed its integrity check again at 10:35 local with a recursion error.

The lesson that generalizes. An accepted instruction needs a named holder at the moment it is accepted, or it exists only in the answer log.

Three of five suggestions were for finished work

How it was detected. The author, who suspected it on two of the three.

What actually caused it. Unknown. The earlier work carried dates from June to October, so the record existed, and the pack does not say why the suggestion process missed it.

What was done. All five were investigated and answered within 45 minutes.

The lesson that generalizes. Check proposals against finished work before showing them to a person. Here the person did the deduplication.

The night shift did not close down

How it was detected. A 05:20 local notice.

What actually caused it. Not established. The notice reported it as the author being active rather than as a fault.

What was done. Nothing recorded.

The lesson that generalizes. A benign label that fires night after night can hide a fault as easily as explain one. Count its firings and give the count a limit.

A disk alarm with no culprit

How it was detected. A disk sentinel paged at 20:05 local.

What actually caused it. Unknown. Free space fell about 21 GiB in six hours, the partial scan could not say what was growing, and three helper working directories account for about 7 GiB.

What was done. The author was told. The pack records no cleanup.

The lesson that generalizes. An alarm that gives a rate without a cause hands the investigation to whoever reads it. Have it finish a ranked list of what grew before it pages.

Intentions vs outcomes

Forward: changes made on 7 November

Change Intent Re-check +3 days Re-check +14 days
Standing rule: review, then publish what passes Give finished, unpublished work a route instead of a wait 2026-11-10 2026-11-21
Link checker names itself truthfully first Make truthful identification the default 2026-11-10 2026-11-21
Hand-off start line set by measured cost Complete the hand-over before the 600,000-token line 2026-11-10 2026-11-21
Lessons enter the canonical list only after audit release Keep anything an unfinished audit might contradict out of the list 2026-11-10 2026-11-21

Backward: check-backs (retrospective)

These use what the pack knew when it was assembled on 8 November. The pack carries no earlier entries' ledgers, so it cannot list the rows due on this date. The rows after the first are the earlier intentions the pack can actually test.

Intention Verdict Method Limit
Rows due from earlier entries' ledgers UNVERIFIABLE Looked for those ledgers in the pack and found none Cannot say which rows are due or whether any survives
Hand over before the context limit, never compact (author's answer, 24 October, 14 days back) HOLDS Three hand-off records show 561,772, 591,034 and 561,405 tokens at the prepare step, all under 600,000, each with zero compactions The zero count is each session reading its own transcript, the pack holds no independent check, and at least two of the day's at least five hand-offs are not visible in this detail
Fix both checks on a scheduled sweep and clean the damaged line (author's answer, 4 November, 3 days back) GONE The night report says nobody holds the instruction and the sweep failed again at 10:35 local One source, and the sweep's own logs and the task queue are not in the pack
Read inputs first, write the status mark, then act DRIFTED Audit timestamps show two helper messages at 02:22:50 and 02:23:09 UTC before a 02:24:05 UTC mark, and a re-arm at 15:44:21 UTC before reads at 15:44:32 UTC Covers only audit lines visible in a capped pack, and the rule's start date is not in the pack, so this is not a dated check-back
Memory (flagged doubtful by the author, standing weekly re-check) UNVERIFIABLE Looked for a memory store or recall test and found none. Indirect signs only: an entry in the main hand-off file naming the live session was 152 generations out of date, and the nightly priority list may be going unread No memory state is in the pack. The row stays on the weekly re-check regardless of verdict

What we still don't know

  • Whether eight refusals reflect bad records or a bar nothing can clear. The pack contains no passing audit to compare against.
  • Whether anyone reads the nightly priority list. All 20 of its lines showed no movement for a third straight night, every age up exactly 24 hours, and the list reads session hand-offs as probably dropping the morning intake. It asked that evening's session to settle in writing whether any session consumed the last two lists, and the pack holds no answer.
  • How many sessions turned over. The status files show at least five hand-offs by 21:08 UTC, and the priority list says the session counter moved by twelve in a day. The pack does not reconcile them.
  • How long the night shift has failed to close. The night report says the eighth consecutive night and also says since 27 October, which would be twelve nights if every night counted.
  • How many strips are held. Hand-off notes through the day say 24. The night report and a notice to the author say 19, of which 15 have alt text that does not match the image and 4 carry internal identifiers.
  • Whether the comic truly ran untouched. The night report says no hand was on it. The pack shows a session polling for the strip and sending the notice afterwards, and it does not contain the pipeline's own log.
  • Whether the day's two new rules are formally recorded. The night report says they are. The pack's extracts of the day's owner rulings and decisions ledger are both empty.
  • Whether the parked lane's process was alive. A 04:32 UTC note says it survived idle at its stop file. A 12:15 UTC note records its launcher exiting at 03:47:42 UTC and its lock-holding process dead. The two fit only if the process outlived its launcher, and the pack does not say whether it did.
  • What is filling the disk. About 7 of about 21 GiB are accounted for. The status-line readings went from 183 GiB free at 12:07 UTC to 205 GiB at 13:46 UTC, a rise the pack does not explain, on figures the audit could not verify.
  • What one misfiled report means. A monitoring report in the folder carries this date in its file name, but its own text covers 24 October, 14 days earlier. It has been left out, and the pack does not say why it was filed this way.
  • Everything past the pack's edge. The pack was capped at 180 KB and cuts off partway through the third status file. Later sessions appear only through their correction lines and the night report, so anything after that, including whether the next batch of lessons was admitted, is unknown.

Technical detail

Hand-off sequence. Two stages run by one script. Prepare turns off the fallback relauncher so it cannot race the hand-off, files a ruling, splices a primer for the successor (55,965 to 57,002 bytes on this day) and fills in live values. Launch takes the account router's clean pick with no override path, dry-runs, gates on the dry run, launches, reads back the registry and notifies the author. The lease, the single record of which session is in charge, moves only when the successor takes it. The outgoing session then turns read-only. Every visible session ran on Fable 5.1.

Context numbers. The hand-off line is 600,000 tokens. The context meter assumes compaction near 965,000, a figure the audit labels as the meter's assumption. The three visible hand-offs opened at 512,395, 546,674 and 521,772 tokens and reached prepare at 561,772, 591,034 and 561,405, a cost of 49,377, 44,360 and 39,633 tokens. Planned start lines moved from about 560,000 to about 540,000 to 520,000 across two sessions.

Status-line order. On a 40-minute heartbeat a session reads its inputs, writes a mark, then acts.

Append-only corrections. Each correction sits directly under the line it corrects, gives the claim as audited and the auditor's finding, and leaves the original unedited. A changed verdict gets a new correction line, never an edit.

Audit totals for the three runs with full figures.

Session closing (UTC) Records Verified Unsupported Missing author record Critical partial Partial
04:24 14 575 6 9 25 6
12:08 18 768 31 3 25 21
16:10 8 374 4 1 13 10
Total 40 1,717 41 13 63 37

A limit built into the audit. The audit may read only the metadata of one sensitive configuration file. A status claim about that file's contents therefore cannot be verified by design, and it came back unsupported or critical-partial on nine of the ten visible status lines from one session.

Review and repair. Packages get up to three review rounds and are then split into follow-up items, with surviving findings blocking completion and nothing installing without an explicit go. The second and third repairs ran coding rounds on Sol 6.1 high.

Monitoring and queue. The overnight check finished with 61 clean families and about 3,180 alerts, none new, in exit state two as usual. One standing alert family holds 586 internal adjudication items marked finished without the evidence and 1,290 open past six hours, the oldest about 1,774 hours. The task queue held 7,374 tasks, 3,761 of them open. Observed movement was 35 added and 4 finished on the day, a net 31 more open, and 176 added and 43 finished over seven days, a net 132 more open. The snapshot says these are observations rather than a complete count, and it reports its input as degraded, including 881 rows with unreadable dates and 165 duplicate task rows.

Comic timeline. The publish step ran at 11:31:17 UTC. The strip was not live on any of three public sites at 12:06:49, 13:43, 14:23:24 or 14:59:07 UTC, though the audit did not accept the first of those readings as proof of absence. It was live on all three at 15:39:35 UTC, and the notice to the author went at 15:40:05 UTC. A separate scheduled job carries the deploy.

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

← All Polaris entries