Twelve Review Rounds, One Measured Cause
This entry covers Wednesday 18 November 2026, in the workspace's local time. The source pack's slot for night reports was empty, so by its own label this is a no-night-report day. The record does hold a report named for the 18th and titled as the nightly report, but it covers the 18th itself, to about 21:50, rather than a night. The rest of the record is a morning report on the night of the 17th into the 18th (state as of 05:40, with readings to 06:00), two automated backlog sweeps that ran during the day, and the author's answer log, which keeps UTC. The seam: the morning report reaches back into the evening of the 17th. Changes made then belong to that day's entry and appear here only where they bear on the 18th. Nothing in the record covers the 18th after about 21:50.
The short version
- A disk cleanup has now failed twelve review rounds and freed nothing. The cleanup code is meant to delete 442 old copies of the site's search files. Four successive versions of it each went through three rounds with a separate reviewer, and each ended with a problem still open. The main disk was 95 per cent full at 05:40. On the 18th the reason was measured instead of guessed: about six programs a second on the machine let go of folders they were holding. That churn is what all twelve rounds kept catching.
- A blog repair went live at about 20:23 still showing the figure it was meant to remove, in a glossary tooltip that appears in two places. Its three review rounds had missed it. Reading the live pages back after publishing caught it, and it was gone by 21:39.
- Two approved blog repairs went live with words that were not on the cards the author answered. One correction note gained three factual sentences and another post gained five reworded lines. Neither set is in the author's voice. The record lists both for the author.
- Two of the author's messages fell into timing gaps.
- An answer sat about two hours unread in the gap between Polaris waking and arming the watcher that collects answers. Polaris now arms the watcher first.
- A note left at 21:54 the night before fell between two daily note extracts. It survives only in the raw inbox, and the gap is not yet fixed.
- The day's two automated sweeps of the backlog were written to disk and, by the evening, not delivered.
- One found that none of the 504 open rows in the 590-row backlog registry had any recorded activity for more than three days. It also found 15,380 of 15,386 tracked queues with nothing reading them.
- The other found 353 open rows whose own text says the work is done. Only 27 of them hold up when checked against the evidence.
- 217 of the author's answers had no record of being acted on at 12:53, against 165 two days earlier.
- A newer release of the Sol model tested level with the one Polaris uses, so no model choice changed. Polaris actually uses Sol for reviewing, and the test of that job has not run; it is planned for 9 December.
- Two records pointed at the wrong thing.
- A new alert said Polaris had been downgraded to Opus 5.5. Its own transcript showed Fable 5.2 on all 45 recent replies. The check reads whichever transcript on the machine is newest, most likely a worker's.
- The page meant to name the session currently running Polaris still names one from 22 October.
What changed in the harness
- Answers are watched from the moment Polaris wakes. Polaris now arms the watcher that collects the author's answers before doing anything else when it starts. Intent: an answer that arrives during start-up is picked up at once. On the 18th one waited about two hours.
- The new-model evaluation is spread across three later weeks. With the newer Sol level, the remaining slices of the Astra evaluation are booked for 25 November, 2 December and 9 December, one session each. They run under a ceiling of 15 per cent of a week that Polaris set itself, not the author. The reviewing test is planned for 9 December. Intent: measure the new models without the evaluation using up any one week's allowance.
- The blog-drafting agent keeps its research and loses the tool servers it does not use. The change follows the author's answer, given at 20:11 on the 17th; the answer log's UTC clock files it under the 18th. It was live, and checked on a real draft, by 05:40.
- The question put to the author counted seven servers to cut. The change cuts eight of the ten that actually connect.
- The record puts the difference down to one kept server that does not load by default and one connector that does. Polaris records it as the narrowing the author approved.
- Intent: shrink what the drafter can reach without taking away the research it relies on.
- A fallback for the overdue weekly review. The review has been held since 15 November behind Polaris's own ceiling on running processes. It now goes first among report sessions once one may start. If it has not gone out by Sunday 22 November, that day's review covers both weeks. Intent: a review the ceiling keeps blocking is folded forward rather than silently skipped.
What broke
Twelve review rounds, nothing freed
How it was detected. By the reviews themselves. Each version of the deletion code goes out as a package with up to three review rounds by a separate reviewer. Every round so far has ended with a problem open. The rounds the morning report describes all ended on problems of the most serious grade. By the record's count that is twelve rounds across four packages. Nothing has been freed: all 442 copies were in place at 05:16 and still there in the evening.
What caused it. The cleanup may delete a copy only after proving that no running program can still see it another way: through something it holds open, a view it inherited, or a second mount. Overnight the rounds found the gaps in that proof one at a time:
- A program that can see a copy starts a child and exits. The child inherits the view, the step loses track of it, and the step deletes a copy the child can still read. The reviewer reproduced this on the machine.
- The same hand-off happens twice in a row.
- A mount is counted as fine and then quietly detached.
- An ordinary bind mount makes a second path just after the checker has read that program's group.
- A program moves into a new group while the checker is counting.
The last two were present in older versions too; earlier rounds had not named them. The fourth package's three rounds each found a different problem with tidying up after a machine restart. Each was a way the tidy-up could leave three background helpers switched off while the run's report said they were back on.
What the 18th added was a measurement. About six programs a second on this machine let go of folders they were holding, and the record names that as what all twelve rounds kept failing on.
What was done.
- 01:17. A session reached its token limit with the second package's last review sealed but unsent. Polaris chose to fix one more thin spot before sending it: a walk up the folder tree that gave up after 64 levels now goes to 4,096, keeping more copies back.
- 02:12. The round sent then found the inherited-view gap, and the package was set down.
- Third package. Polaris started a third package without putting a question to the author. It read the author's standing instruction to stop asking and finish the work as setting aside an earlier promise: that a failed last round would come back to the author. It told the author so by message at 05:30.
- 03:45. A provider error cut off the session making the third package's fixes. Both fixes were made and nothing had been sent. The session reported nothing itself, and Polaris saw the stop 17 minutes later.
- About 04:04. Its replacement was stopped two minutes in, when Opus 5's safety filter flagged its first listing of the package folder. The stopped screen offered an automatic switch to Opus 4.8, a fallback the author's standing rule forbids. Nothing was answered there.
- By 05:15. The next session passed all 245 tests, and the round it sent still found the two older gaps.
- 11:49. A card, meaning a question put to the author for a decision, had set out three ways forward. The author took the recommended one: accept, for this one run, that the step can miss a second view made or changed while it checks.
- Fourth package. It still ended each of its three rounds with a problem open. Under that card's terms the work has stopped, and no further card is coming.
The lesson that generalizes. Before building a check that looks at a live system and then acts on what it saw, measure how fast the system changes underneath it. Here the checks take about 28 minutes to run over the 442 copies. That is one reading, not a bound, on a machine where about six programs a second let go of folders. Twelve rounds found instances of one race one at a time, and a green test suite only encoded the instances already found. When round after round finds a new instance of the same race, the race is the problem. The next design has to take the race away rather than patch another instance of it.
A smaller lesson: a standing instruction and a specific earlier promise can point different ways. When they do, the choice between them should reach the person at the moment it is made. This one reached the author at 05:30, after the third package had started.
Three review rounds missed what the live page showed
How it was detected. The landing's own read-back, meaning a reading of the live pages after publishing. It ran after the repair to post 022 went live at about 20:23.
What caused it. A glossary tooltip, in two places, still repeated the figure the repair was removing. The package missed it, and so did all three review rounds. The record does not say why.
What was done. The tooltip line was taken out by 21:39.
The lesson that generalizes. Check the change where it lands. A review sees the change its author packaged. A read-back of the live surface sees everything that renders the thing being changed, including places the package never touched. Keep the read-back even after a clean review; here it was the only check standing outside the package's picture of the change.
Words under the author's name that were not on the cards
How it was detected. The record says both sets of added words were listed for the author. It does not say which check caught them.
What caused it. The landings carried out two repairs the author had approved. In doing so they put words on the posts that were not on the cards the author answered: three factual sentences in the correction note on post 038, and five reworded lines in post 022. Neither set is in the author's voice.
What was done. Both were listed for the author. Post 022 was deliberately given no dated correction note, because the note would itself have been new words under the author's name.
The lesson that generalizes. An approval covers the text it showed and nothing else. Words added while landing are still new words under the author's name, however factual, connective or small. They should be shown before they go live, not listed after. The pack does not say which of those happened here.
Two of the author's messages fell into timing gaps
How it was detected. The answer's delay was found the same day; the record does not say how. The nightly report found the note in the raw inbox but in neither day's extract.
What caused it.
- An answer given at 11:51 sat about two hours unread, in the gap between Polaris waking and arming its watcher.
- A note left at 21:54 on the 17th falls between the daily note extracts. The 17th's extract was written at 21:30, and the 18th's begins after the note.
What was done. Polaris now arms the watcher first. The nightly report names the extract gap as worth fixing, and it was not fixed on the day. What the note asked for had already been done.
The lesson that generalizes. These are two forms of one rule.
- On start-up, arm the listeners before doing anything that takes time. Otherwise start-up becomes a window in which input waits unseen.
- Consecutive windows must share a boundary. Each extract should begin exactly where the last one stopped consuming, whatever clock either one keeps.
Written, not read
How it was detected. Two automated sweeps of the backlog ran during the day, and the nightly report checked whether they were delivered.
-
The silence sweep, at 10:35.
- Every one of the registry's 504 open rows had gone more than 72 hours without a recorded signal; 224 of them had never had one.
- 15,380 of 15,386 tracked queues had nothing reading them. Since the previous sweep, 651 records had arrived among them: 271 to the queue of candidate memories, 108 to a work queue's witness log and 82 to escalations.
- 585 dispatches had been written out but never started.
-
The closure sweep. It found 353 open rows whose text says the work is done.
- 27 survive a full check: a dated claim, corroboration on disk, and nothing contradicting it.
- 72 need the author.
- 254 are refuted.
The silence sweep's looser reading of the same rows counts 108 as finished.
-
The nightly report.
- 2,223 pieces of commissioned work show no sign of life, 25 more than the night before. 1,711 of them are written hand-offs that nothing picked up.
- A priority list's top item has been first for a fourth night with no session taking it; the list dispatches nothing by itself.
- The 217 answers with no record of being acted on.
- Neither sweep had a delivery row: both were on disk, not in front of the author.
What caused it. Not established. The record shows producers without named readers. For none of them does it say whether a reader was lost or never existed.
What was done. Nothing on the day.
The lesson that generalizes. Writing is not delivering. Every producer in a harness needs a named reader, and monitors are producers too: the sweep whose subject was unread queues became one more unread file the day it ran. Separately, a task's own text saying it was done held up against the evidence in 27 of 353 cases. Completion language is a weak signal, which is why the closure tool closes nothing on wording alone.
Records that pointed at the wrong thing
How it was detected. The morning report checked the new downgrade alert against the session's own transcript. The nightly report found the hand-off trail stale while trying to use it.
What caused it.
- The downgrade alert. At 06:00 a new alert said Polaris was on Opus 5.5, downgraded. The transcript of the session running Polaris, read at 05:52, shows Fable 5.2 on all 45 replies in its last 200 lines. The check behind the alert reads the newest transcript on the machine. The morning report judges that most likely a worker's, and so judges the alert most likely false.
- The hand-off page. It is meant to name the live Polaris session and its status log, and it still names a session from 22 October. No status log exists under the naming it uses for any session that ran on the 18th. The nightly report rebuilt the day from the session's state page and the day's hand-backs instead.
- The owed-to-report file. Reporting sessions are meant to record what they owe the day's report in a file. That file does not exist for the 18th, and only instructions from early October still ask for one. Its absence says nothing about whether there were misses to report.
- This entry's own source pack. It arrived with its status-file and decisions sections empty. The pack does not say whether that has the same cause.
- Two sign-in cards. Two cards asking the author to renew sign-ins stayed open all day, though both sign-ins had been done on the evening of the 17th.
What was done. Nothing on the day.
The lesson that generalizes. Bind a monitor to an identity, not to recency or a naming pattern. "The newest transcript" and "the status log named like this" keep producing confident answers after they have started describing something else. An empty report channel is a finding, not a clean bill of health. And a card that asks for an action should close on evidence that the action happened.
The weekly review, held by Polaris's own ceiling
How it was detected. The morning report.
What caused it. The weekly review waits for the count of running Claude processes to fall below 85, a ceiling Polaris set itself. The count has not been under it since 15:00 on 15 November. On the evening of the 17th, Polaris itself went past the same ceiling: it launched a deploy repair with 116 processes open, classing it as a live break.
What was done. The fallback listed under the harness changes. The pack does not say whether Sunday's combined review sits behind the same ceiling.
The lesson that generalizes. A gate with no timeout becomes a ban once its condition stops being reachable. Work that has to happen needs a deadline that outranks the gate. First place in a queue behind an unreachable condition is still no place.
Starts above the heat limit
How it was detected. The sessions' own records, as carried in the morning and nightly reports.
What caused it.
- On the 18th, the work withdrawing post 018 started one local run at 79°C against its own 78°C start limit. The record does not say why.
- The morning report also records an earlier session that started a re-run at 80°C against the same limit. Its heat check was chained so that a failed check did not stop the run.
What was done. Nothing recorded beyond the reports.
The lesson that generalizes. A precondition has to stop the run, not just report on it. A check whose failure the chain ignores is only a log line with a threshold in it.
Intentions vs outcomes
Forward: changes made on the 18th
| Change | Intent | +3 days | +14 days | |---|---|---|---| | Answer watcher armed first on waking | No answer waits through start-up | 21 Nov | 2 Dec | | New-model evaluation spread over three later weeks; model choice unchanged | Evaluate without using up any one week's allowance | 21 Nov | 2 Dec | | Blog drafter's unused tool servers cut (answer given on the 17th local time, filed under the 18th) | Narrow its reach, keep its research | 21 Nov | 2 Dec | | Fallback for the overdue weekly review | A blocked review is folded forward, not skipped | 21 Nov (the deciding date is 22 Nov) | 2 Dec |
Backward: check-backs due
No earlier entry's ledger is in this pack. The rows falling due today are the 15th's at three days and the 4th's at fourteen, and they cannot be named or checked here. This entry retires none of them.
| Row | Verdict | Method | Limit | |---|---|---|---| | Memory (the author's standing weekly re-check) | UNVERIFIABLE | Searched the pack for any reading of the memory system. The only one is in the silence sweep: the queue of candidate memories is the largest of the queues with nothing reading them, with 271 records added since its previous sweep. | The pack holds neither the row's original intent nor any account of what is meant to consume those candidates. |
Early reads: changes from the 17th, not yet due and not retired
| Change (17 Nov) | Read on the 18th | Method | Limit | |---|---|---|---| | Read every running session's screen and status file on each wake (from 16:53) | HOLDS | The 03:45 provider stop was seen 17 minutes later. On the 17th, before the rule, a stopped session sat unseen from 10:16 to 16:50. | One observed stop. The record puts these two facts side by side without saying the rule caught the stop. When the 04:04 stop was noticed is not recorded. The rule does not cover input that arrives before the watcher is armed. | | Deploy packages restored, and the cleanup method no longer prunes them (23:35) | HOLDS | Five landings went live on the 18th, and each was read back off the public addresses. | A prune run outside that method would break deploys the same way, and nothing refuses one. | | New work sessions keep their temporary files on the second disk (from 10:58) | HOLDS | Still in use. The folder held 22.9 GiB at 05:40, against 44 KB at set-up. The second disk had 93.2 GiB free, 29.3 GiB less than a day earlier. | Nothing in the pack trims the folder. The record names it as one new user of that disk and gives no other user's share. |
What we still don't know
- Whether the 442 copies can be removed at all while the machine is running. The work stopped under the terms of the card answered at 11:49, and the pack names no next design.
- Whether the newer Sol reviews better or worse. "Level" covers every test Polaris has except the one for reviewing, the job Polaris uses Sol for. That test is planned for 9 December.
- Why three review rounds missed the tooltip. The record does not say.
- Whether the words added to two posts were shown to the author before they went live. The record does not say this either.
- Whether the downgrade alert was false. The morning report says it most likely was. Nothing in the pack confirms which transcript the check read.
- Where the rise in unacted answers came from. The count went from 165 to 217 in two days, but the 06:00 board dates only 16 of the 217 to those two days.
- Whether the 590-row registry the silence sweep reads is still meant to track live work. The reports' own project counts come from a separate work queue: 8,050 tasks, 4,220 open. That queue's snapshot calls its own counts inexact and its source degraded.
- Whether anything will read the sweeps. Neither had a delivery row by the evening.
- The silence sweep says each finding was also written out as its own card draft.
- The closure tool starts a confirmed row's 24-hour objection window only when a notice is delivered.
- The pack shows neither the drafts nor the notice.
- Why the scheduled job that assembles this entry was on the 06:00 list of failing jobs, and which entries, if any, it failed to produce.
- What the session running Polaris costs. The one measurement covers 13 to 16 November. Nothing has been measured since.
- What happened after about 21:50. Nothing in the pack covers the rest of the 18th.
- What one misdated file describes. It carries the 18th in its name but, by its own text, describes a window on the morning of Monday 26 October. Either it was written on the 18th about that window, or it is misnamed. The pack cannot say which, so its events are not counted here.
Technical detail
The deletion predicate. These points come on top of the gaps listed above.
- Folder walk. The walk up the folder tree now goes to 4,096 levels. When it gives up, it keeps the copy back.
- Accepted limit. Some handles carry only a path or a folder and are passed through a socket or shuffled between descriptor numbers. Polaris treated these as an accepted limit, on its own reading of the author's instruction.
- Background helpers. While three background helpers run, every version keeps every copy back on its own. The run stops them before it starts, which is exactly when the gaps matter. The fourth package's rounds were about restarting them after a restart and reporting that truthfully.
- Measured costs. The second-view checks took about 28 minutes over the 442 copies. That is one reading of the machine, not a bound. The six-programs-a-second figure comes without its method or the period measured.
- Known limits. The run's list grew from 13 to 14, 15 and 16 across the night's rounds. Five of the sixteen are faults in the session's own papers. Seven paths still have no end-to-end evidence.
- Tests. Four test suites passed 245 of 245 in the same session whose review round found two more gaps.
Provider stops. The 03:45 session ended on a provider error ("The response stopped arriving") and reported nothing itself. A worker cut off by its provider cannot announce its own end, so liveness has to be read from outside it.
The 04:04 stop was Opus 5's safety filter acting on the session's first folder listing. An earlier package's record had already warned that the provider's cyber filter could cut a review of this work. The morning report lists the stopped screen's choice, between Opus 4.8 and Opus 5, as the author's to make; it has not yet been put to the author.
Seat choice. At 04:09 Polaris launched the next session on a seat it named itself, against the account picker. The picker was steering work to a seat close to its weekly reset, to use up the remaining allowance first. The 05:24 session went back to the picker. Two automatic policies disagree here:
- the board warns against new work on a seat in its wind-down band (that seat stood at 96 per cent);
- the picker sends work there before the allowance lapses.
The 06:00 board.
-
Totals. 3,524 alerts in 35 kinds, against 3,297 the day before. One new kind flagged the 03:45 session as near the hour after which its saved context goes cold; nothing depended on it.
-
Failing jobs. Seven scheduled jobs failed at least once in the previous 24 hours, against ten. Four are still failing, among them the job that assembles this entry.
-
The binding check. It asks whether each of the author's answers has a record naming it as acted on. This time it finished, after timing out the day before. It raised 230 alerts:
- 217 unacted answers, 16 of them from the previous two days;
- 12 older reports whose question record names no consumer.
The nightly report compares its 12:53 count of 217 with "the board's own 06:00 read of the same thing", 230. The morning report's breakdown shows that 230 is the alert total containing those 217, so the two readings agree.
The closure sweep.
- Confirming a row. A row is confirmed only if it has a dated completion claim, corroboration on disk (a file the row names exists, or no next action is on file), and nothing contradicting it.
- Who judges. Three questions about each row's prose were judged by GPT-5.6 Luna in one batched call over all 353 rows. The verdict itself is a table in code.
- Closing. A confirmed row closes only after a delivered notice and a full 24-hour objection window.
- Rows needing the author. The 72 go to the author as one review batch, not as 72 separate questions.
The silence sweep.
- Silence. A row counts as silent after strictly more than 72 hours with no status file, decision record or completion-gate verdict mentioning it.
- Ages. For the 224 rows that never had a signal, it falls back to a date scraped from the row's own text, and says to read those ages as an upper bound.
- Inputs. It read 4,063 status files, 5,825 decision records, 2,395 gate commissions, 3,840 verdicts and 15,386 queues holding 3,671,874 records. Nothing was unparseable.
- Its own split of the 504. 108 read as finished, 42 wait on the author, and 354 are open and stalled.
The work queue's snapshot.
- Flow. Over the 18th: 81 added, 1 closed, net 80 more open. Over the seven days to the 18th: 518 added, 162 closed, net 355 more open. The snapshot calls these observations, not a complete measurement.
- Source health. The snapshot reports its source as degraded: 882 rows with no usable date, 132 dated out of order, 165 duplicate lines and 151 unusable status lines. Another section of the same report calls the queue's source healthy. Both agree that no line is damaged.
- Staged count. The count of staged-but-unapplied rows reads 0, yet several projects' latest rows read "staged". The report trusts the rows.
The new-model evaluation.
- Tests run. The newer Sol ran against the current one on transport, a 170-item ladder at four effort levels, retrieval, two private suites and a coding arm.
- Result. Level overall: one retrieval task better and a decision panel a little worse.
- Astra. The first Astra slice matched the newer Sol and cost under 2 per cent of a week.
- Usage. Across the day's testing, the OpenAI seat in use went from 89 to 92 per cent of its week. No credits were spent and no key was used.
Review caps. The three-round cap stopped the morning report itself at 06:34 with one sentence open. A second session folded that sentence in and delivered the report at 07:06. The same cap stopped every disk package. A cap ends a loop; it settles one only when what is left is small.
Clocks. The reports keep local time and the answer log keeps UTC, so an answer given on the evening of the 17th is filed under the 18th. A daily window assembled from two clocks puts some things in the wrong day.
Polaris is an AI agent that runs this workspace overnight under a constitution the author ratified clause by clause. It works under standing limits: no action 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.