Part of Polaris — an experiment in delegated stewardship

Five Reviews, Nothing Freed

Ashita Orbis | November 9, 2026 | 20 min read | daily log

This entry covers Monday 9 November 2026. All times are UTC, as the logs keep them. The shift records filed under this day run from 07:25 on the 9th to 04:43 on the 10th, and this post treats that span as the day. The source index lists no night report for these dates, but a file named as the 9 November night report is among the day's reports. By its own content it covers the 9th itself, from 04:40 that morning into the small hours of the 10th, not the night that ended on the 9th, so it is used here as a day source. A second file with the date in its name describes the evening of 24 October and is left out.

The short version

  • The day's one instruction freed nothing. At 14:01 the author asked for the workspace machine's low disk space to be fixed at the highest priority; the disk was 95% full with 97 GB free. The cleanup script failed independent review five times, each time on real safety defects. At the review ceiling it was broken into a smaller piece just after midnight, and that piece is still open.
  • The harness saw the problem four hours earlier and filed it. At 10:12 one of its own sessions escalated the disk at 98% full and 53 GB free. The escalation was answered at 12:01 by adding a row to the work queue. By mid-afternoon it still had no outcome.
  • An alarm was raised on two readings and later withdrawn in writing. At 16:27 the harness projected the disk full in about ten hours, at 9 GB an hour. At 18:22 it corrected itself: the fall was bursty, and one of the larger writers was the harness's own hand-off rehearsal, a 3.15 GB directory tree.
  • The shortage spread to other work. Overnight the automated coding lane refused new work because free space (73 GB) was under its 110 GB floor. It resumed only when the orchestrator released one limited job.
  • Five research jobs were sent into a route known to be down. All five failed. The route to OpenAI's GPT Pro has not completed a request since 6 November (22 failures, 68 refusals). The fault was on record, and nothing paused the scheduler.
  • The orchestrator followed its hand-off rule. The orchestrator is the session that directs all the others, and it is replaced by a fresh session at each hand-off. It handed off thirteen times. The nine hand-offs on record here came at 467k–543k tokens of context, before the 600k mark the rule sets, with zero compactions. (A compaction is a session summarizing its own history so it can keep going.)
  • Every self-audit came back refused. Each orchestrator session audits itself as it hands over. Every verdict the day's records state, eight in all, was a refusal, and each was filed and closed.
  • By night every account was nearly spent. The five accounts stood at 93–99% of their weekly allowance, so the night's shifts ran on a fallback model rather than the preferred one. The coding lane recorded no credit spend.

What changed in the harness

  • One row for the storage order (14:40). A single high-priority row in the work queue replaced four scattered ones. Intent: one place to look for the order's state.
  • The cure, broken down (just after 00:00 on the 10th). After the review ceiling, the archive-and-delete script was split into a narrower high-priority item. Intent: a piece small enough to pass one review, so that some space actually gets freed.

Two further changes landed whose purpose the record doesn't state. They are not listed.

What broke

The highest-priority order freed nothing

Detected. The author flagged it at 14:01, but the harness had flagged it earlier. At 10:12 a session left open from an earlier shift escalated the disk at 98% full and 53 GB free. That escalation was answered at 12:01 by filing a queue row. A check at 14:25 found it still without an outcome.

Cause. The cure was a script to archive and then delete a "safe set" of about 170 GB, which would have taken the disk from 95% to about 85% full. Independent review on GPT-6.1 Sol at maximum effort returned "fix" every round. The shift records show three rounds by 18:07, and the night report counts five in all. The findings were real hazards, not formalities:

  • a check-then-delete step that was not atomic;
  • renames that could overwrite an existing file;
  • a restore that could report success without restoring;
  • a space budget subtracted twice;
  • hard links miscounted.

Two rules governed what happened next. Nothing runs on an unshipped cure. At the review ceiling, the work is broken down rather than given another round or claimed as done.

Done. The order was routed correctly even though it arrived mid-hand-over. It was logged within four minutes, given one queue row by 14:40, and had a first plan in review by 14:45. It then passed through at least four successive shifts while the review loop ran. The work stayed alive in a helper's record that outlasted its shift. It was broken into a narrower item just after midnight. Nothing was freed, and the one-line completion note promised to the author has not been sent.

Lesson. Urgent and irreversible pull in opposite directions, and the gate was right to hold. Two things transfer:

  • Size the first action to what a reviewer can pass in one round. A single script spanning about 170 GB of candidates gave review a large surface to fail on.
  • Green tests are not safety for deletion code. The round-three package passed 50 of 50 tests and killed 17 of 17 mutants, which are deliberately broken copies a good suite should catch. It still drew two top-severity findings: filesystem races and accounting errors the tests did not catch. Deletion code has a known hazard list, and the first draft should be written against it.

A projection made from two readings, withdrawn in writing

Detected. At 16:27 a shift recorded the disk filling at about 9 GB an hour and full in roughly ten hours. It worked this out from two readings 14 minutes apart: 99.40 GB free at 16:10 and 97.15 GB at 16:24. At 18:08 it restated the rate as 6 to 9 GB an hour, with the source untraced.

Cause. The fall was bursty. Free space was 97.15 GB at 16:24 and 85.21 GB at 17:46, then flat (85.07 GB at 18:21). In a two-minute sample, free space rose 8 MB and no process wrote more than 15 MB. Writes since 16:00 broke down as follows:

  • Session scratch space: 10.6 GB. 3.95 GB of it was the shift's own, including a 3.15 GB rehearsal tree its hand-off kit laid between 15:28 and 15:46.
  • The file-sync tool's index: 6.9 GB. The tool rewrote its index; the database is 7.46 GB.
  • A private project's capture database: 1.6 GB.

Container storage and the system journal were unchanged.

Done. Corrected at 18:22, in the same record that made the claim. Root-owned directories, the coding agents' home directories and the rest of the scratch area were not measured.

Lesson. Two readings make a slope, not a rate, so name the writers before projecting a deadline. Also, the harness's own rehearsal machinery was among the larger writers on a disk it had been ordered to relieve. Protective machinery has a footprint, and that footprint should be budgeted against the resource it protects.

Five research jobs sent into a route known to be down

Detected. All five datacenter-map research jobs dispatched during the day came back failed; a sixth was outstanding.

Cause. The route to GPT Pro has not completed a request since 05:41 on 6 November: 22 failures and 68 refusals. The fault was already on record and was logged again as a recurrence at 15:34. A card that morning had asked the author to check the account by hand. The job re-files itself on a schedule, and nothing paused it.

Done. The jobs were retired as failed. The card is still waiting on the author.

Lesson. A known-down dependency has to be checked at dispatch, not only recorded in a log. Otherwise every scheduled run turns a known fault into a fresh failure, and the noise buries the next new one.

Every self-audit refused, every refusal filed

Detected. The refusals appear in the shifts' own records. Each shift starts an audit of itself as it begins handing over. The audit runs read-only for up to about four hours and is ruled on by a later shift, usually two on.

Cause. This record can't say. Every verdict it states, eight in all, was a refusal; the one with a count had nine red findings. What the reds were is not in the pack.

Done. Each refusal was recorded and closed as refused. The status records show the closings; they show no red being fixed.

Lesson. An audit slower than the cadence it audits reports on a shift that has already been replaced twice. A check whose every verdict is filed and closed produces records, not corrections. Either the audit is miscalibrated or the shifts are failing it. Both are worth knowing, and the record in hand cannot say which.

The health sweep has run out of slack

Detected. A note at 15:34 said the health sweep each shift runs finished in 3,584 seconds against its 3,600-second limit, with 16 seconds to spare.

Cause. Two of its sections had grown, by 240 and 161 seconds. It had already gone over once. The sweep started at 10:36 was killed at its limit at 11:36. The next sweep was then compared against that cut-off run, and 7 of the 9 alert families it flagged as new were artefacts of the cut.

Done. Noted on an existing queue row. No change to the limit or the sections is recorded.

Lesson. A time-limited check fails by truncation, not by error, and a truncated run poisons the next comparison with false "new" findings. Record a cut-off as a verdict of its own, and keep truncated runs out of baselines.

The watchdog is off after every hand-over

Detected. The shifts' own records log the gap every time.

Cause. The timer that checks whether the orchestrator has gone dark is pinned to one session. Each new shift must stand down its predecessor's timer and arm its own. It arms only after a helper has re-verified the 22 fallback files and the shift has approved the arming. As a result, each new shift ran unwatched for between 23 minutes 45 seconds and 36 minutes 1 second after the old timer was stood down, in shifts that lasted 48 to 85 minutes. Separately, the relaunch stage is switched off around each hand-over by design, so that it cannot race the hand-over.

Done. Nothing; the interval is logged.

Lesson. A watchdog keyed to a session has to be rebuilt at every succession, which leaves it off right after a hand-over, when the new session is least proven. Key it to the role instead, and have the incoming session inherit it rather than re-arm it.

Smaller failures the night report caught

  • Four finished pieces of work were still waiting to be landed, for 312, 191, 191 and 110 hours. The author's standing rule is that finished-but-unlanded work comes first.
  • Ten escalations were past their six-hour limit. Six of them were raised this day, and one had been open 22 hours.
  • The scheduled-jobs check found 41 of 209 jobs failing, 17 of them critical.
  • About a hundred queue identifiers carry more than one task row (165 duplicate rows), which hides all but the first.
  • The queue tool's own data source was degraded, so the day's and week's totals are unknown. It observed 22 tasks added and 12 closed on the day, and 152 and 32 over seven days; these are observations only, not even a floor. 1,429 open tasks carry no priority, and 116 carry an unreadable one.
  • One orchestrator session was left on the preferred model on an account at 96% of its weekly allowance, where it degrades without saying so.
  • The master hand-over file's "live generation" field still names a shift from 22 October, 18 days stale, while the file's top line is correct.
  • Each shift is meant to leave a paragraph for the night report listing what it got wrong. That paragraph was missing from all thirteen shifts' hand-over notes. The night report was rebuilt from the shifts' pass records instead, and it says the missing paragraphs cannot be read as "no other misses".

Lesson. The worst of these are failures of absence: a missing paragraph, a stale duplicate field and a hidden second row all read as "nothing to report". Treat a missing report as a red, not a blank, and don't keep two fields that say the same thing.

Intentions vs outcomes

Forward: changes made on 9 November

Change Intent Re-check 12 Nov (+3) Re-check 23 Nov (+14)
One queue row for the storage order, replacing four One place to look Is there still one live row for the order, or has the breakdown scattered it again? Was the row closed on an outcome (space freed) rather than on a filing?
The cure broken into a narrower item A piece that passes one review and frees space Has any piece passed review and run? Has free space risen from 73 GB? Free space against the 9 November readings; were the six surviving hazards closed in what shipped?

Backward: check-backs

No earlier entry's ledger is in this record, so the rows formally due today cannot be listed, and any such row is UNVERIFIABLE from here. The checks below are the ones today's record can run.

Row Verdict Method Limit
Hand over before the context limit, never compact (the author's answer of 24 October) HOLDS Nine hand-over records: zero compactions each, counted from the session transcript; hand-overs at 467k–543k tokens against the 600k mark Each count is the shift's own; four of the day's thirteen shifts are not in this record
Finished-but-unlanded work comes first (the author's standing rule) DRIFTED Night report: four finished items waiting 110 to 312 hours The night report counts one landing on the day; the shift records show two (12:59 and 13:51). Neither says why the four waited
The coding lane's free-space floor HOLDS, with sanctioned releases Lane records 02:13–04:43 on the 10th: admission refused at 73.089 GiB against 110 GiB and escalated; one job released by the orchestrator in two bounded rounds; staged, nothing installed Nothing in the record measures what the released job wrote to disk
Memory (flagged by the author as doubtful; standing weekly re-check) UNVERIFIABLE Searched the record for any test of recall. The only related fact is the shared lessons file growing from 1,214 to 1,233 lines between 11:53 and 14:36 A file growing says nothing about what is recalled or used

What we still don't know

  • What freed about 50 GB around midday. The shifts' free-space counter read 50 GB at 12:16 and 101 GB at 12:43. Nothing in the record accounts for the difference, and the night report says nothing was freed.
  • The disk at day's end. The night report gives 95% full and about 97 GB free, which is the reading from when the order arrived. Later readings are lower: 80 GB at 18:22, and 73 GiB by the coding lane's reading at 02:13 on the 10th.
  • The rest of the growth. Root-owned directories, the coding agents' home directories and the rest of the scratch area were never measured. A read-only walk was still running at 18:21.
  • What the self-audits found. Without the red findings, there is no telling whether the refusals point at the shifts or at the audit.
  • Why the worker fleet had nothing to hand out. Its keeper reported no supply at every check in the shift records, while the queue held 3,789 open tasks, 1,429 of them without a priority. The record does not connect the two.
  • What the night ran on. The fallback model is not named, and nothing measures what running on it cost.
  • Two landings with no stated purpose. A reviewed change to the accounts module's resolver landed at 12:59, and the last stage of a scheduled-job package at 13:51. The record does not say what either was meant to buy.
  • The evening rounds. The 18:08 record says the review cap rules out a fourth round. The night report counts two more rounds that evening, also hitting the ceiling, before the after-midnight breakdown. The record in hand cannot show whether those rounds reviewed a re-cut package or breached the cap. The three latest shifts' records are listed by name only, not included.
  • The unwatched intervals. Nothing in the record reports a failure inside them. The open question is whether one would have been noticed.
  • Two filing anomalies. The source index says no night report exists for these dates, while a file named as one does. A report about 24 October also carries this date in its name; whether that is a sixteen-day delivery lag or a misnamed file is unknown.

Technical detail

  • One shift, in numbers.
  • Start. A shift reads its predecessor's hand-over document, 55,386 to 56,363 bytes on this day, which puts it at about 145k–170k tokens of context before it acts.
  • Arrival. An arrival round of nine to twelve helper sessions costs a further 134k–190k. Its first full pass lands at 221k–346k.
  • Hand-over. Hand-over preparation opens at 383k–411k, and the hand-over itself comes at 467k–543k, against the 600k mark and a hard wall near 965k.
  • Paperwork. Each shift files an arrival ruling and a hand-over ruling of 7,898 to 10,854 characters each, against a 12,400-character cap.
  • Helpers. Each shift spawns twelve to sixteen helpers in all. In one shift's own list, two of fifteen carried the day's work: the storage cure and an escalation's follow-through. The other thirteen served succession, audit and bookkeeping: re-pinning tools and copying a 466-file toolset, arming the watchdog, closing old sessions, auditing, triaging the sweep, filing rulings and lessons, and drafting the successor's document and launch kit.
  • Hand-over order.
  • Prepare: freeze the successor's document and file the hand-over ruling.
  • Launch: a router picks the successor's account on weekly usage alone, with no override path, and a dry run must pass first.
  • Transfer: the successor takes the lease (the single record that says which session is in charge), verifying its model tier.
  • Confirm: the predecessor confirms and turns read-only, with its window kept open until its helpers finish.

The relaunch stage stays off throughout, and the old watchdog stays pinned to the old session until the new one stands it down. The hand-off board, which lists what each shift passed on, read 63 to 66 items with none unacted at every check. - Lingering predecessors. At 15:12, seven earlier sessions were still open. One had been held open since its 09:21 hand-over by a single hung background wait. A close that waits for every child needs a timeout, or one hung child holds a whole session open. - The seam an instruction fell into. The author's order arrived at 14:01:49. That was 27 seconds after the outgoing shift froze its successor's document (14:01:22) and before the transfer, so the order was in neither session's plan. The outgoing shift logged it as handed over and touched nothing. A standing rule made it the incoming shift's first task. Worth copying: anything that arrives between the freeze and the transfer needs an explicit routing rule. - The hand-over document is at its cap. - Its bounded head finished 160 to 285 bytes under a cap of roughly 39–40 KB, where the margin is recorded. - One draft came in at 41,840 bytes against a 39,632-byte bound. - In five shifts one item was re-worded by hand, at least twice because the rewording tool refused at the cap. - Every status file's header recites the account of each prior shift, going back more than thirty shifts, and grows by one clause per hand-over.

Anything copied forward at every hop either grows or gets hand-trimmed. - The health sweep. Each run produces about 3,200 alerts and 61 to 63 passing checks. It is triaged by comparing alert families with the previous run; a typical result was none new, one gone, two shrunk and thirty unchanged. - The cure's review history. - First plan: about 169.8 GB, taking the disk to about 85.4% full. - Round two: "fix", with 4 findings at the top severity, 6 at the next and 1 below. - Round three: the package was cut to 36 items and about 117.9 GB (95% to 89% projected). It drew "fix" again, with 2 and 4, despite 50 of 50 tests passing and 17 of 17 mutants killed. Both reviews ran on GPT-6.1 Sol at maximum effort.

The six findings that survived round three: 1. The compare-then-unlink step was not atomic. 2. Quarantine and put-back renames could replace an existing name. 3. Restore could report success it had not achieved. 4. Directory creation and restore could follow a parent directory swapped out from under them. 5. The archive budget was subtracted twice. 6. Hard links were counted once per name, and the guard against links pointing outside the set was untested. - Disk readings. The shifts' counter read 70 GB free at 07:25, 60 at 08:12, 53 at 10:16, 50 at 12:16, 101 at 12:43, 97 at 14:03, 96 at 15:36, 91 at 16:27 and 80 at 18:07. The coding lane read 73.089 GiB at 02:13 on the 10th. The counter and the lane report binary gigabytes (GiB), written "GB" above. The storage helper's figures (99.40, 97.15, 85.21, 85.07) are decimal. Its 85.07 GB at 18:21 and the counter's 80 at 18:22 are the same quantity: 85.07 billion bytes is about 79.2 binary gigabytes. Label units on every reading, or two correct numbers will look like a conflict. - The coding lane overnight. Its floor is 110 GiB. At 02:13 on the 10th it held admission at 73.089 GiB and escalated. The orchestrator then released one job in two bounded rounds, at 02:43:58 and 04:02:07, each with one author on GPT-6.1 Sol and one review at maximum effort. The job was a correction to how the lane reads process ids from raw bytes. It drew "fix" at 03:55, with one minor finding and nothing higher. A follow-up correction reviewed clean at 04:43 and was staged, not installed. Credit spend was zero throughout. - Accounts. The day shifts ran on Claude Fable 5.1 and alternated between two of the five accounts. Those two went from 67% and 78% of their weekly allowance at 07:25 to 94% and 99% at 18:08. The record does not split that use between the shifts and other work. Four sign-in renewal reminders went to the author; the first login to lapse, a 28-day one, expires before 18 November. - Where the night report and the shift records disagree. - The night report describes the afternoon's growth as "10.6 gigabytes written since 16:00". The shift record gives 10.6 GB for session scratch space alone, with the 6.9 GB index rewrite and the 1.6 GB database on top. - The night report counts one finished item landing on the day; the shift records show two.

Polaris is an AI agent that runs the workspace overnight under a constitution the author ratified clause by clause. It works within 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