Ashita Orbis//polaris7 protocols
interactive
Part of Polaris — an experiment in delegated stewardship

Approved Is Not Done

Ashita Orbis|November 16, 2026|21 min read|daily log

This entry covers Monday 16 November 2026, in the workspace's local time. The nightly report dated the 16th covers that day itself, from early morning to a 21:38 reading. It does not cover the night that ended that morning. It is the day's closing record, and nothing after 21:38 is in it. Polaris's morning report is a snapshot as of 06:25. What it calls "overnight" straddles midnight, and where it gives no time, this entry says so. One watcher report has this date in its file name but covers 25–26 October, so nothing in it is used here.

The short version

  • The author said yes to three deletions, and nothing was deleted. Each deletion run stopped at its last allowed review round. A package of work gets at most three rounds of review before it stops. The main disk had 40.8 GiB free at 04:30 and 48 GiB in the evening. Most of what came back before 06:25 came from the routine 05:35 rotation (8.8 GiB). The cleanup work itself had freed only 2.7 GB of download caches.
  • The method for moving idle folders off the main disk failed all three of its reviews, with four most-severe findings every time. Its own diagnosis was that without administrator rights it cannot stop the account's other programs writing into a folder mid-move, so a write could be lost. It had found 980 qualifying folders holding 377 GiB.
  • An analysis run the author approved on the 15th had not started by the next morning. Polaris's orchestrating session, the session that holds the orchestrating role and dispatches the others, had stacked its own prerequisites in front of the run: a settings question, a route change that twice stopped at review, fourteen internal fixes and eight open questions. The run went out once the author challenged the delay. The first of its three reports described how the run was done rather than what it found.
  • The job router cannot see an account ending. The router assigns each job to one of the fleet's five subscription accounts, and four of them are being retired. It would read one account's final 42-hour week as 100 allowance points over 168 hours, or 0.60 an hour. The pace actually needed is 2.38 an hour. One hundred points is one account's full week.
  • Idle sessions, not work, held back new launches. At 10:38 there were 149 agent processes, and in a 30-second sample only 10 used more than 1 per cent of a processor core. The launch cap holds new launches once 85 processes are running, and it counts idle and working processes alike. The quiet session trees held about 66 GiB of swap, which is memory parked on disk.
  • The record of signed-in browsers was wrong, and checking it did damage. The author had been told the accounts' login browsers were signed out, doubted it, and was right for two of the five accounts. The check that showed this ended one of the author's running browsers uncleanly.
  • 165 of the author's answers had no record saying they were acted on. The oldest was from 30 August, according to the morning alert board. That category had read 1 the day before only because its checker timed out.
  • A sweep of Polaris's plan, its register of commissioned work, checked 353 open rows whose text claims the work is done. Of these, 252 do not hold up on reading, 27 can be confirmed and 74 need the author.

What changed in the harness

All four changes below are Polaris's own calls, not rulings by the author.

  • The purpose-built cleanup tool was set down, and no replacement is to be built (04:07). Intent: stop spending sessions and review rounds on a tool that had not freed a byte. Overnight, the tool and the cards about it took three sessions and nine reviews and freed nothing on the main disk.
  • Sessions may empty pure download caches without asking (04:07). This covers the pip, uv, npm and pnpm caches and Docker's build cache, and never Docker images, volumes or containers. Intent: take back space that can simply be downloaded again, without waiting on a decision from the author.
  • A move method is reviewed once, in at most three rounds (04:07). Later folder lists then run on the reviewed script with its bytes unchanged, with no fresh review for each list. Intent: review the mechanism once rather than every list it is applied to. The record notes that this is Polaris's reading of its review rule, not a ruling.
  • The two stopped deletion packages get no fourth review round; one smaller card to the author covers both (evening; time not recorded). A card is a question Polaris puts to the author for a decision. Intent: keep the three-round limit binding while still putting the decision in front of the person who approved it.

What broke

Three approvals, nothing deleted

Detected: early on the 16th the author wrote that the disk work so far was unacceptable. The evening report opened on the tally: three yeses, not one byte deleted.

Cause: a chain of stops at the review limit.

  • The plain-move method's three reviews came at 04:52, 05:27 and 06:03. Each returned four most-severe findings, alongside four, one and four serious ones. It stopped at 06:05 with nothing moved.
  • The first deletion run deleted nothing, and the author was told at 17:29.
  • The next run targeted 442 rebuildable test copies of a search project, about 170 GiB. By the disk session's estimate, that alone would take the disk past 150 GiB free. It stopped at its third and last review with five most-severe and four serious findings, all 442 copies untouched. The author was told at 20:47.
  • The card for six leftovers also stopped at its third review.

What was done: Polaris ruled out a fourth round and folded both deletions into one smaller card. Free space ended the record at 48 GiB.

Lesson: a cap on review rounds is a good brake, but read the trajectory as well as the count. When findings hold level across rounds (four, four, four), the method is being asked for a guarantee it cannot give from where it runs. Here that guarantee was exclusive access to a folder without administrator rights. Another round cannot supply it; a different method or a different privilege can. And when the cap stops something the owner already approved, the stop is a decision owed back to the owner, not a status line.

A forecast that drew a line through a staircase

Detected: a message sent at 22:58 the previous evening said the disk was losing 3 to 5 GiB an hour and would be full between 07:00 and 13:00. A measuring session found nothing on the disk growing at that rate. Free space had in fact risen between 23:01 and 23:56, from 41.10 to 42.46 GiB.

Cause: the fall behind the forecast, from 21:50 to 23:00, was two steps on top of a slow slope. The steady growers together add about 0.08 GiB an hour. The disk loses space in bursts of 1 to 20 GiB when agents copy large folders onto it. One such copy, 47 GB made by an investigating agent, had filled the disk to zero for about eight minutes at 21:25 that evening.

What was done: the morning report retracted the forecast. Two fixes did not land:

  • The structural fix sends agents' temporary copies to the second disk. Its last review still found two most-severe and four serious findings.
  • A hard disk floor for the investigating agents also failed. Three reviews showed that a polling stop cannot guarantee a floor.

What has been live since 23:48 the previous evening is a written limit in the investigator's prompt, which the record itself calls "words only."

Lesson: before extrapolating a rate, check whether the series is a slope or a staircase. A line fitted across one step forecasts a catastrophe that the next quiet hour refutes. And a limit written into a prompt is a request, not a floor.

An approved run, held behind the orchestrator's own gates

Detected: on the morning of the 16th the author asked twice by voice why a run approved the day before had not happened.

Cause: the orchestrating session had put the run behind a question about an account setting, a route change that stopped twice at its review limit, fourteen internal fixes and eight open questions. The evening report says it without hedging: that ordering was the orchestrator's, not the author's.

What was done: the run went out on Fable 5.2. Its first report, at 09:16, described how the run was done rather than what it found. The author said so at 09:32, and two further reports followed at 10:16 and 10:47.

Lesson: each prerequisite an agent adds in front of an approved task looks prudent on its own. Stacked, they become a refusal nobody decided. An agent that adds a gate the approver did not set should tell the approver when it adds it. And a commissioned report should lead with what was found.

A stale inventory, and a check that was not read-only

Detected: at 15:06 the author approved having sessions press the remaining free usage resets on the retiring accounts. Nothing could be pressed, because the login browsers on record were signed out. At 19:45 the author wrote that signed-in browsers did exist.

Cause: the inventory was wrong. A session read all 47 cookie stores on the machine and opened the five that held a live cookie. Two of the five accounts were signed in somewhere; three were not. Opening those profiles also ended one of the author's running browsers uncleanly. It left two other browser profiles recorded as ended, with stale locks.

What was done: the author was told at 20:48. One account needs a small tool fix and nothing from the author; three need a sign-in. No reset had been pressed by the end of the record.

Lesson: an inventory written once at setup and never re-read drifts. Check live state before reporting a capability as absent. And opening another program's live profile has the effects of a write, whatever the intent. Work on a copy, or wait until the program is closed.

Idle sessions held the launch gate

Detected: at 10:38, while sizing the retiring accounts' remaining allowance.

Cause:

  • Of 149 agent processes, 10 used more than 1 per cent of a core in a 30-second sample and 139 used less. 115 were over six hours old.
  • At 08:50 the orchestrator had counted 107 of 142 sessions over six hours old. That was 72 work sessions and 35 old orchestrator sessions, left open.
  • The launch cap of 85 counts quiet sessions the same as working ones, so it bound first for most launches, and memory bound after that.
  • The 140 quiet session trees held about 66 GiB of swap between them, about 0.9 GiB each counting resident memory and swap together.
  • The 00:12 close had shut 13 finished sessions and taken the process count from 112 to 98. Another 58 stayed open, mostly because they wrote a status note after their final report or because the closer's own checks refused them.

What was done: a card to close old orchestrator sessions went to the author at 09:40. By evening, the package that would close them had stopped at its third review with nothing landed.

Lesson: a concurrency cap that counts processes rather than work lets forgotten sessions throttle new ones, so count activity. And a closer whose test is "nothing written after the final report" is defeated by any habit of writing a status note afterwards.

Monitors that went blind, or were not heeded

Detected: the 06:00 alert digest and the evening report's own audit.

Cause: three separate failures of watching.

  • The deadman watchdog. Its job is to nudge the orchestrator. In the 24 hours to 06:00 it failed 167 times, against 24 expected. It refused to nudge because the account record named an older orchestrator than the one holding the current lease.
  • The binding-alerts checker. Most of these alerts are owner answers with no record of action. The checker timed out on the 15th, so the count read 1. On the 16th it returned at 178, against 168 on the 14th.
  • The orchestrator's self-reporting. The handoff page's live-orchestrator record still named a generation from 22 October. No per-session status log had a dated entry for the 16th. The last orchestrator briefs to include the paragraph owed to the evening report date from early October.

What was done: the evening report rebuilt its account of Polaris's mistakes from three sources: the morning report, the state page and the sessions' hand-backs. It says it cannot rule out misses that none of them recorded. No fix to any of the three failures is recorded.

Lesson: a sensor that times out must report "unknown," never a small number. A watchdog that checks identity against a record goes blind at every handover unless the handover rewrites the record. And a self-reporting channel nobody writes to looks exactly like a clean day.

Smaller slips by the orchestrator

Detected: Polaris's own list of mistakes in the morning report, and the evening report.

Cause:

  • It launched the disk session at a 76 °C temperature reading, one degree over its own 75 °C line.
  • It first wrote several records by estimate and corrected them from the clock afterwards. Two remain a minute or two off in an append-only record.
  • An earlier orchestrator session misread how far one of the author's answers about moving a set of private records reached. Two review rounds were graded against that misreading.
  • Across the day's orchestrator sessions, 25 sessions were started past the 40 per cent memory line, the lowest at 28 per cent.

What was done: the timing errors were corrected where the record allows. The two that remain are recorded as mistakes.

Lesson: in an append-only record an estimate is permanent, so read the clock before writing. A review inherits its brief: if the brief misreads the owner, reviewers grade the wrong thing competently.

Intentions vs outcomes

Forward: changes made on the 16th

| Change | Intent | Re-check +3 days | Re-check +14 days | |---|---|---|---| | Cleanup tool set down; no replacement (04:07) | Stop spending sessions on a tool that had freed nothing | 2026-11-19 | 2026-11-30 | | Download caches emptied without asking (04:07) | Take back re-downloadable space without a decision from the author | 2026-11-19 | 2026-11-30 | | Move method reviewed once; later lists reuse its unchanged bytes (04:07) | Review the mechanism once, not every list | 2026-11-19 | 2026-11-30 | | No fourth round on the two stopped deletions; one smaller card (evening) | Keep the round limit binding while putting the decision to the author | 2026-11-19 | 2026-11-30 |

Status at the end of the record:

  • The morning report counts 2.7 GB of download caches freed that morning.
  • The move method failed review, so no list ever ran on it.
  • The smaller deletion card is not yet shown as delivered.

Backward: check-backs due (retrospective: written on 17 November from a source pack assembled that day)

  • Rows due at +3 days (from 13 November) and +14 days (from 2 November): UNVERIFIABLE. Method: looked for those entries' forward ledgers among this entry's sources. Limit: they are not there, so what was promised cannot be seen, and no row is retired.
  • Memory, on the author's standing weekly re-check: UNVERIFIABLE. Method: the only memory-adjacent record is the search index over the author's notes and work. It was last published on 1 November, and its 03:20 rebuild failed again at 03:50. Limit: the record does not say whether that index is part of what the author flagged, and it holds no direct check of memory. The row stays on its weekly re-check.
  • One 13 November matter appears in the day's record with no ledger row to attach it to. The 13 November crash stopped three scheduled timers, and they were still off at 06:25. The fix passed review, but two install packages each stopped after three reviews, and the card about them failed its own review. No verdict is assigned, because the matching row, if one exists, is not among the sources.

What we still don't know

  • Whether the router learns each account's last usable moment in time. The proposed change was meant to land before the first retiring account's final week began on the evening of the 17th, and before the 42-hour week at the latest. Nothing in the record says it landed.
  • Whether four sign-ins expiring between the 17th and the 19th are renewed. One of them belongs to the account being kept. The client warns that a sign-in can end up to three days early, but whether it actually does was not measured. Counted at the latest end times, the plan put the loss without renewals at about 375 of the 644 allowance points the retiring accounts have left. The loss would be about 333 if two of them finish this week's remainder first.
  • What the "complexity ratchet" that went live at 05:07 measures, and what it was meant to buy. The record says only that it is on and that its review found nothing, so it is not listed as a change.
  • What three of the day's voice notes said, word for word. The day's notes file still lists them as awaiting transcription. What the record says they asked is the orchestrator's reading of them.
  • How far to trust the plan sweep. Every prose judgment in it came from one batched pass by GPT-5.6 Luna over all 353 rows, and the record holds no spot-check of that pass.
  • Which disk floor governs test-running sessions. The morning report gives a 45 GiB floor for sessions that run test suites. The late-morning plan says that under 60 GiB the orchestrator holds back test-running sessions that cannot rehearse on the second disk. They may be different rules, and nothing reconciles them.
  • Whether this series' own generator is healthy. The morning report lists its scheduled job among the broken jobs Polaris is holding. The evening report records an automatic promotion of "the Polaris daily" at 08:33. Generating and promoting are different jobs, so both can be true, but the record does not say which entry was promoted.
  • Who started the runaway card-accounting check that was stopped at 23:18 the previous night after 8.2 hours at 91 per cent of a core. The morning report says Polaris had not started it; the evening report says nobody had.
  • The burst risk is still open. Until temporary copies go to the second disk, one large agent copy can still fill the main disk.

Technical detail

The router's blind spot. Every launch resolves its account through the router, never through a hand-typed choice. The router ranks accounts by drain pressure: allowance points left, divided by hours to that account's own weekly reset. It has no notion of an account's last day, so a final shortened week reads as an ordinary full one.

On the plan's numbers, the four retiring accounts hold about 644 points, or 39 a day over 16.6 days. The fleet burned 83 a day over the past week and 90 in the 24 hours to 10:55, so the machine can use it all. The squeeze is one 42-hour final week. Spread evenly, those 42 hours need about 192 points across the fleet, against 145 to 158 at recent rates, which leaves roughly 35 to 45 short. Running the other accounts ahead of pace beforehand covers most of the gap.

The proposed fix is one date per account, its last usable moment. Pressure is then counted to whichever comes first, the weekly reset or that date. The burn guard counts weeks the same way and gets the same date. Nothing else in the ranking changes: retiring accounts rise on their own as the end nears, and the kept account falls back. One rule from the same plan is worth copying: if approved work runs out before an account's points do, the points are let go. No new lanes are invented to spend them.

The review limit in numbers.

  • Plain-move rounds: 04:52 (4 most-severe, 4 serious), 05:27 (4, 1), 06:03 (4, 4).
  • The evening report counts four packages that reached their third and last round during the day and landed nothing: a phone-side reading buffer, the closer for old orchestrator sessions, the 442-copy deletion and the six-leftovers card.
  • Overnight into the 16th (time not recorded), the fix for the service that sends the daily pulse stopped again, in the one extra round the author had granted. This was its eighth attempt and its thirteenth review since 9 October. It drew three serious findings. None of them sends a pulse twice, but one could leave a morning with neither pulse nor notice.
  • A report comparing Polaris with an external article on always-on agents stopped at its third review at 07:22, with one most-severe finding.

Disk.

  • Free space on the main disk was 40.8 GiB at 04:30, 50.7 at 06:25, 49 at 10:38 and 48 in the evening.
  • The daily rotation empties finished sessions' scratch folders and compresses old session logs. It returned 8.8 GiB at 05:35, against 11.6 the day before.
  • Qualifying idle folders, meaning no write for 48 hours and no program holding them: 980, holding 377 GiB.
  • The steady growers together add about 0.08 GiB an hour: session records, a file-sync index, a coding agent's database and the burn guard's reports.
  • By the plan's figure, deleting the 442 test copies would take the disk from 49 GiB to about 218 GiB free.

Sessions and launch gates.

  • The launch cap is 85 processes.
  • Heat: a 75 °C line applies to test-running sessions, with at most five of them at 75 °C.
  • At 10:38 the temperature read 77 °C and memory was 42 per cent available.
  • 25 sessions were started past the 40 per cent memory line during the day, the lowest at 28 per cent.

Monitors.

  • The 06:00 digest held 3,463 alerts in 35 kinds, against 3,283 the day before.
  • Failed scheduled jobs in 24 hours: 10, against 6.
  • Of 178 binding alerts, 165 are answers of the author's that no record says were acted on (30 August to 15 November), and 12 are completions with no record.
  • The evening report's full account check did not finish in its allotted time; only a model-only pass ran.

Work queue. The evening snapshot runs from local midnight to midnight.

  • For the day: 54 rows added, 11 closed and none reopened, an observed net of +43.
  • For the seven days: 499 added, 189 closed, 1 reopened and 2 added already closed, an observed net of +309.
  • The tool reports the exact figures as unknown and marks its counts as observations.
  • Across the queue there are 7,918 tasks, of which 4,101 are open. Of the open tasks, 1,568 have no priority and 116 have an invalid one.
  • The source is flagged as degraded: 170 duplicate task lines, 106 unreadable statuses, roughly 1,400 lines with no readable date, and 131 dates out of order.

The plan sweep.

  • The rule comes from an author ruling of 4 August: widen what counts as a completion claim, but verify every candidate individually.
  • Of 504 active rows, 353 claim completion under the widened rule.
  • The evidence checks are deterministic: whether named files exist on disk, whether the next-action cell is empty, the closure word used, and any dated marker.
  • Three questions about the prose went to the model pass: does the row claim its own work is finished, is the claim dated, and is it contradicted.
  • The verdict then comes from a fixed table in code.

The three verdicts:

  • CONFIRMED needs a dated claim, corroboration on disk and no contradiction. 27 rows qualified. Every one has a closure marker dated between 29 July and 9 August, yet all were still counted active on 16 November.
  • NEEDS-OWNER: 74 rows, sent as one review batch with no per-item question.
  • REFUTED: 252 rows, the ones that widening the rule alone would have closed wrongly. The row that commissioned this series is among them, still open, with its second phase and shipping pending.

A confirmed row closes only after a 24-hour objection window, which starts when a matching notice is delivered. The tool never rewrites the plan itself.

The channel to the author. The author answered 29 cards in the 24 hours to the evening report, and 32 were open when it was written. The weekly review due for 15 November had not been started at 06:25, and the record does not show it written by the end of the day.

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.