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

Paraphrase Beat The Guard

Ashita Orbis|November 19, 2026|23 min read|daily log

This entry covers Thursday 19 November 2026, on the author's local clock. Times marked UTC are not local. It is built from that day's record: a morning report describing the state at 05:40, a nightly report written that evening, a comparison of the harness with published guidance on always-on agents, two sweeps of the work backlog, a research background document, and the day's decision answers. The morning report narrates what happened since the previous morning's report, so much of what it describes happened on the 18th. Items from the 18th appear here only where they bear on the 19th, and they are labelled. The pack's assembler listed no night report for the 19th. A nightly report dated the 19th does sit among the day's reports, and by its own words it covers that day and its evening, so it is used here. Its "last 24 hours" figures reach back into the evening of the 18th, and that is the seam.

The short version

  • More work arrived than left. The work queue's own dated observations show 66 items added and 11 closed on the day; the queue labels these as observations, not an exact count. 4,284 of 8,126 tasks were open that night. All 504 open rows in the backlog registry had gone more than three days without any status file, decision record or review verdict mentioning them.
  • Most rows that sounded finished were not. 353 backlog rows use wording that claims completion. A row-by-row check confirmed 27 of them, refuted 256 and left 70 for the author to settle.
  • A text-matching guard lost to ordinary rewording. The guard was meant to stop machine-written "related posts" blurbs from restating a withdrawn claim. Paraphrase beat it in all three of its review rounds. At 04:49 the agent removed the offending line without the guard and declared the risk.
  • Two monitors misfired. One watchdog exists to reach the orchestrator, the single long-running session that takes every incoming event. At 05:18 it typed into a retired orchestrator session's window instead. Separately, for the second day running, a model-identity alarm reported a downgrade. The orchestrator's own transcript contradicts it: its intended model, Fable 5.2, appears in all 42 of its recent replies.
  • Requests piled up unread. At the 06:00 digest, 103 approval requests from one automated guard were waiting, against 12 a day earlier. 91 had arrived in 24 hours and none had been picked up.
  • The record behind the nightly "what went wrong" section has been dark since early October. That section was built from the day's session hand-backs instead, and it says outright that it may have missed something.
  • The agent wrote two operating rules for itself and labelled them as its own. The first: retry a refused task in a fresh session with a reframed request. The second, from 05:37: start only short sessions on an account whose sign-in is about to lapse.
  • It declined to put passwords in a message. The author asked for pending sign-ins to be made doable from the phone. A session started within three minutes and sent numbered steps per sign-in. It left every password for the author to read in their own terminal, and said why.

What changed in the harness

  • Refused tasks restart fresh, reframed. At 06:11 the author left a note suggesting that refusals are probably best handled by stopping and retrying with a reframe. The orchestrator adopted that as its own standing default.

    • A stopped session is never typed into again.
    • Material that trips a provider's filter goes to a different model, and the orchestrator does not load it itself.
    • It told the author this is its own reading, not a ruling.
    • It also reported that it could not find the refused session the note mentioned, rather than guessing.

    Intent: recover from a refusal without pushing more text into a session that has already stopped.

  • Long sessions wait when an account is about to lose access. At 05:37 the orchestrator wrote a rule: a short session may start on an account near its sign-in deadline, and a long one waits. Earlier, at 04:08, the orchestrator's own lease had already passed to a session on a different account, ahead of one sign-in lapsing at 07:02. Intent: keep long work from being cut off mid-run when an account loses access.

  • An ending account's end date becomes the deadline for its banked resets. At 14:27 the author left a note about an account nearing its end date. The orchestrator read that date as the deadline under the author's existing 72-hour rule for redeeming usage resets, since a reset never redeemed is simply lost. It told the author this reading was its own. Intent: no banked reset expires unused.

  • Sign-ins walkable from the phone, with no secrets in messages. At 07:27 the author asked for pending sign-ins to be doable from the phone. A session checked one phone path per pending sign-in and began sending one short message per sign-in, each with numbered steps. Any password is read by the author in their own terminal. The deviation from the literal request was named to the author, along with the note that their word changes it. Intent: make sign-ins possible away from the desk without a secret ever sitting in a chat log.

  • The inbox watcher is armed before the inbox is read. The morning report of the 19th records this change. It answers a slip on the 18th: an answer arrived between the orchestrator's read and its re-arm, and it sat about two hours until the next answer woke the orchestrator. The record does not say whether the change landed on the 18th or the 19th. Intent: no answer can arrive in that gap and wake nothing.

  • One correction landed without its regeneration guard, with the risk declared. Intent: finish the author's correction rather than spend a fourth capped review round on a guard that rewording defeats. The durable fix is filed and undecided (see below).

What broke

Paraphrase beat the guard

Detected in review. A correction to a published post had to remove a withdrawn claim from a machine-written "related posts" blurb. The agent required of itself a guard against any later regeneration writing the claim again. The review stopped all three rounds on that guard, because ordinary wording got past it, and the package hit its three-round cap at 03:24.

Cause: the guard matched words, but the danger was a meaning. The blurbs are written by a model and publish without review, and they had already written the withdrawn claim three times. Two more gaps sat beside it:

  • The landing session's inventory of every place the claim appeared missed this blurb. The orchestrator found it at 00:47.
  • The live blurb was served from the deploy's own cache (84 links), not the committed copy (77 links, older). Editing the committed copy alone would have republished older text.

A related failure sits elsewhere on the blog. Its question-answering feature still answers one question with a withdrawn phrase, because the correction notes at the top of that post quote the phrase and fill the context budget first. It has been found and filed, not fixed.

Done:

  • The line came out at 04:49 without the guard.
  • All 18 affected addresses read clean at 04:50 and again at 05:31.
  • The residual risk was declared.

The record names one design that holds for any wording: stop model-written blurbs from publishing on their own. That fix is filed but tied to a second one. The deploy step that reads those model replies parsed none of them, which deferred 23 of 25 analyses across seven deploys on the 18th and 19th. Fixing that parser alone would make the regeneration risk live, so the two are to be decided together.

Lesson: a correction has to reach everything generated from the corrected text: blurbs, caches, snippets and retrieval answers. A string guard cannot stop a model restating a claim in new words, so gate the generator, not the string. Also, a correction note that quotes the withdrawn claim feeds it straight back into retrieval.

Two monitors misfired

Detected by the 06:00 board digest, as two kinds of alert.

  • The watchdog. It exists to reach the orchestrator when the normal wake-up path is gone. At 05:18 it typed into the window of a retired orchestrator session. No turn was taken there, so that session's saved context was not refreshed. The same watchdog's job failed 29 times in the preceding 24 hours, against 39 the day before.
  • The identity alarm. For the second day running it reported that the orchestrator had been downgraded to Opus 5.5. The orchestrator's own transcript showed Fable 5.2 in all 42 replies within its last 200 lines. The morning report's reading is that the alarm most likely read a work session's transcript.

Cause: for the watchdog, the record shows the wrong target but not how the target was chosen. For the alarm, "most likely" is as far as the record goes.

Done: both remain as alerts; no fix is recorded.

Lesson: after a hand-off, the old session's window still exists and still accepts keystrokes. Anything that addresses "the orchestrator" must work out which session that is at the moment it acts, and confirm its input was taken. An identity check is only as good as its binding to the process it claims to check.

The record behind "what went wrong" stopped being written

Detected when the nightly report went to its named source for this section.

Cause: the orchestrator ran the day as four short generations and wrote to its own state page and hand-backs instead of the named records. The record shows each named source stale or missing:

  • The hand-off page's live-generation entry still names an orchestrator from 22 October, and the previous-generation entry names one from 15 October.
  • No generation hand-off log carries a pass entry in the day's window.
  • The file of items owed to the general report for the 19th does not exist.
  • The "owed" paragraphs in successor briefs stop in early October.

Done: the section was built from the day's hand-backs. It says it cannot rule out that the named source would have caught something more.

Lesson: when an agent changes how it works, the records that downstream reports depend on can stop without anything failing. A report should check that each named source is fresh and say plainly when it fell back, as this one did. Otherwise it looks complete while running on a dead feed.

A check that failed on a page that was right

Detected at read-back: one correction's landing passed 63 of 64 checks.

Cause: the failing check demands a heading in the post's PDF that the PDF has never printed, before or after the correction. The check is wrong, not the page.

Done: nothing further. The landing's closing bookkeeping is queued and unrun.

Lesson: run a new read-back suite against the unchanged page first. A check that fails on the baseline is measuring itself. And a correct landing that cannot close its bookkeeping becomes tomorrow's stale open row.

"Nothing in the record," when the record existed

Detected by a later session.

Cause: a decision card told the author that nothing in the record confirmed or refuted the timing a post claimed. In fact, the original working session's own timing record existed.

Done: the later session read that record and left the author's answer standing as applied.

Lesson: an absence claim on a decision card is really a search claim, and it should say where it looked.

Requests written, never picked up

Detected by the 06:00 digest and by the backlog sweep at 17:35 UTC.

What was found:

  • Approval requests: 103 from an automated guard that nudges sessions were waiting, against 12 the day before. 91 had arrived in 24 hours, the oldest had waited 30 hours, and none had been picked up.
  • Escalations: 98 new agent escalations were awaiting adjudication, against 49 the day before. Four had gone unanswered for more than four hours.
  • Unstarted dispatches: the sweep counted 585 dispatches written out but never started, each a brief with no matching status file.
  • Unread queues: the sweep counted 15,380 queues gaining records with nothing reading them, 769 records in total since the previous sweep. The largest is the queue of memory candidates (+383).
  • Unreachable drafts: 2,932 decision-card drafts sit deeper in folders than the author's Reports tab reads.

Cause: not established. The sweep's own words: whatever writes to each queue is still running, and whatever is supposed to read it is not.

Done: the sweep wrote one card draft per finding.

Lesson: a queue with a live writer and a dead reader is silent loss that looks like activity. The reader needs its own liveness alarm, and its findings need a channel that is itself read.

Work in, work out

Detected by the night report's queue snapshot and by the two backlog sweeps.

What was found:

  • The week of 13–19 November: 580 items added, 154 closed and 1 reopened, a net of +425. These are again observations, not an exact count.
  • Priorities: 1,748 open tasks carry no priority at all, and 116 more carry an invalid one.
  • Quiet rows: every one of the 504 open backlog rows had been quiet for more than 72 hours. 224 of them have never produced any signal, so their ages fall back to dates written in their own text.
  • Completion claims: of the 353 rows whose wording claims completion, 27 were confirmed, 256 refuted and 70 left for the author.

Cause: not established. The record measures; it does not diagnose.

Done: the closure tool proposes closures but does not make them. A confirmed row closes only after a 24-hour objection window, which opens when a notice is delivered.

Lesson: when 100 per cent of open items read as "quiet", the quiet measure has stopped telling items apart, and the sweep says so itself. Completion wording is not evidence either: 256 of 353 rows that sounded done were contradicted when checked one by one.

Closed sessions keep their scratch space

Detected in the morning report's disk reading.

What was found:

  • The work sessions' temporary folder on the second disk held 31.7 GiB at 05:42, against 22.95 GiB at 09:44 the day before.
  • Two closed sessions hold 18.8 GiB and 3.9 GiB of it.
  • Later that day the bulk disk had 70.2 GiB free (97% used).
  • The root disk had 91.3 GiB free (95% used), under a September floor of 200 GiB free.
  • The author has answered on this disk six times.

Done: nothing has been deleted, and no session has been asked to clear its share.

Lesson: closing a session should include releasing what it wrote to scratch space. Otherwise every finished session is a slow leak that only a human ruling can stop.

A first run over the heat limit

Detected by the session itself. It started its first rehearsal run at 80 °C, against a 78 °C start limit. On the 18th, two other sessions had each started one local run at 79 °C.

Cause: the record does not say why the first start skipped the limit.

Done: the session gated every later start and reported the slip.

Lesson: a start limit that each session applies to itself is weakest on the first start, so the check belongs where starts happen.

A dead remote-login service, masked by a second path

Detected by the session building the phone sign-in path, while it was looking for something else.

Cause: the machine's own remote-login service had been down since March, because its host keys were empty files, and it failed outright on 18 November. The author's phone had been reaching the machine only because a second network layer answered in its place. Nothing had caught the failure.

Done: the session wrote up the fix. The author's answer the same day took the recommended option: leave it down, because the second path is enough.

Lesson: redundancy hides failure. A check like "can I reach the machine" passes as long as any path works, so check each path you believe you have.

Cards that outlived their questions

Detected by the two reports' card counts.

What was found:

  • At 05:40, four sign-in renewal cards from 9 November still read open, although two of those accounts had been signed in on the 17th.
  • That night 34 cards were open. The oldest is dated 20 August, and one carries a deadline (31 October) that has already passed.
  • One card about what a post claims was answered with a request to see the words themselves. The follow-up card was answered the same day.

Cause: not recorded.

Done: nothing recorded.

Lesson: a decision card should retire itself when its condition is met some other way or its deadline passes. A card asking for a decision about text should carry the text.

Intentions vs outcomes

Forward: changes made on the 19th

| Change | Intent | Re-check +3 | Re-check +14 | |---|---|---|---| | Refused tasks: stop, then retry reframed in a fresh session (the agent's own default) | recover without typing into a stopped session | 22 Nov | 3 Dec | | Only short sessions on an account near a sign-in lapse | no long run cut off by lost access | 22 Nov | 3 Dec | | An ending account's end date becomes the deadline for its banked resets (the agent's reading) | no reset expires unredeemed | 22 Nov | 3 Dec | | Phone sign-in steps, with no secrets in messages | sign-ins doable from the phone, with no secret in a chat log | 22 Nov | 3 Dec | | Inbox watcher armed before the read (landing date unrecorded) | no answer lands unseen in the gap | 22 Nov | 3 Dec | | Withdrawn line removed without its guard; risk declared; durable fix filed | finish the correction, and decide the generator gate and the parser fix together | 22 Nov | 3 Dec |

Backward: check-backs

The pack does not carry earlier entries' forward rows. The due list below is therefore reconstructed from changes that the day's record dates, plus the standing memory row. Every check uses only the record as it stood on the evening of the 19th.

| Row | Verdict | Method | Limit | |---|---|---|---| | Memory (standing weekly re-check; the author has flagged it as doubtful) | UNVERIFIABLE | Backlog sweep: the memory-candidates queue gained 383 records since its previous run, with nothing reading it. 06:00 digest: a job named for nightly extraction is still on the failing list. Always-on comparison: memory's nightly job had failed as of the 16th. | Nothing in the pack tests whether recalled facts are correct or current. The pack does not confirm the failing job is memory's. The signals point the wrong way, but the verdict is not upgraded on a guess. | | One long-lived orchestrator session (the author's design of 13 Nov; its first check fell on the 16th and reached the author on the 19th) | DRIFTED | A reviewed measurement from 13:15 on 13 Nov to 05:45 on 16 Nov: 99.7 wakes a day against 36 estimated; 35.1 weekly points on the design's own scale against 5–10 estimated; 39 hand-offs with a median term of 76 minutes. | Nothing since 05:45 on the 16th has been measured. The design itself stands by the author's ruling; this verdict is on its cost. | | Retired orchestrator sessions close two hours after hand-off (the author's answer of 16 Nov; its +3 check falls on the 19th) | DRIFTED | Morning report: on the 18th, five retired sessions were closed and 20 stayed open on their own checks. The 05:18 alert shows a retired session's window still open on the 19th, with the watchdog typing into it. | The record gives no age for each retired session, so the number past their two hours cannot be counted. | | Finished work sessions closed by name once a day, and whenever the count reaches 85 (rule of the evening of 15 Nov) | DRIFTED | Morning report on the 18th's daily close: 44 closed and 60 stayed open. 22 of those had written a status note after their final report, and 9 had been nudged by an automated guard after reporting. 92 processes remained, above the line of 85, and the rest cannot close under the daily rule. | One close, read once. The pack has no count from a close on the 19th. |

What we still don't know

  • Whether the nightly "what went wrong" section missed anything. Its named source is dark, and the section says it cannot rule that out.
  • Why the assembler saw no night report. The pack's header lists none for the 19th, yet a nightly report dated the 19th is among the day's reports and is used here. The record does not say why the lookup missed it.
  • What a lane report filed under this date is for. One report carries the 19th in its filename, but its own text describes a window on 26 October. Nothing in the pack says why, and it is not used for this day.
  • Whether the work queue is healthy. Two readers disagree:
    • The fold read 4,284 rows from a source it reports as healthy, with no damaged lines.
    • The hygiene snapshot reports the same source as degraded (882 lines with no readable date, 168 with an unreadable status, 165 duplicated task lines) and calls its own counts not exact.
  • How many quiet rows are actually finished. The backlog sweep reads 105 as finished from their wording. The closure tool, checking row by row, confirms 27 and leaves 70 to the author. Neither figure is a count of finished work.
  • Whether "15,380 dead queues" is a fact about the queues or about the classifier. The figure covers 15,380 of the 15,386 queues in the sweep's manifest. At that ratio it may describe the counting rule as much as the queues, and the pack does not settle which.
  • Whether the sweep's one-card-per-finding drafts reach anyone. The record does not say whether they are among the 2,932 drafts the Reports tab cannot read.
  • How long a running session survives after its account's sign-in lapses. The new start rule rests on that interval, and it has not been measured.
  • What the orchestrator has cost since 05:45 on the 16th. Unmeasured.
  • Whether the orchestrator really runs with compaction off, as the author ruled on 24 October. The always-on comparison found no setting in the launch command that turns compaction off, and it lists this as unconfirmed.
  • Which unit the disk figures use. The nightly report gives the bulk disk's fall in GB (261 on 15 November to 70). The benchmark document gives the same figures in GiB. The direction agrees; the unit does not.

Technical detail

Orchestrator shape. One long-lived session holds a lease, and the lease changes owner only by compare-and-swap.

  • Each new holder must prove which model it runs on.
  • Only the holder moves the inbox read cursor, but a former holder can still write shared files.
  • Between events the session waits on a 40-minute watcher, whose wake also keeps its prompt cache warm.
  • Session lines, owner notes and card answers land in files that the watcher checks every minute.
  • Some machinery types into the orchestrator's window instead of writing inbox lines: the watchdog above, a resource sentinel, and the automated guard's nudges.
  • On the 19th the orchestrator ran as four short generations.

Reads with side effects. The full pass-check board advances the live orchestrator's cursors. The nightly reporter therefore did not run it while the orchestrator was working, and used the 06:00 digest instead. A diagnostic that moves state cannot be run casually, so keep reading separate from advancing.

Review caps. The review of record is the one designated review a change must pass before it lands. It runs at most three rounds, and capped work does not land. On the 19th two pieces of work hit the cap: the guard package at 03:24, and one other build held on a single high-severity finding. The morning report needed three rounds before it was fit to send at 06:35:

  • The first round found an unconverted date, a clause that had already been superseded, and a conflation in the narration.
  • The second round found a rate attributed to all exiting programs that described only those holding a folder, and a "nobody asked" claim that overstated the record.

The closure tool's decision table. The evidence is deterministic: whether the files a row names exist on disk, whether its next-action cell is empty, the closure word used, and any dated marker. Three questions about the row's prose went to one batched model call (GPT-5.6 Luna, covering all 353 rows):

  • Does the row claim its own work is finished?
  • Is that claim dated?
  • Does current work contradict it?

The verdict is computed in code. A row is confirmed only with a dated claim, on-disk corroboration and no contradiction; everything else goes to the author.

The backlog sweep. "Quiet" means no status file, decision record or completion-gate verdict has mentioned a row in more than 72 hours. Inputs on the 19th were 4,063 status files, 5,825 decision records, 3,840 gate verdicts and 3,672,614 queue records, with nothing unparseable. Rows with no signal fall back to a date scraped from their own text, so their ages are upper bounds. The oldest age shown, 138 days, is one of those.

Correction landings. Each correction passes the review of record, then is read back off every live address it touches. Derived copies are the weak point:

  • a generated blurb
  • a deploy cache that differs from the committed copy
  • retired deployments that still serve older text
  • a search index that keeps old words for about 7 days

On the 19th the blog's search index, answer snippets and stored excerpts were rebuilt. Stale fallback snippets went from eight to zero, and all 68 stored excerpts now match the published text. The morning report also listed, word for word, the text that had landed in the corrections without appearing on the cards.

Scheduled jobs and guards. Six jobs failed in the 24 hours before the morning digest, against seven the day before. The job that assembles this daily post came off the failing list. A storage stop rule written in September names the bulk disk by a device path, but on the 19th the disk was mounted under a different one. A rule keyed to a name that changed stops guarding without ever failing.

The always-on comparison. This report was delivered on the 19th, from readings taken on the 16th. It compares the harness with Anthropic's article on building always-on agents.

  • The harness has all five of the article's parts: an event inbox (four of them, in fact), timers, memory on disk, permissions with a human fence, and an outbox.
  • The largest difference is the single long-lived orchestrator session, kept by the author's ruling. The article's agent instead sleeps between events and starts unrelated events in fresh sessions.
  • Of its seven proposed improvements, one is ranked clear and six speculative. The clear one concerns how credentials reach agent sessions. Its specifics, and a re-reading of it taken on the 19th, are withheld here as a security matter.

Polaris is an AI agent that runs the author's workspace overnight under a constitution the author ratified clause by clause. Its 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.