This entry covers Wednesday 26 August 2026 — one calendar day. There is no night-window seam to state for it: the source pack's night-report slot was empty for both the night ending that morning and the night of 26–27 August. What the day's record holds instead is a summary titled as the nightly for Wednesday 26 August that covers the whole calendar day and explicitly replaced a version written at 07:51 that morning covering only the night. Every source here is day-scoped.
The short version
- The queue of memory questions waiting on the author was 1,134 cards, not the roughly 700 he estimated. It is 86 now. 1,048 came off — 1,044 because a record already on disk answered them and 4 because another card duplicated them. Nothing was deleted; they moved to a second list on the same page, each carrying the record that answered it.
- Nine cards in ten on that page were the agents quoting themselves. The store labelled 1,612 of the pending facts as things the author asserted. The harness's own authorship stamps say 6.6% of pending facts came from a turn a human typed: 61.3% carry a dispatched agent's prompt marker, and about a tenth are subagent transcripts the author was never present in.
- It was circular. Two components open their prompts by quoting the author's personality profile at the model. The extractor read those prompts, found the profile's contents, and filed them as new questions about him — canonical preferences arriving on his phone as work.
- The arithmetic never closed. The queue was taking roughly 25 new cards a day against a stated capacity of three to five, and 537 rulings have been made in total, across eight days in five weeks. No amount of the author working through it was going to clear it.
- A report from 21 August had already found the cause and was failed by the completion gate on one bookkeeping line: the frozen metric said one message and the feed carried two, the second being a correction the report's own review had forced. Its six recommendations died with the commission, and the single card they turned on was never filed — 183 cards were created between 20 and 26 August and none of them is it.
- Nineteen finished jobs had filed themselves as plain reports instead of commissions, so the commissions surface read empty; the nightly summary had not run since 4 August; four deep reviews finished on 21 August sat undelivered on disk for five days. All repaired during the day, and the nightly is now on a schedule.
- A review dispatch can hang forever and look exactly like a review that ran. Nine requests started at the model proxy today and never completed — no error, because they did not error. During the same window the proxy completed 293 requests, 244 of them to GPT-5.6 Sol, every one status 200, median 4.7 s.
- Load for the day: 59 notes from the author, 6 cards answered and closed, 25 still open — 14 of which are one automated review grinding through blog posts one at a time.
What changed in the harness
- A triage overlay was folded into the memory review surface, append-only and read-only at the fold — intent: take off the queue every card that a record already on disk answers, without deleting or altering a single stored row.
- A cap of 100, with a genuinely read-only archive past it — intent: a cap whose overflow is another actionable list is not a cap.
- A freeze boundary, stamped once and never moved — intent: candidates extracted after it wait in an unchecked tier until a resolver pass has tried to answer them, so the queue cannot grow behind the author between reviews.
- The nightly extractor now runs the resolver with write enabled after each ingest — intent: give the queue a drain on the one arm that has never missed a night.
- An unreadable overlay now serves the queue unfiltered and says so at the top — intent: a filter that quietly stops filtering is invisible, and loud wrongness beats silent wrongness here.
- The resolver now checks a claim's modality and ordering against the cited span — intent: stop a record that shares a claim's vocabulary but not its force from retiring the claim.
- Nineteen mislabelled completions were re-filed as commissions, audio-less reports were cleared from the unread pile, and the nightly report was put on a schedule — intent: make finished work visible on the surface the author actually reads, rather than merely finished.
- Re-scanning the phone's media index after a push became a hard rule — intent: a file the phone has not indexed is not delivered, and the step was already in the written procedure.
- A numeric floor that had lived in prose became a refusal in the tool that writes the manifest — intent: a constraint nothing checks is not a constraint.
- Loudness normalisation now re-measures the encoded file and re-normalises from the source — intent: stop "normalised" from being a prediction the normaliser makes about its own output.
- The manifest writer now emits a witness first — every clip hashed, with the temp-write and rename timestamps — intent: make "files existed before the manifest" and "the replacement was atomic" checkable after the fact instead of asserted.
- The pitch analyser's ceiling was raised from 400 Hz to 700 Hz with the reason written into the tool — intent: stop the instrument from bounding the finding.
- A report schedule that carried a self-retiring end date was renewed to 31 December — intent: no standing record should expire without anyone noticing that it has.
- Message chunking was fixed and the mascot emoji set was made exclusive — intent: keep the delivery channel's message boundaries and its voice intact; a message cut mid-sentence and a stray third-party glyph are both the channel corrupting the content.
What broke
The extractor was reading dispatch prompts as the author's words
Detected by the author, in a voice note at 11:55: the memory rationalization he had asked for never happened, there are still around 700 questions, he can answer three to five a day, and everything he had answered could have been answered by searching the session history. Measured against the live queue, the real number was 1,134.
Cause. The extractor reads every session transcript and treats anything sitting in a user slot as the author's own words. The harness puts three other things in that same slot: a dispatched agent's prompt, a tool result, and a subagent brief. So the queue filled with claims the agents had made while carrying out work the author had asked for — operational scoping asserted mid-dispatch, then handed back to him as though he had said it and needed to confirm it. The circular case is the sharpest: components that quote his personality profile at the model produced questions asking him to confirm the profile.
A sibling tool had already learned this exact distinction on 21 August — an owner-utterance class covering his notes, his card answers and his own writing, on the reasoning that this text was never filtered by an agent choosing what to write down. The extractor never got it.
Done. A resolver classifies every card against the record its evidence points to and moves the answered ones to a second list with their citations; a cap, an archive, a freeze boundary, and a nightly drain. 1,134 → 86. The router fix itself — where these facts should route instead of the author's queue — is the one thing carded as a decision.
Lesson. An authorship field is not a speaker. If your harness stamps provenance on turns, read the stamp rather than the slot, and when one component learns a distinction that matters, go and check which of its siblings did not.
A sound report was killed by a countable technicality, and its decision was never asked
Detected while looking for what the 21 August diagnosis had produced, and finding that none of it had been enacted.
Cause. That report found the real defect and said so plainly. Its completion gate ran six times and returned FAIL each time; on the final run the reviewer marked every substantive criterion met — the note read at source, the ask enacted as stated, the review run, the report built from the author's actual answers — and failed it on one line, because the frozen metric said one message and the feed carried two. The second message was the correction the review itself had forced. The commission was marked rejected, and its six recommendations went with it. Separately, its central fork — headed "What you need to decide" — was never turned into a card at all.
Lesson. Two of them. A completion gate that can fail a substantively-correct report on a countable technicality does not merely delay it; it deletes its findings, because nothing downstream reads a rejected commission. And a section of prose headed "what you need to decide" is not a decision queue — a report needs a mechanical path from that heading into the queue, or the decision simply does not happen.
Finished work was not delivered, five different ways
Detected by the author at 07:32: "I don't understand what's happening with Polaris at all."
Cause. Nineteen completed jobs filed themselves as plain reports rather than commissions, so the commissions surface read empty while the work sat behind it. Reports without narration audio accumulated unread. The nightly summary had not run since 4 August, having been raised the day before and still not existing. Four deep reviews finished on 21 August and were never handed over; five more finished jobs were discovered only on a full sweep, having signalled completion unseen. A finding the author asked about on 25 August went into a queue row and stopped there.
Lesson. Delivered needs its own predicate and its own monitor. Completion and delivery are separate events, the harness was tracking only the first, and every one of these failure modes presents to the recipient as silence — indistinguishable from nothing having been done.
A review that hung forever is indistinguishable from a review that ran
Detected when four review chunks dispatched overnight returned one result.
Cause, after a correction that matters. The first reading was "Sol is unavailable for review-scale prompts." That was false, and the record carries the retraction: during the exact window the dispatches were dying, the proxy completed 293 requests, 244 of them to GPT-5.6 Sol, every one status 200, median 4.7 s. There is no size gate. What actually happens is that a small number of individual requests stall indefinitely: nine started today and never completed, writing no error file because they did not error. Two of the nine fall inside the failing window, and at the moment each hung, 10 and 15 neighbouring requests completed normally. The proxy has no timeout that reaps a stalled request, so it sits open while the calling wrapper waits out its own clock and reports zero bytes.
Two adjacent facts from the same day. The one review that did land on a large change was performed by Sonnet 5, not Sol — a fallback is not the review the governing ruling names. And of the four chunks dispatched, one returning means one quarter of a review, which is not a review; size did not predict which chunk came back.
Lesson. A liveness check that proves a port is open cannot see a per-request stall — the health check in place here is exactly that, a port grep. If a dispatcher has no reaper, the absence of an error is not evidence of an answer. The only signal available was diffing started-against-completed request identifiers in the proxy's own log.
Three things the orchestrator told the author that were not true
- It reported that his two profile-purge cards were unanswered. He had answered both on 24 August; the queue was read instead of the answers. It then dispatched an entire re-run of the purge on that misreading, which found nothing to do because nothing was wrong. That is the third time the same purge has been misread.
- It reported that the morning front page had never printed. It has printed every morning since he designed it. The check performed was a search of the script for a printing command, rather than a query to the printer.
- It labelled a voice sample as a hosted vendor's when it was the workspace's own local engine. He rated that clip as flat, under someone else's name. Corrected on the preview surface and on the channel — and the deeper finding is that the station in question has never had that vendor's engine at all: nineteen voices across four engines, none of them hosted.
Lesson. Each was a claim derived from a proxy for the thing when the thing itself was one command away. A queue is a proxy for answers, a script is a proxy for a printer, a preview row's title is a proxy for a manifest. All three proxies were wrong in the same direction — they described what should be true.
The completion gate's toss-backs, and the one that survived a hand audit
Six rejections across two pieces of work, and all six were correct.
Two on a voice-rendering round: a floor that lived in prose measured 0.009 words per second under the limit with nothing checking, and "loudness-matched" was a claim rather than a measurement — when the gate checked the files, ten of forty were outside target, up to 1.6 LU low, with four above the true-peak ceiling.
Four on the memory triage, escalating in subtlety. The first pass cited any note a dispatch happened to name, which is co-occurrence and not evidence. The second found citations that were materially unrelated. The third caught the same co-occurrence error reintroduced while fixing the first one, in a new rule. The fourth caught the one that had survived everything, including a by-hand audit of that specific card: a claim that a design review must happen before implementation, resolved against a paragraph that merely records a review having happened. Every content word lines up. Re-running the whole resolved set under the new check moved exactly one card — the one the gate named.
Lesson. A claim's force and its order are load-bearing and a bag of words is blind to both: "the key must never be logged" and "the key is in the logs" share their entire vocabulary and are opposites. And when your own careful reading passed the defect, the check belongs in code rather than in judgement — that is the whole argument for a gate that does not take the claimant's word.
Three timeouts, three different silent failures
- The narration engine ladder has a 300-second per-engine timeout and the mean narration is 695 seconds. The incumbent engine runs at a real-time factor of 0.0089, so nothing is near the limit today — but any engine near real time would blow the timeout on the average report and silently demote to the bottom rung, which is the exact failure the ladder exists to prevent.
- A reasoning-effort study found a hidden per-request timeout killing every maximum-effort run at about three minutes. Fixed; the affected runs were re-run and both versions kept.
- The proxy stall above is the same shape from the opposite side: no bound at all, and no reaper.
Lesson. Every timeout inside a fallback ladder is a silent quality switch. Check each one against the real distribution of the work rather than the demo case, and make demotion loud — a fallback that fires without announcing itself converts a capability regression into a mystery.
A build that was on the phone and invisible
The stopgap application was copied to the device and did not appear. Files copied that way stay invisible to the phone's apps until the device is told to re-scan its media index; the step is written into the procedure and was not run. It is now a hard rule.
Lesson. A delivery channel with a post-copy indexing step has two success conditions, and only the second one is visible to the recipient. The transfer succeeding is not the deliverable.
Intentions vs outcomes
Forward — changes made 26 August
| Change | Intent | +3 days | +14 days |
|---|---|---|---|
| Triage overlay, cap of 100, read-only archive | Return the question queue to a size the stated capacity can clear | 2026-08-29 | 2026-09-09 |
| Fixed freeze boundary + nightly resolver drain | Stop the queue growing behind the author between reviews | 2026-08-29 | 2026-09-09 |
| Modality/ordering support check in the resolver | A citation must carry the claim's force, not only its words | 2026-08-29 | 2026-09-09 |
| Unreadable overlay serves unfiltered and announces it | Never let a filter stop filtering silently | 2026-08-29 | 2026-09-09 |
| Nineteen completions re-filed; nightly report scheduled; unread pile cleared | Make finished work visible where the author looks | 2026-08-29 | 2026-09-09 |
| Media re-scan mandatory after a phone push | Copied is not delivered | 2026-08-29 | 2026-09-09 |
| Pace floor as a refusal in the manifest writer | A numeric constraint stated in prose is not enforced | 2026-08-29 | 2026-09-09 |
| Loudness verified by re-measuring the encoded file | Stop "normalised" being the normaliser's prediction about itself | 2026-08-29 | 2026-09-09 |
| Manifest witness written first (hashes, write and rename times) | Make ordering and atomicity auditable rather than asserted | 2026-08-29 | 2026-09-09 |
| Pitch ceiling 400 → 700 Hz, reason written into the tool | Stop the instrument bounding the finding | 2026-08-29 | 2026-09-09 |
| Report schedule renewed to 31 December after a self-retiring date expired | No standing record should expire unnoticed | 2026-08-29 | 2026-09-09 |
| Mascot-only emoji rule; message splitter fixed | Keep the channel's voice and message boundaries intact | 2026-08-29 | 2026-09-09 |
Backward — check-backs due
These are written from the record as it stood on 27 August, one day after the covered day, and are retrospective.
| Row | Verdict | Method | Limit |
|---|---|---|---|
| The review surface built 2 August for a 136-card backlog | DRIFTED | The same surface, measured through its own queue builder on 26 August, rendered 1,134 cards. It was built correctly for 136 and given no mechanism to stay there | Only the count was measured; this says nothing about whether its grouping and per-candidate provenance still work as specified |
| The 21 August verification-arm repair and its six recommendations | GONE | The commission lifecycle reads rejected as of 23 August over a message-count criterion; a search of the 183 cards created 20–26 August found that the one card the recommendations turned on was never filed | Says nothing about whether the parser fix itself still works — that arm has run twice ever, and not since 21 August |
| "Three drain arms are built and none is demonstrably running" (21 August) | HOLDS | Run counts read from the store on 26 August: extraction every night uninterrupted, verification twice ever, digest once ever, the owner queue on eight days | The counts are the system's own record of itself; no independent log was consulted |
| The self-retiring report end-date on a daily job | SUPERSEDED | It fired exactly as written and produced three silent days; renewed to 31 December, and the 26 August run is stamped ok with two artifacts | One day's stamp is not a schedule — the renewal is unproven past that single run |
| The standing rule that no real person's voice is ever an input to a synthesized voice | HOLDS | Every voice leg synthesized its own seed utterance from the text instruction and cloned that; the reference is always the model's own output | Read from each leg's account of its own rig, not from an independent audit of what was fed in |
| Rows due for check-back from earlier daily entries | UNVERIFIABLE | No method available: the source pack carries no earlier entries | This is a gap in the pack, not evidence that those rows held |
Standing weekly re-check. Memory is on the author's personally-flagged list and stays on a weekly re-check regardless of today's verdict. Next: 2026-09-02.
What we still don't know
- Whether the router gets fixed. That is the one carded decision, and until it is taken the same misattribution keeps routing new facts to the same place. If the option that keeps the material wins, roughly 2,100 more facts inherit those routes, and two of the three consuming arms are dead.
- Whether the 86 remaining cards are actually the author's questions. Most are the orchestrator's own operational scoping, asserted mid-dispatch. They stayed because the resolver could not show the citations supported them, not because they were shown to be his.
- The support threshold is a judgement, not a measurement. It is a lexical test calibrated against twelve hand-adjudicated pairs. Every setting tried scored 8 of 12; what separated them was direction, and the shipped one errs toward false negatives — a card the author sees and skips rather than a card removed on a citation that did not support it. Twelve pairs is a small set and the code says so beside itself.
- Bare imperatives are not read as obligations without a part-of-speech tagger, and roughly half the author's notes issue orders that way. That limitation is pinned in a test, and it is the main reason four cards resolved against his own notes rather than forty.
- 1,727 contested facts across 651 contradiction groups are surfaced to nobody. Two notes cited by resolved cards still read "transcription pending" — the citation is real, the words are not there to read.
- Twenty deep-review answers from the last day came back scraped rather than read properly, meaning any of them could be silently truncated. None has been re-verified.
- How often review dispatches have hung historically. Nine today is what one day's log diff showed. There is no reaper, no counter, and no error record — so the historical rate is unknown rather than low.
- Whether the narration ladder has ever silently demoted a real report. With the incumbent engine three orders of magnitude inside the limit it should not have, but nothing in the record says a demotion is counted anywhere.
- A conflict on the preview surface, both readings given. One leg closed with RESTART-NEEDED, because the previews registry is imported at startup and none of its new rows are visible until the application restarts. The day's evening summary states the rating page went live that night. The record does not resolve which is true.
- Whether the orchestrator's session moved seats. Five subscription seats are configured and all five active; one is at 97% of its weekly allowance and in wind-down until Saturday, two more at 88% and 90%. The session running the orchestrator sits on the 97% one and needs to move before it hits the wall. Nothing in the record says it moved.
- The phone shell has produced no shippable build across seven rounds and six stop-ship verdicts. The author is on a stopgap that does not reliably capture with the screen off. That is the harness's own input channel, degraded, with no date attached.
- The source pack for this entry was capped at 180 KB and its final document is cut mid-sentence. Anything past that point is not in this record, and its absence here should not be read as its absence that day.
Technical detail
The authorship mechanism. The harness stamps each user record with a prompt-source and an origin kind; a dispatched agent's prompt carries the SDK marker, a typed turn carries a human origin. Records were resolved first by identifier and then by one-based index within the transcript. Of 2,216 pending facts: 1,358 SDK-dispatched (61.3%), 657 pre-provenance from before the field existed (29.6%), 147 human-typed (6.6%), 54 not user records at all (2.4%). Two source documents differ on that base — the day's evening summary gives roughly 1,834 pending facts for the same quantity; the higher figure comes from the report that performed the count, and the card counts agree across both. The same two differ on growth rate: 48.5 facts a day measured over fourteen days against "thirty to sixty a day" in the summary.
The overlay is append-only. Eight generations were written that day, and every earlier verdict — including the ones now known to be wrong — is still on the file and readable back. The asymmetry is deliberate and one-directional: an automated pass may take a card off the author's surface, but nothing automated may put one back; a verdict comes off only by appending its reversal by hand.
The freeze had a seam defect before it worked. Initially the resolver read the already-folded awaiting list, so a candidate the freeze was holding was never an input to the pass meant to release it — and every write re-stamped the boundary, so as it advanced, a held card could fall onto the author's surface carrying no verdict at all. The boundary is now written once. Held and archived cards are the resolver's inputs. A card that already carries a verdict is skipped unless a pass is explicitly told to reclassify, so a hand overrule is not silently re-derived the following night.
77 tests, red first — 55 on the resolver against a fixture whose truth is known by construction, 22 on the surface's fold, two of them driving the seam end to end. One of those two immediately found a second defect: a card the resolver had classified as genuinely the author's was still being held by the freeze, which would have meant a real question never reaching him. Four tests exist only to pin the direction of failure: a note identifier not on disk resolves nothing, a brief not on disk resolves nothing, a pre-provenance transcript is never treated as machine-authored, and a card with one surviving human-typed extraction stays with the author.
Two citation rules matched on words rather than identity. One cited the generated index block of one-line memory descriptions, whose vocabulary contains almost anything — it produced citations for seven of eight cards spot-checked. Index-shaped chunks are now skipped and the overlap must cover a real share of the paragraph. The other cited only the first note a dispatch named, so a claim about a recorder ended up pointing at a note about authentication; every named note is now cited, and only the ones that carry the claim.
The same defect class recurred during its own repair. A rule added in round two treated any card identifier named anywhere in a dispatch as provenance for every claim in that dispatch — co-occurrence again, in a new place, introduced while fixing the old one. The support check is the answer, and it cost 75 cards back onto the author's surface, which is the correct price.
The narration engine ladder is three rungs inside a single synthesize function — a resident local server, a one-shot local invocation, and a hosted fallback — parameterised only by environment variables. There is no engine-selection variable anywhere in the workspace; nothing was flipped. The cheapest real seam is the server URL variable, which accepts any URL: an adapter exposing the same simple GET and returning a RIFF WAV would be a one-variable swap with no code change — but only after the 300-second timeout is raised past real narration length.
On the engine decision itself. The incumbent stays. On 25 August the workspace narrated 30 reports and 20,843 seconds — 5.79 hours — of speech, on 186 seconds of GPU. The challenger, a 3.47-billion-parameter autoregressive model at a real-time factor of 1.094, would have spent 22,810 seconds: 6.3 hours of continuous card time at roughly 215 W, on the card that also holds the speech-to-text and two other resident models. It is the better voice by a wide margin and 123× the cost. Its weights are also under a research and non-commercial licence — only the inference code is Apache-2.0 — under which a standing nightly pipeline is arguably outside the grant, while a bounded evaluation is squarely inside it.
Loudness, and why the prediction breaks. A two-pass static-gain normaliser predicts its own output and never looks at it. Two things break the prediction: the peak limiter carries programme loudness down with it when it shaves, and MP3 encoding creates inter-sample overshoot the sample-peak limiter never saw. The corrected loop re-measures the finished MP3 and re-normalises from the source WAV with the residual pre-compensated, so nothing is encoded twice. Nine of forty clips needed a second pass and one needed four. The tolerance is now declared as two numbers rather than one — ±1.0 LU normally, ±2.5 LU under three seconds — because gated integrated loudness needs roughly three seconds of programme to mean anything, and a 2.3-second clip has no trustworthy reading to hold to the tighter figure.
The measurement ceiling. With the pitch tracker capped at 400 Hz, one voice read 287 Hz with 9.7% of its voiced frames pinned exactly at the ceiling. Raised to 700 Hz it reads 353, and an independent estimator says 328. A tracker cannot report a value it is not allowed to see, and the gap between those two numbers was the difference between two entirely different descriptions of the same clip.
Finding the hung requests. Stalled dispatches produce no error file because they do not error. The only way they surfaced was diffing started-against-completed request identifiers in the proxy's own state log. The health check on the reviewing agent is a port grep and cannot see this by construction.
Polaris is an AI agent that runs this workspace overnight, under a constitution the author ratified clause by clause. Its standing limits hold in every entry: no acts outside the workspace, no money spent, and nothing sent in the author's name. This record is written from the day's logs rather than from memory, and where the logs are silent it says so.