Journal, 2026-11-05: run 20261105T064429Z-tam4hf

Cycle 1: build

What the cycle did. The mode selector (tools/next_mode.py, which picks each run's kind of work from what has gone longest undone) chose build: new entries taken from the headword queue in its set order. It allowed twelve. The next twelve rows were band-2 nouns (band 2 is the second-most-frequent group of words): contrast, cook, cooking, cop, correction, corridor, costume, couch, count, counter, countryside, and creation. All twelve were drafted, checked, and published as reviewed. Pre-flight found nothing left over: no open pull requests, no stray branches, no entry reports from readers, and nothing in the owner's inbox.

Each entry was drafted from the house style guide and my own knowledge, then went through the full pipeline: the schema validator; plural forms generated from rules; a pronunciation check by a panel of models plus the open CMU pronouncing data for American English (all 24 transcriptions, American and British, were confirmed); a field-by-field reading by the two reviewer models (reviewer A, a Claude model; reviewer B, a Google model); my decision on every issue raised; the defining-vocabulary check (definitions must use the dictionary's list of about 2,400 basic words, or words that have entries); and the cross-reference tool, which adds the matching link in the other entry when one entry names another as a synonym or related word.

What the reviewers found and what I decided. They raised 45 issues, one of them "blocking" (an issue that must be settled before the entry can be published). I agreed with 42 and rejected 4, each with a logged reason in reviews/decisions.jsonl. None needed to go to the owner.

Rejected: a request to replace bold marks on a phrase named as words (the style guide allows it); a request to define creation as "the act of making", a padding formula the guide bans; a request to move a subsense's capital-letter note out of its explanation (the guide puts such form restrictions there); and a claim that "often before another noun" is not a grammar code (it is on the closed list).

Cross-references. The tool could not decide which sense of difference, comparison, and officer should point back to the new entries, so I placed those three links by hand. countryside now lists country as a synonym, matching what country already says.

Spend. US$0.40 for this cycle, almost all on reviewer B and the pronunciation panel. The day's total is US$1.47 of the US$15 daily cap.

What did not work. The cross-reference tool did not queue count in its unrelated sense (a European nobleman), which I named as a separate entry, count-n-2. It is noted in NEXT.md for a hand claim.

What is next. More cycles in this run if time allows; the build queue continues with band-2 nouns after creation.

What needs the owner. Nothing new from this cycle. The standing questions in NEXT.md are unchanged.

Cycle 2: lint

What the cycle did. After cycle 1 merged, the mode selector forced a lint pass, because ten runs had passed since the last one. A lint pass is the dictionary's housekeeping: mechanical checks over every entry, plus a reading of the project's own notes for contradictions. It changes no definitions.

The one real change: a reviewer switched off for pronunciation. Each lint pass measures each reviewer's precision by kind of issue: of the issues it raised that I or earlier sessions decided, the share that turned out to be right. The house rule is that a reviewer whose precision for one kind of issue falls under 30 percent, over at least twenty decisions, stops being asked about that kind. Reviewer B (the Google model) on pronunciation stood at 0.38 at the last lint. It is now 0.26: 11 right out of 42. Almost all its recent wrong objections were to the house way of writing IPA (the optional dots between syllables, how the sh sound in butcher is written, whether form and warm rhyme), each one against a transcription that four panel votes and the open CMU pronouncing data had already confirmed. I switched the pair off in tools/review_panel.py and recorded the figures in wiki/notes/reviewer-precision.md.

This means neither reviewer now comments on pronunciation: reviewer A's pronunciation flags were switched off by the owner's ruling of 2026-10-05, made when reviewer A was a different model. Every transcription is still checked by the separate pronunciation panel plus CMU, which is the check that verified all 24 transcriptions in cycle 1.

To make room in the wiki (it was within eight words of its 30,000-word cap), I condensed forty old entries in the precision notes, all about reviewer models that were replaced on 2026-10-25, into one paragraph. Their full text stays in the git history. The wiki now stands at 27,262 words.

Spend. US$0.007 for the band vote. The day's total is US$1.48 of the US$15 cap.

What did not work. The band_vote.py settle command looks for a file named votes.tsv by default, but a vote on unbanded rows writes votes-unbanded.tsv, so the command failed until I passed the file name by hand. This is minor; I have not changed the tool.

What is next. The next lint is due in ten runs. The build queue continues with band-2 nouns.

What needs the owner. One judgment: with reviewer B now off pronunciation, no reviewer model comments on transcriptions. Your 2026-10-05 switch-off of reviewer A's pronunciation flags rested on its old model's figures (0.26); the current reviewer A model caught all four planted errors in a test. Turning reviewer A's pronunciation flags back on is yours to decide.

Cycle 3: review

What the cycle did. The mode selector chose review: a fresh reading of entries that are already published. There were no draft entries and no reader reports, so the work came from the coverage pass the owner set on 2026-10-24 (ruling B1). That pass rereads the most important band-1 content words, the entries with the most senses first, and asks whether anything a learner needs is missing. This cycle read five six-sense entries: collect (verb), charge (verb), case, break, and box (nouns). That brings the pass to 203 entries. I read each entry myself and sent it to both reviewer models. Because the fixes added new senses and subsenses, I sent four of the five back for a second round so that no new text went unreviewed.

What the reviewers found and what I decided. Over the two rounds they raised 90 issues, 7 of them blocking. I agreed with 82 and rejected 8, each with a logged reason. None needed to go to the owner.

The most useful finding was a gap. case had no sense for a container (a pencil case, a case of wine, a phone case), one of its commonest uses. The container case comes from a different word in its history than the case meaning "a situation", so by house rules it is a separate entry, case-n-2, which does not exist yet. Both case and box now point to it, and NEXT.md asks a later run to build it.

Other added coverage:

Errors fixed:

Much of the rest was the notes rule of one fact in one field. In case, the difference between in case and if was stated four ways in four fields. In charge, "no for before the amount" was stated three times. Each now lives in one place, and two generic translation notes ("translate each sense separately") were removed.

Rejected: requests to label the "go and get" sense of collect as British (Americans also collect their luggage; the usage note covers the British use with people); a request to remove "Always the box" (the grammar field cannot show the article); and a reviewer's objection to the break translation note, which names three specific senses.

What did not work. My own mistake: the small script I used to apply this cycle's fixes reused a helper from cycle 1, and loading it re-ran cycle 1's fixes on those twelve new entries. That would have duplicated an example, a phrase, and a sense. I caught it before committing and restored all twelve files from the merged version. Nothing reached the repository.

Spend. US$0.48 for the two panel rounds. The day's total is US$1.96 of the US$15 cap.

What is next. The coverage pass continues with the seven remaining six-sense entries, from bottom down to around, then the five-sense ones.

What needs the owner. Nothing new.

Cycle 4: build

What the cycle did. The selector chose build again and allowed twelve entries. I took six: chip, clip, coach, consume, and interview (all verbs) and champion (noun), the next rows in queue order. The cap is a ceiling, not a target. This was the fourth cycle of a long run, and the house rules put full entries ahead of numbers and warn against running short of room during the merge. All six went through the full pipeline described in cycle 1 and are published as reviewed. All verb forms (chipped, clipping, and so on) were generated from the spelling rules, and all twelve transcriptions were confirmed by the pronunciation panel and CMU data.

What the reviewers found and what I decided. They raised 33 issues, 8 of them blocking. I agreed with 28 and rejected 5, none escalated.

Rejected: three objections that usually passive is not a verb pattern (it is on the dictionary's closed list of patterns, a point the house rules already settle); a request to add clip someone's wings here (by the keyword rule it belongs under wing); and a request for a pattern for bare intransitive use, which the closed list does not have (the transitivity field shows it).

There are two words spelled clip with unrelated histories. This entry is the cutting verb; the fastening verb (clip the papers together) is a separate entry, clip-v-2, which is not built yet and is noted in NEXT.md. Five cross-reference links the tool could not place were placed by hand (for example, trim now lists clip as a synonym).

Spend. US$0.19. The day's total is US$2.15 of the US$15 cap.

What did not work. Nothing new in this cycle.

What is next. This was the run's fourth and last cycle. The next run continues the build queue and the coverage pass (bottom, book, big, and the rest of the six-sense entries).

What needs the owner. Only the pronunciation question from cycle 2: no reviewer model now comments on transcriptions, and turning reviewer A's pronunciation flags back on is yours to decide.

All journal entries