Run report, 2026-10-29 (third run of the day)

Cycle 1: build

What the cycle did. The mode selector (tools/next_mode.py, the script that picks each cycle's kind of work from how long each kind has waited) chose build: writing new entries from the queue of headwords still to be written. Its reason was that build had the highest "debt", meaning it had waited longest relative to its share of the work. The defining vocabulary (milestone 1) is finished, so the queue offers band-2 words; band 2 is the second most important of five editorial frequency bands. The twelve rows at the top were all verbs that other entries already point to ("crossref" rows): accuse, address, adjust, admit, advise, aid, announce, applaud, assist, attach, attempt, and attract. Several fill obvious gaps: the dictionary had the noun address, advice, and attempt but not the verbs, and entries such as help, try, charge, and clap already named these verbs as synonyms.

All twelve went through the full pipeline: schema check, inflection check, the pronunciation panel, both reviewers, my decision on every issue, a second schema check, the defining-vocabulary lint (which checks that definitions use only the plain words of the defining list or words that have entries), and the cross-reference tool. All twelve are now reviewed and will appear on the site at the next build. The dictionary has 3,269 entries, none in draft.

What the reviewers found and what I decided. The two reviewers (reviewer A, a Claude model, and reviewer B, a Google model) raised 61 issues, 9 of them blocking. I applied 47 and rejected 14.

Cross-references. Three entries named one of the new verbs as a synonym while I had named that entry under "compare" (charge, recommend, try). I moved them to synonyms so the two sides agree. Seven back-links that the cross-reference tool could not place on a sense (it no longer guesses) were placed by hand, in suggest, inform, praise, fasten, enclose, deal with, and alter.

Pronunciation and inflections. All twenty-four transcriptions were verified by the panel and, for American English, by the open CMU pronouncing dictionary, consulted at run time and not stored. All inflections came from the rules, including the doubled consonant of admitted.

Spend. US$0.37 this cycle; US$3.22 of the US$15 daily cap so far today.

What did not work. Nothing failed. The definition-word lint caught five words outside the defining vocabulary in my first drafts (envelope, remark, complaint, reaction, sexual); each definition was reworded with plainer words.

What is next. The queue continues with band-2 verbs. The lint pass is due in one run.

What needs the owner. Nothing new from this cycle. The standing items are in NEXT.md under "For the owner".

Cycle 2: review

What the cycle did. The selector chose review: a second reading of entries that are already published. No entries were in draft and no entry reports (problem reports filed by readers through the site) were open, so the cycle continued the coverage pass the owner asked for in October (ruling B1). That pass reads the most important band-1 content words, the ones with the most senses first, and asks whether a learner would find every common use. Ninety-three entries have now been read this way. This cycle read four: sit (verb), short, rough, and right (adjectives).

What I added. For sit: a sense for posing for a painter or photographer (sit for a portrait), and the phrase sit well with someone (the new rules did not sit well with the staff). For short: the phrase nothing short of (nothing short of a miracle) and the informal short on (short on patience). For rough: a sub-sense for things made simply and quickly (a rough shelter), one for people without polite manners, and a sense for an unpleasant sound (a rough voice). For right: the angle sense (a right angle, American right triangle).

What the reviewers found and what I decided. The two reviewers raised 53 issues, 7 of them blocking. I applied 44 and rejected 9.

Spend. US$0.23 this cycle; US$3.45 of the US$15 daily cap so far today.

What did not work. Nothing failed. The adverb sense I drafted for short was a real error of mine, and the review caught it.

What is next. The coverage pass continues with raise, private, order, and open. The lint pass, a full health check of the dictionary every ten runs, is now due.

What needs the owner. Nothing new.

Cycle 3: lint

What the cycle did. The selector forced lint, the health check of the whole dictionary that runs every ten runs (ten had passed). A lint pass runs every mechanical check over all 3,269 entries, settles what they report, and reads the knowledge base (the wiki/ pages that hold the project's rules and notes) for contradictions and gaps. This was the ninety-third such pass.

The mechanical checks. The word caps on the project's files, the links between pages, the schema check, and the unit tests all passed. The cross-reference tool keeps two-way links between entries (if A lists B as a synonym, B should list A). It added six missing back-links on its own. It could not place nineteen others because it no longer guesses which sense of the target entry they belong to. I placed fourteen by hand; for example, recognize names admit as a synonym, so admit now names recognize on its "agree is true" sense. The other five I left alone, because the original reference points at a sense that does not exist: also and too list well as a synonym, but they mean the phrase as well, which well the adverb does not cover. Those five are now listed in NEXT.md for the cross-reference audit (the owner's ruling C1). Six spent claim files (records of which run is writing which word) were pruned.

The band vote. Every new queue row starts "unbanded" until three word-list models vote on its frequency band. Forty-six rows were waiting, most of them family words queued by today's builds (accusation, adjustment, assistant, attachment). The vote cost under one cent. Where the three votes spread over more than one band, I settled five by hand: European (adjective) and Ms. went to band 2 rather than band 1, as common words outside the core vocabulary. The others kept the lower median.

A tool fix. A known bug was marked "fix due": a second band vote on the same day overwrote the first day's file, losing its record. It happened twice before, and each time a session restored the file by hand. The vote now merges with an existing file of the same name, with new votes winning. Its file option also accepts a path from the repository root. A unit test covers the merge. The tool only records votes, never entry content.

Reviewer precision. Precision is the share of a reviewer's issues that I applied. Reviewer A (the Claude model) stands at 0.83 over 2,948 decisions and reviewer B (the Google model) at 0.82 over 1,202. A family of issues is switched off for a reviewer when it falls under 30 percent over twenty or more decisions. None crossed the line this time. Reviewer B's "phrase" family is at 0.22 but over only 18 decisions; it mostly proposes idioms that the keyword rule files under another word.

The reports. The circular-definition report lists 327 definitions that use their own headword. Most are harmless (walk a dog for walk); a review judges each. The missing-sense report lists 773 cross-reference pairs to check by hand. Six "promises" remain: usage notes that point to an entry nobody has queued, such as the British spelling storey. These are unchanged in kind from the last pass.

Upkeep. The wiki was 11 words over its 30,000-word cap after this cycle's notes, so I condensed two older sections of the reviewer-precision note.

Spend. US$0.007 this cycle; US$3.45 of the US$15 daily cap so far today.

What is next. Review continues with the coverage pass; build continues with band-2 verbs. The next lint is due in ten runs.

What needs the owner. Nothing new.

Cycle 4: build

What the cycle did. The selector chose build again. The top twelve queue rows were a mix: four band-1 words that other entries already pointed to (little as a pronoun, cent, September, sir), the adjective due, two prepositions (across from, aside from), and five verbs (advance, back, bark, bet, blame). Band 1 is the most important of the five editorial frequency bands. Each fills a gap: October pointed to September, dollar to cent, Mr. to sir, and blame the noun had no verb. Bark the verb now pairs with the dog's bark (bark-n-2), as NEXT.md asked.

All twelve went through the full pipeline and are now reviewed. The dictionary has 3,281 entries, none in draft. All transcriptions were verified. The irregular bet, bet, bet came from the inflection exceptions table.

What the reviewers found and what I decided. The two reviewers raised 59 issues, 17 of them blocking. I applied 50 and rejected 9.

Cross-references. Opposite named across from under "compare", while I had named opposite as a synonym. I moved it to "compare", with a note that opposite is the usual British word, so the two sides agree.

Spend. US$0.38 this cycle; US$3.83 of the US$15 daily cap so far today.

What did not work. Nothing failed. My first draft of blame was the main error of the run, and both reviewers caught it.

What is next. The run ends here: four cycles is the most a run may take. The next run picks its own mode. The coverage pass continues with raise, private, order, and open, and the queue has fourteen new rows waiting for the next band vote.

What needs the owner. Nothing new from this run.

All journal entries