Run report, 2026-10-28 (second run of the day)

Cycle 1: lint

What a lint pass is. Every ten runs the project stops adding to the dictionary and checks its own house: whether the word caps on its rule files hold, whether every link inside the knowledge base (the wiki/ folder of rules and notes) still points somewhere, whether cross-references between entries agree in both directions, and how often each of the two automated reviewers is right. The selector tool forced this pass because ten runs had passed since the last one (2026-10-27). This is the ninety-first such pass.

What was checked and what changed.

Spend. US$0.008 this cycle; US$1.53 of the US$15 daily budget used so far today.

What did not work. Nothing failed. The repeated mis-queueing of plurals is the one standing tool problem worth fixing soon.

What is next. The selector picks the next cycle's mode. The queue's standing items are the coverage pass over the most important band-1 entries (next: serve, school, save, rise), building the rows of the closure queue, and the cross-reference audit.

For the owner. Nothing new. The open items are unchanged: your ruling on names after titles (Ms. Kim), on continent names in definitions, and on the former and the following; and the line in PROJECT.md that still says both reviewers are non-Anthropic, which only you can edit. The three leftover claude/gallant-carson-* branches can be deleted at any time.

Cycle 2: build

What a build cycle is. A build cycle drafts new dictionary entries from the queue (the ordered list of words still to write). Each entry then goes through the full pipeline: a format check, generated word forms, a pronunciation check by three outside models, and a field-by-field reading by two reviewers from different labs (reviewer A, a Claude model; reviewer B, Gemini 4 Argon). I settle every objection myself before an entry is marked reviewed. The selector chose this mode because building had the largest backlog among the modes it could pick. It allowed up to twelve entries. I took ten, so that each could be full.

What was built. Ten common nouns from band 1 (the most frequent words): kid, lady, living room, luck, magazine, model, north, November, orange (the fruit and the color), and passport. All twenty transcriptions (American and British for each word) were confirmed by the pronunciation panel. The richest entries are luck (two senses and eleven fixed phrases, from good luck to the luck of the draw), model (seven senses, from a small copy to a fashion model to a computer model), and lady (five senses, plus ladies and gentlemen and First Lady).

What the reviewers found and what I decided. I made 60 decisions: 53 accepted, 7 rejected. Eight objections were "blocking", meaning the entry cannot be marked reviewed until they are settled. The main ones:

The rejections were these. A request for a "figurative" label, a value the closed label sets do not have. A request for a "rude" attitude label for the way some Americans address a stranger (Hey, lady): no such value exists either, and the explanation already says it. Reviewer B's view that a note on where magazine is stressed repeats the transcriptions; house practice keeps such notes because learners cannot read IPA (the phonetic alphabet). A request to reorder model's senses. And reviewer B's objection to bold marks on a phrase named as a phrase, which the style guide allows.

Other checks. The checker that keeps definitions inside the defining vocabulary (the core word list that definitions are written in) failed at first on Britain, lord, knight, and eleventh. I rewrote those definitions: the British region label now carries the place, and November is "the month before December, near the end of the year". Two bad queue rows that the vocabulary tool added were marked duplicates (advertisements, eleventh as a noun). Cross-references were made symmetric: copy, system, and woman now point back to the new entries.

Spend. US$0.36 this cycle; US$1.89 of US$15 for the day.

What is next. The next build rows are percent, pound, region, and response.

For the owner. Nothing new. The curator question about list names used as family names after a title now also covers Lady Kim and Lord and Lady Lee in lady-n.

Cycle 3: review

What a review cycle is. A review cycle rereads entries that are already published and fixes what is missing or wrong. Under your ruling B1, the current job is the "coverage pass". It goes through the most important band-1 words (band 1 is the most frequent band), starting with those that have the most senses, and asks whether a learner would find every common use. Both reviewers read each entry again, field by field, and I settle what they raise. This cycle covered four verbs and nouns with seven senses each: serve, school, save, and rise. That brings the pass to sixty-nine entries.

What was added.

What the reviewers found and what I decided. There were 40 issues, none blocking. I accepted 33 and rejected 7. Among the accepted ones: a definition of school that used the word school twice; a vague sun-and-moon definition ("appear in the sky", now "come up into the sky"); an example placing a rising river under "move upward" rather than "increase"; and a British-sounding example (it'll save me going out).

The rejections, with reasons:

Other checks. The defining-vocabulary checker caught the word evil in the new save definition, and I reworded it. All cross-references remain symmetric.

Spend. US$0.24 this cycle; US$2.13 of US$15 for the day.

What is next. The coverage pass continues with regular, read, range, and pull.

For the owner. Nothing new. The curator line on list names used as family names after a title now also lists lady-n.

Cycle 4: build

What this cycle did. This was a second build cycle, with the same pipeline as cycle 2: drafting, generated word forms, a pronunciation check, both reviewers, and my adjudication. The selector allowed twelve entries. I took eight, because several of them (shot, spot, pound) need many senses and phrases to be complete.

What was built. Eight band-1 nouns: percent, pound, region, response, salad, shot, smartphone, and spot. All sixteen transcriptions were confirmed. The fullest are shot, with seven senses (from a gun, in sport, an attempt, a photograph, an injection, a drink, and a person who shoots) and eight phrases (a long shot, call the shots, a shot in the dark, a cheap shot, and others), and spot, with five senses and six phrases (on the spot, hit the spot, a blind spot, and others). One line was added to the project's table of irregular word forms: percent has no plural after a number (ten percent).

What the reviewers found and what I decided. I made 56 decisions: 52 accepted, 4 rejected. Five objections were blocking.

The other rejections were the "formal" label on in the region of (it suits reports and news) and the second and third requests for the dog-pound sense.

Other checks. The defining-vocabulary checker caught six words outside the core list in my definitions: grams, kingdom, lettuce, pasta, alcoholic, and mobile. I rewrote each definition in plainer words. For example, smartphone is now "a phone that you carry with you and that can also do many of the things a computer does". Two back-links were placed by hand (attempt to shot, position to spot).

Spend. US$0.28 this cycle; US$2.42 of US$15 for the day. The whole run (four cycles) cost about US$0.89.

What is next. Build rows supermarket, T-shirt, tonight (noun), and TV. The coverage pass continues with regular, read, range, and pull. bat-n-2 and pound-n-2 wait to be built by hand.

For the owner. Nothing new beyond cycle 1's list. This run used all four cycles; the four pull requests merged cleanly.

All journal entries