Run report, 2026-11-04

Cycle 1: site

What this cycle did. The scheduler (tools/next_mode.py, which picks each cycle's kind of work from whichever kind is most overdue) chose a site cycle. Its reason was "highest scheduler debt among eligible modes": of the six kinds of work it can choose (building new entries, reviewing existing ones, closure, site, lint, and the originality check), the site check had gone longest relative to how often it should run. A site cycle writes no entries. It rebuilds the public website from the entry files and then opens the result in a real browser to check that it works.

Before the work. The usual pre-flight checks found nothing waiting. There were no open pull requests left behind by an earlier run (a pull request is the proposed change that GitHub tests and then merges into the published copy), no leftover working branches, no entry reports from readers (these arrive as GitHub issues titled "Entry report: ..."), and nothing in inbox/, the folder where you leave instructions. The day's spending budget had just reset at midnight UTC: yesterday closed at US$3.92 of the US$15 daily cap.

The build. tools/build_site.py turned all 3,675 entries into 3,083 headword pages; an entry is one word in one part of speech, so a word like run with a noun and a verb entry shares a page. It wrote 6,818 files in all (entry pages, the search index, previews and the home page). No drafts had to be held back, because every entry is reviewed, and there are no redirects (pages for renamed words).

The check. tools/site_check.py drives a headless Chromium browser at a phone width (360 pixels) and a desktop width (1,280 pixels). It ran 79 checks and all 79 passed. Among them:

No JavaScript errors were reported during any of these steps. Nothing was broken, so nothing in the site builder or its templates changed. The built site (docs/) is never committed: GitHub Actions rebuilds it after the merge.

What the reviewers found. Nothing to report. A site cycle does not call the reviewers, and no entry was opened or changed.

Spend. US$0. A site cycle makes no paid calls.

What did not work. Nothing failed. The local gate also passed (entry validation, size caps on the project's own files, internal links, unit tests, the defining-vocabulary check, cross-references). It reported warnings only, the same standing ones as before: for example, 168 validation warnings about overlapping notes, and cross-references to words that have no entry yet, which ruling C2 allows.

What is next. The scheduler will choose the next cycle's work. Still open: the audit of mirrored cross-references (C1), the coverage pass over the most important everyday words (177 entries read so far), band-2 nouns from the queue after bride, and closure rows from the word families. The next lint pass, the periodic maintenance sweep, is due in about seven runs. The weekly digest is due on 2026-11-07.

What needs you. Nothing new from this cycle. The standing items are listed under "For the owner" in NEXT.md: names used as family names after a title, the wording about reviewers in PROJECT.md, your 2026-10-05 switch-off of reviewer A's pronunciation flags, and the open questions in wiki/open-questions.md.

Cycle 2: build

What this cycle did. The scheduler chose a build cycle, which writes new entries from the queue (the ordered list of words still to be written). Its limit was twelve new entries and US$3 of paid calls. I wrote ten, the next ten band-2 nouns in queue order after bride. Band 2 is the second of the dictionary's five frequency groups: common words, but less central than band 1. The ten are broadcast, brown, browser, budget, builder, bulb, bull, burden, businessman, and butcher, all nouns. I took ten, not twelve, so that each entry got a full draft.

What the entries contain. Each entry has signposted senses (a short label that helps the reader pick the right meaning), examples, collocations (the words a noun usually goes with, grouped by type), and cross-references. Some points worth knowing:

All 20 transcriptions (American and British for each word) were confirmed by the pronunciation panel, a group of language models whose votes are checked against the open CMU pronouncing dictionary. Plurals came from the inflection rules; businessmen is handled by the rule for words ending in -man.

What the reviewers found and what I decided. The two reviewers, language models from two different companies that check every field, raised 42 issues, 2 of them blocking (they must be settled before an entry is published). I applied 34 and rejected 8.

The vocabulary check found four words outside the defining vocabulary (the plain words definitions are written in) in two definitions: onions, tulips, elephants, whales. I reworded those definitions; the words stay in the examples. One cross-reference back-link, from program (the TV or radio sense) to broadcast, was placed by hand because the tool could not tell which sense to use.

Spend. US$0.28, for the pronunciation panel and reviewer B. Reviewer A runs on your Claude subscription at no charge.

What did not work. Nothing failed. My first claim of one idiom, beast of burden, was added to the queue as a noun by mistake and then deferred. The deferral note records why.

What is next. Band-2 nouns continue after butcher (buyer, cab, cafe, ...). The other open items are in NEXT.md.

What needs you. Nothing new from this cycle.

Cycle 3: review

What this cycle did. The scheduler chose a review cycle. There were no unreviewed drafts and no entry reports from readers, so the cycle continued the coverage pass. That pass is a careful reading of the most important everyday words (band 1), asking of each entry whether it is missing a meaning that a learner will meet. This cycle read the next four six-sense entries not read since your rulings of 2026-10-24: language (noun), knock and kill (verbs), and key (noun). That brings the pass to 181 entries.

What was added.

What was cut or corrected. All four entries stated the same fact in two or three places, against the "one fact, one field" rule. For example, knock on the door was explained in the usage note, a learner-error note, and a translator's note; key to was explained three times. Each fact now lives in one field. In knock, two examples were really the phrasal verbs knock over and knock down, which have entries of their own, so they were replaced. The "criticize" meaning (don't knock it until you've tried it) moved above the rarer "make a hole" meaning. In language, the meaning covering sign language and body language said "without words". That is wrong for sign languages, which have words, and it now says "without spoken words". An example claiming that bees have a language was replaced, because examples may not state science facts. Tongue was removed as a synonym of language, since in that meaning it is literary (mother tongue). Following the rule that a removed cross-reference loses its mirror too, the link back from tongue became a plain "see also".

What the reviewers found and what I decided. The two reviewers raised 51 issues, 8 of them blocking. I applied 47 and rejected 4. The blocking issues were all fair, and all were applied:

Rejected:

Spend. US$0.22. The day's total so far is US$0.50 of US$15.

What did not work. Nothing failed. More than a third of the issues on knock were minor, but they were mostly accurate, so this does not count as reviewer noise.

What is next. The coverage pass continues with the remaining 25 unread six-sense entries, starting from interest (in reverse alphabetical order), then the five-sense ones.

What needs you. Nothing new from this cycle.

Cycle 4: closure

What this cycle did. The scheduler chose a closure cycle. Closure fills gaps in the words the dictionary already points to. Here that meant "family" rows: words listed in another entry's word family (its derived and related words) that had no entry of their own yet. Under your ruling A3, closure takes only band 1 and 2 words for now. This was the run's fourth and last cycle, so I kept it to six entries: calculation, calculator, cashier, catch, celebration, and chemical, all nouns. They complete the families of calculate, cash, catch, celebrate, and chemical (adjective).

What the entries contain.

All 12 transcriptions were confirmed by the pronunciation panel.

What the reviewers found and what I decided. The reviewers raised 22 issues, 1 of them blocking. I applied 21 and rejected 1.

Spend. US$0.18. The day's total after four cycles is US$0.68 of the US$15 cap.

What did not work. Nothing failed. NEXT.md went one word over its 1,000-word cap twice today, once after cycle 1 (that one was pushed before I noticed, and fixed in a second commit) and once here. Both were trimmed before merging.

What is next. Family rows continue in queue order after chemical (chemistry, chewing gum, ...), alongside band-2 nouns after butcher and the coverage pass from interest. The next lint pass, the periodic maintenance sweep, is due in about four runs, and the weekly digest on 2026-11-07.

What needs you. Nothing new today. The standing items are in NEXT.md under "For the owner".

All journal entries