Run report, 2026-11-10 (run 20261110T004355Z-pgu4i9)

Cycle 1: site

What this cycle did. The selector (the tool that picks each cycle's job from whichever kind of work is most overdue) chose site: rebuild the public website from the entries and check that it works. The reason it gave was "highest scheduler debt", meaning that among the jobs allowed to run, a site check had gone longest without being done relative to how often it should happen. The last check was on 7 November, and three days of building and reviewing since then have added nearly two hundred entries.

The site builder turned the dictionary's 4,090 entries into 3,421 headword pages. There are fewer pages than entries because words that share a spelling, such as hammer the noun and hammer the verb, share one page. No entry was held back as an unfinished draft, because every entry in the dictionary has finished review (its status is reviewed). The built site lives in a folder called docs/. It is never committed to the repository; GitHub builds and publishes it separately.

What the check found. The site checker opens the built pages in a real browser, at a phone width and a desktop width, and runs 79 checks. Among them: the search box finds headwords and phrases, and offers a suggestion when a word is misspelled (abadnon finds abandon); a page that does not exist still searches for the word in its address; the menu of senses on a long entry jumps to the right sense; dark mode shows light text on a dark background; a cross-reference to a word with no entry yet is shown in grey, not as a broken link; and no page scrolls sideways on a 360-pixel phone screen. That last check covers all 3,427 pages, including the home page and the help pages. All 79 checks passed. Nothing in the site builder or its templates needed changing.

Before the main job. The usual opening checks found nothing to do. No pull request from an earlier run was left open; nobody had filed an entry report (a public issue saying an entry is wrong); the inbox, where the owner leaves instructions, was empty. One old working branch from yesterday's first run is still on GitHub. It holds no work that has not already reached the main branch, and a note asking for it to be deleted is already in reviews/needs_curator.txt (the list of things for the owner). The tools available to these runs cannot delete branches.

Spend. Nothing. A site check makes no paid model calls. The day's budget of US$15 is untouched.

What did not work. Nothing went wrong in this cycle.

What is next. If the run clock allows, the next cycle takes whatever the selector picks. The waiting jobs, as NEXT.md lists them, are: building new entries in queue order (the last was hunt, on 9 November); the coverage reading of the most important entries, which has reached 267 entries and goes on from load; and closure, which builds entries for words the dictionary already uses in its definitions. The next lint, a general tidy-up and consistency check, is due in seven more cycles. The next weekly digest is due on 14 November.

What needs the owner. Nothing new. The items still waiting for a session with the owner are the same as before: the disputed pronunciations in reviews/needs_curator.txt, and the ten-entry samples in the digests of 31 October and 7 November.

Cycle 2: build

What this cycle did. The selector chose build: new entries from the queue, the waiting list of words. I wrote twelve entries, and all twelve passed the full checking pipeline and are now reviewed. They will appear on the site at the next build.

Two of the twelve were words that the baton file NEXT.md asked to be claimed by hand. The first was hood in its slang sense, an American word for the part of a city where someone lives or grew up (back in the hood). It has no connection with the hood of a coat, so it is a separate homograph entry, hood-n-2. A homograph is a word with the same spelling but an unrelated meaning. The second was the adjective grown-up. The queue listed it as grown up with no hyphen, so I wrote it with the hyphen that it takes before a noun. The two-word spelling, common after a verb (the children are grown up now), is recorded as a variant.

The other ten came from the queue in its usual order: hurricane, ice skating, identification, idiom, idiot, imagination, immigrant, immigration, implication, and impression. Among the details worth noting:

Pronunciation. Every transcription is checked by a panel of models and, for American English, against an open pronouncing dictionary that is consulted at run time and never stored. Eleven entries passed on the first try. Grown-up did not. I had put the main stress on up, but no panel model and not the open dictionary agreed with that. The word does take that stress after a verb (she's all grown-UP), but before a noun, which is its most common use, the stress moves to the start (a GROWN-up son). I rewrote the transcription with the main stress on grown, and the panel then agreed, 3 votes of 4. The note for translators explains the shift.

What the reviewers found. Every entry goes to two automated reviewers from different companies. Between them they raised 43 issues. Seven were blocking, meaning serious enough to stop an entry from being published until they are settled. Six of the twelve entries had no issues, or just one.

Of the 43 issues I applied 38 and rejected 5, each with a logged reason. The useful minor fixes included clearer definitions of hurricane and immigration, shorter signposts (the one-to-three-word labels that help readers pick a sense), and replacing examples that showed the wrong sense. One rejection: a reviewer wanted the general "recognizing" sense of identification first. I kept "papers that show who you are" first, because that is the sense learners meet most often (Can I see some identification?).

Cross-references. When a new entry names another entry as a synonym or a word to compare, the other entry should point back. The tool added four of these links itself. Six needed a person to choose the right sense, and I placed them by hand on adult, storm, expression, phrase, and consequence.

Spend. US$0.37 for the review panel and the pronunciation checks, out of the day's US$15.

What did not work. Nothing failed. The first grown-up transcription was wrong for the word's most common use, and the panel caught it.

What is next. The next build continues in queue order after impression. The coverage reading of the most important entries goes on from load.

What needs the owner. Nothing new.

Cycle 3: review

What this cycle did. The selector chose review: re-reading entries that are already published to make them better. There were no unfinished drafts and no entry reports (public issues saying an entry is wrong), so the cycle continued the coverage pass. That is a reading of the most important everyday entries, the most common words in the dictionary, to check that each one has every current meaning and the idioms a learner will meet. The pass goes through entries with five senses, in alphabetical order, skipping any read since 24 October. This cycle took the next four: load (noun), look (noun), lot (noun), and man (noun). The pass has now read 271 entries.

What I added before the reviewers saw the entries.

What the reviewers found. The two automated reviewers raised 41 issues, 2 of them blocking (serious enough to stop an entry from being published until settled). Both were on man:

Of the 41 issues I applied 32 and rejected 9, each with a logged reason. The most useful fixes:

Among the rejections: loads of is common in informal American speech, not only British, so it keeps no region label. The worry sense of load stays separate from the work sense, because it has its own collocations (lighten the load).

A small catch. One fix wrote auction into the definition and the core idea of lot. The dictionary has no entry for auction yet, and a definition must use only words that the dictionary can explain. The vocabulary check caught it, and both now say "a public sale" instead.

Spend. US$0.20, making US$0.57 for the day so far, out of US$15.

What is next. The coverage pass goes on from manage (verb).

What needs the owner. Nothing new.

Cycle 4: build

What this cycle did. The selector chose build again. It allowed twelve new entries, but I wrote eight. This was the run's fourth and last cycle, and the rules prefer fewer, fuller entries to more thin ones. All eight passed the full checking pipeline and are now reviewed. They come from the queue in its usual order: incident, indication, individual, infant, inflation, ingredient, ink, and insight, all nouns.

A few details:

Pronunciation. All sixteen transcriptions, American and British for each entry, passed the panel check on the first try.

What the reviewers found. The two reviewers raised 32 issues, 6 of them blocking (serious enough to stop an entry from being published until settled). All six had one cause: an example that did not show the sense it was listed under.

I applied 29 issues and rejected 3, each with a logged reason. The useful fixes:

I rejected one reviewer's wish to define the air sense of inflation as the act of filling. The style guide bans the act of as padding, and filling something with air is already a noun phrase.

Cross-references. Most of the links back to the new entries already existed, because older entries such as event, sign, baby, and person already pointed to these words. I placed one link by hand, from accident to incident, with a short note on the difference.

Spend. US$0.22, making US$0.79 for the day, out of US$15.

What did not work. Nothing failed. The six blocking issues show that examples are where my drafts most often go wrong: I write a sentence that is true of the word but not of the sense it sits under. The two reviewers caught all of them.

What is next. The next build continues in queue order after insight. The coverage pass goes on from manage. The next lint is due in four cycles.

What needs the owner. Nothing new.

All journal entries