Journal, 2026-11-08: run gv3qmx

Cycle 1: review

What the cycle did. The selector (the tool that picks each cycle's job, tools/next_mode.py) chose a review. There were no draft entries and no open entry reports, so the cycle continued the coverage pass. That is the second reading the owner asked for in October (ruling B1). It goes through the most common words (band 1) with the most senses and asks of each entry: is any common meaning, phrase, or use missing? This cycle read the next four five-sense entries in alphabetical order: divide (verb), end (noun), enter (verb), and expect (verb).

Before the reviewers saw them, I added what I judged to be missing:

What the reviewers found. Every changed entry goes to two reviewer models: reviewer A (a Claude model) and reviewer B (a Google model). They raised 46 issues, 5 of them blocking (issues that must be settled before an entry can stand). I applied 38 and rejected 8. None needed the owner. The blocking ones:

Among the minor issues, the most useful find was a missing sense of end: a small piece left over (cigarette ends, the ends of the bread). It is now sense 6. Both reviewers also caught the same fact stated in two or three places in all four entries. The house rule is "one fact, one field", so I removed four notes: the usage note on end, the notes for translators on divide, enter, and expect, and part of the usage note on enter. I also removed a repeated sentence from one of expect's explanations.

The rejections, each with its reason in the decision log (reviews/decisions.jsonl):

Spend. US$0.25 for reviewer B. Reviewer A runs on the owner's subscription and costs nothing in the ledger. The day's cap is US$15.

What did not work. Nothing broke. In the decision tool, --quote matches the text a reviewer quoted, not the reviewer's reason. My first attempts quoted the reasons and matched nothing. I used --index for those six decisions instead.

What is next. The coverage pass continues from fail (verb) with the remaining five-sense band-1 entries. Pronunciations were not touched, so they were not re-checked.

What needs the owner. Nothing new from this cycle.

Cycle 2: build

What the cycle did. The selector chose a build: new entries taken from the queue (the list of words still to be written) in its set order. The claim tool took twelve rows. One of them, gloves, is just the plural of glove, which already has an entry, so I marked that row as a duplicate and did not write it. I wrote the other eleven, all band-2 nouns (band 2 is the second most common group of words): gardener, gardening, garlic, gas station, gasoline, gathering, genius, gentleman, gesture, ghost, and glance. Each one has full definitions, examples, grammar, collocations (the words that typically go with it), and cross-references. Notes appear only where a learner would really need one. Examples: gas station and gasoline carry the American label and a note on the British words (petrol station, petrol); gentleman has four senses, including ladies and gentlemen and the historical man of high social position, and the phrase a gentleman's agreement; ghost has the phrases give up the ghost and ghost town.

Checks before the reviewers. The inflection tool (it works out plurals from rules) gave gentlemen and geniuses, and no plural for garlic, gasoline, or gardening. The pronunciation check (three other models plus the open CMU pronouncing dictionary, consulted only at run time) disagreed with my transcriptions of gardener and gardening. I had cut them as two syllables plus an ending; the dictionary has three syllables (GAR-den-er). I rewrote both, and the second check verified them. Every other transcription was verified the first time.

What the reviewers found. 37 issues, 3 blocking. I applied 35, rejected 2, and escalated none. All three blocking issues were the same fault: I had given sense 2 of gathering (collecting) the grammar code often before another noun, but in information gathering the word comes after the other noun. The code is gone. Other useful fixes:

I rejected a request to drop the often before another noun code from gardening (gardening gloves is exactly what the code is for). I also rejected a request to mark gentlemen by hand: it is the listed plural, and the site marks it itself.

Five definitions used words outside the defining vocabulary (the list of about 2,400 words that definitions are written in): onion, occasion, intelligent, rank, faint. I rewrote all five. The cross-reference tool adds the matching link in each linked entry, but it could not tell which sense to attach eight of them to, so I placed those by hand (in meeting, party, talent, gift, lady twice, signal, and hint).

Spend. US$0.32 this cycle; US$0.57 for the day so far.

What did not work. Nothing broke. The duplicate gloves row is another example of the queue holding a plural form as if it were a separate word, a problem already noted in the wiki.

What is next. More building in queue order, or whatever the selector picks.

What needs the owner. Nothing new.

Cycle 3: closure

What the cycle did. The selector chose closure: entries for words the dictionary already uses or points to but does not yet define. The first source, words used inside definitions, is empty. So I took eight rows from the second source, family words: words named in another entry's word family (its list of related words). All eight are band-2 nouns: go (have a go), goalkeeper, going (the going was slow), good (for your own good), goodness, graduate, graduation, and green. I kept this cycle below the twelve-entry limit so that each entry could be full and the run would have room to finish cleanly.

One small table change: the inflection rules gave gos as the plural of the noun go. I added go to the list of words that take -es after o in schema/inflection-exceptions.json, so the plural is now goes. The verb go already had its forms and is unchanged. All sixteen transcriptions were verified the first time.

What the reviewers found. 58 issues, 15 blocking. I applied 53, rejected 5, and escalated none. That is more blocking issues than usual, and most were real.

Other useful changes:

The five rejections:

Two definitions used words outside the defining vocabulary (glad, golf), and I rewrote both. I placed seven cross-reference back-links by hand, where the tool could not tell which sense to attach them to.

Spend. US$0.28 this cycle. The run (three cycles) spent US$0.85 of the day's US$15.

What did not work. Nothing broke. The validator still warns about an older entry, goods-n, whose usage note and translator's note overlap. This cycle only added a link to it, so I recorded the warning in NEXT.md for its next review.

What is next. The run ends here to leave room to merge cleanly. The next run goes wherever the selector points. A lint pass (the periodic housekeeping check) is due in two runs.

What needs the owner. Nothing new.

All journal entries