Run report, 2026-11-11, third run (run 20261111T164342Z-wibk39)

Cycle 1: build

What this cycle did. The selector (the tool that picks each cycle's job from what is most overdue) chose build: new entries taken from the queue, the ordered list of words still waiting for an entry. Its reason was that build had the highest "scheduler debt", meaning it had gone longest without its fair share of runs. It allowed up to twelve entries and US$3 of paid model calls. Twelve words were claimed in queue order and all twelve were written, checked, and published: logic, login, loser, loyalty, luggage, magic, maintenance, make (the noun, as in What make is your car?), making, male (the noun), mall, and management. All are nouns in band 2 (band 2 is the second most common group of words).

Before the work. No open pull requests were left by an earlier run, inbox/ (the folder where you leave instructions) was empty, and there were no "Entry report" issues from readers. The one old work branch, claude/friendly-mayer-mdy32d, still holds nothing that is not already in main. It already has its line in reviews/needs_curator.txt asking you to delete it, because these runs cannot delete branches.

What the entries contain. A few points worth knowing:

What the reviewers found and what I decided. Each entry goes to two reviewer models from different companies, which check every field and mark each problem as blocking (must be settled before publishing) or minor. They raised 37 issues, 7 of them blocking. I accepted 30 and rejected 7, and every decision is logged with a reason.

Other fixes. Two definitions used words that have no entry yet (philosophy in logic, brand in loyalty), so I reworded both. The cross-reference tool (it adds the matching link in the other entry) could not decide which sense to link in four cases, so I placed those links by hand: in password, model, man, and leadership. Logic now points to reason as a "compare" word, matching what the reason entry already said.

Spend. US$0.36 for this cycle (pronunciation panel and reviewer B; reviewer A runs on your subscription at no charge). The day's total is US$1.75 of the US$15 cap.

What did not work. Nothing failed. The making inflection gap has now come up six times. When a run has room for it, the tool fix described in the notes page is worth doing.

What is next. The queue continues after management. The next lint pass (the periodic housekeeping check) is due within a run or two, since nine of its ten runs have passed.

What needs you. Nothing new. The pronunciation disputes and the digest samples still wait for a session with you, and the old branch above can be deleted whenever convenient.

Cycle 2: lint

What this cycle did. After the first cycle merged, the selector chose lint. Lint is the periodic housekeeping pass, forced every ten runs (each cycle counts as a run), and ten had passed. It changes no definitions; it checks that the dictionary, the queue, and the wiki agree with one another and with the rules.

The mechanical checks.

The reports. Three standing reports list things for later review runs to judge, not errors to fix now:

Reviewer precision. Precision is the share of a reviewer's flags that I accepted. Reviewer A, the Claude model, is at 0.85 and reviewer B, the Google model, at 0.82. The rule switches off any issue type a reviewer gets right less than 30 percent of the time, once it has twenty or more decisions. One type is near that line: reviewer B on inflections (word forms such as plurals), at 0.32 over 25 decisions. Its recent misses come from a gap in our own tool, not from the reviewer: the tool cannot record a plural like makings when another sense of the noun is uncountable. So no type was switched off, and fixing the tool would remove the cause. This is written up in wiki/notes/reviewer-precision.md.

The judgment pass. I read the wiki index against the pages: every page is listed. The open-questions page has nothing stale, and reviews/needs_curator.txt (the list of things waiting for you) has no duplicate lines. The log has no gaps.

Spend. Under one cent this cycle. The day's total is US$1.75 of the US$15 cap.

What did not work. The settle step of the band-vote tool looked for a votes file under its default name and failed. It worked once I gave it the file name the run step had written (--votes votes-unbanded.tsv). This is a small inconvenience, not a fault in the data.

What is next. The next lint is due after ten more runs. The build queue continues after management.

What needs you. Nothing new.

Cycle 3: review

What this cycle did. The selector chose review: a second reading of published entries, to find what a learner would expect and cannot find. No draft entries were waiting and no reader reports were open. So the cycle continued the coverage pass, the owner's ruling B1 of 2026-10-24. That pass reads the most important band-1 words one by one, the ones with the most senses first (band 1 is the commonest group of words). The next four unread entries, all with five senses, were program (noun), push (verb), quiet, and ready (adjectives). Two earlier entries in the list, practice and reading, had already been read since 2026-10-24 and were skipped. The pass now stands at 291 entries.

What I added before the reviewers saw the entries.

What the reviewers found and what I decided. The two reviewer models raised 57 issues, 10 of them blocking. I accepted 52 and rejected 5. The useful finds:

What I rejected: push your luck is already in the entry for luck, under the house rule that an idiom lives under its first noun. Make ready is formal and rare today. I also rejected two objections to the new pushing fifty sense. One said the grammar label "not used in the passive" is not on the closed list, but it is. The other said the note "used only in the continuous form" belongs elsewhere, but the style guide puts that kind of restriction in the explanation field.

Spend. US$0.21 this cycle; the day's total is US$1.96 of the US$15 cap.

What did not work. While checking how the decision tool reports several issues on one field, I accidentally logged two real decisions with a placeholder note. Both decisions were correct, and I replaced the two notes with real reasons in the ledger before committing.

What is next. The pass continues with receive (verb), whose last reading was on 2026-10-24 itself, or else record (noun).

What needs you. Nothing new.

Cycle 4: build

What this cycle did. The selector chose build again, allowing up to twelve new entries. This was the fourth and last cycle a run may have, so I took eight, as earlier late cycles have done, to keep every entry as full as the first cycle's. The first two words in the queue were brand and media. Both moved into band 1 (the commonest words) in the band vote during this run's lint cycle, so they came first. The other six came from the queue in order: camp (noun and verb), locate, sue, junior (noun), and lecturer. All eight were written, checked, and published.

What the entries contain.

What the reviewers found and what I decided. The two reviewers raised 28 issues, 5 of them blocking. I accepted 27 and rejected 1.

Three definitions had used words that have no entry yet (tent, vacation, rank), so I reworded them. The cross-reference tool could not choose a sense in two places, so I placed those links by hand: in side (the sense for groups against each other) and find (the sense for finding after looking).

Spend. US$0.22 this cycle. The run as a whole spent about US$0.80, and the day's total is US$2.18 of the US$15 cap.

What did not work. Nothing failed.

What is next. The build queue continues after lecturer. The coverage pass continues with receive or record.

What needs you. Nothing new. As before: the pronunciation disputes and digest samples in reviews/needs_curator.txt and the digests wait for a session with you, and the old branch claude/friendly-mayer-mdy32d can be deleted.

All journal entries