Scheduled run, 2026-10-24 (run 20261024T084338Z-jiwbqq)

Cycle 1: build

What this cycle did. The mode selector (tools/next_mode.py, the script that picks each cycle's job) chose build: write new entries from the queue of words still to do. Under your ruling A2 the queue now gives milestone 1 first (an entry for every word the definitions are written in) and, within it, the grammar terms first. So this cycle wrote twelve entries for words the dictionary uses to talk about language: noun, pronoun, consonant, plural (both the noun, the plural of "mouse", and the adjective, a plural noun), preposition, syllable, tense, vowel, passive, singular (adjective), and infinitive. Each went through the full pipeline (the fixed sequence of checks every entry passes: format check, generated word forms, pronunciation check, two outside reviewers, my decisions on their objections, a vocabulary check, and cross-reference links) and is now marked reviewed, so it appears on the site.

Before the build, the cycle also wrote the first weekly digest (journal/digest-2026-10-24.md), which your ruling E3 asked for: the week's numbers, ten entries for you to sample, and what waits on you.

Choices in the entries worth knowing.

What the reviewers found and what I decided. The two reviewers (an OpenAI model and a Google model, called reviewer-a and reviewer-b) raised 44 issues in all, 18 of them marked blocking. I applied 19 and rejected 25; none needed you.

Reviewer-b again added little: of its few objections, almost all were about punctuation and were rejected. That is the pattern your ruling E1 addresses; the trial of another model waits on two lint passes, as ruled.

Pronunciation. All 24 transcriptions (American and British for each entry) were verified by agreement among the pronunciation models and, where it has the word, the open CMU pronouncing data consulted at run time.

Spend. US$0.63 for this cycle; US$4.88 of the US$15 daily cap is spent so far today.

What did not work. Nothing failed. The cross-reference tool reported that active lists passive as a "compare" link while passive lists active as an antonym; both are defensible, and the mismatch is only a warning. It is left for the cross-reference audit (ruling C1).

What is next. Milestone 1 has 270 rows left. Further cycles today will take whatever the selector picks, most likely more building or the review work your rulings queued (the appendix errors first).

What needs you. Nothing new from this cycle. The digest lists the standing items: two open questions and five pronunciation disputes. One more line in reviews/needs_curator.txt says the branch claude/new-session-c4lr78 is fully merged and could be deleted.

Cycle 2: review

What this cycle did. The selector chose review: re-reading existing entries. With no drafts waiting, the first review work under your ruling E2 is the list of errors in the appendix of your rulings of 2026-10-24 (ruling B3). This cycle took the first two groups of that list: the four misplaced back-links and the "wrong or doubtful notes" (all except the numbers two to ten, which are next). It also did three items from the third group (carefully, entertainment), plus deliberately, because removing a cross-reference means removing its mirror too. A back-link is a cross-reference the tools add on the other entry so that the link runs both ways. Nineteen entries went through the full pipeline again, and one more (distinguish) received a back-link placed by hand.

What changed, item by item.

What the reviewers found and what I decided. The two reviewers raised 134 issues, 101 of them blocking (an issue that must be decided before the entry can count as reviewed). I applied 66 and rejected 53; none went to you. Most rejections were requests to add phrasal verbs (run out of, tell off) or compounds (flat-out, house arrest) as senses, which this dictionary keeps as separate headwords, or to put an idiom under the wrong keyword. Useful finds that I applied: can't seem to ("be unable to do something, although you try") now has its own subsense under seem; get gained an informal sense, annoy (what really gets me is...); life gained the phrases in real life and life after death; physical now covers a physical relationship; and several entries lost notes that only repeated another field. I rejected a reviewer's claim that the card suit normally takes a singular verb, because your appendix questions Hearts is trump.

Spend. US$2.62 for this cycle, almost all of it the reviewers reading these long entries; US$7.50 of US$15 is spent today.

What did not work. One fix tried a collocation type that the closed list does not have (verb + adjective); the validator caught it, and the item now uses the type phrase. The logged decision for it names the type that was refused.

What is next. The rest of the appendix, in order: the numbers two to ten, the remaining examples and collocations (deal, like, just, clean up, go on, -y, strict and the surname cases), definitions and codes (proud, light, dull), the five invented learner errors, and the missing senses, which open the coverage pass.

What needs you. Nothing new.

Cycle 3: build

What this cycle did. The selector chose build again. The queue gave the last three grammar terms (participle, and the adjectives countable and uncountable) and then the next milestone-1 words in band order: receive, bar, exam, newspaper, series, lovely, responsible, and eighteen. That makes eleven new entries, all through the full pipeline and now reviewed. Milestone 1, an entry for every word the definitions are written in, now has 258 rows to go, and the grammar terms are done.

One claimed word was not built. The queue held of course as an adverb, but the dictionary already has of course as a phrase under course. By the house rule a fixed phrase lives under its keyword, the first noun, so a separate entry would duplicate it. I marked the queue row as a duplicate. Three entries (certainly, naturally, obviously) still point at the adverb entry that will now never exist; NEXT.md asks a review to point them at the phrase instead.

Choices worth knowing.

What the reviewers found and what I decided. The two reviewers raised 38 issues, 17 of them blocking. I applied 18 and rejected 20; none went to you.

Four back-links could not be placed by the tool, because no sense of the other entry clearly matched. I placed them by hand: receive under accept "take what's offered", series under show "TV or radio", lovely under kind "caring", and responsible under sensible "good judgment".

Spend. US$0.82 for this cycle; US$8.32 of the US$15 daily cap is spent so far today.

What did not work. Nothing failed. Four first-draft definitions used words outside the defining vocabulary (chocolate, collection, reaction, react); the vocabulary check caught them and they were reworded.

What is next. More of milestone 1, or the rest of your appendix, as the selector decides.

What needs you. Nothing new.

Cycle 4: review

What this cycle did. The selector chose review again, and the cycle took the next part of your appendix of errors (ruling B3). That covered the numbers two to ten, the group of wrong examples, collocations, and synonyms (deal, like, just, clean up, go on, the suffix -y, strict), and the cases where a name from the house list of example names (Kim, Lee, Sam...) is used as a family name. It also cleaned up the three entries that pointed at the duplicate of course adverb found in cycle 3. Twenty entries went through the full pipeline, and one more (ceiling) received a hand-placed link.

What changed.

The surname cases. I found seven entries where a list name follows a title. Three can do without the title and were fixed (strict and floor; manner is next). The other four (teacher, title, Dr., doctor) are about titles, so their examples need a family name after Mr. or Dr., and the names policy says nothing about that case. I left them as they are and added one line to the curator queue (reviews/needs_curator.txt). The working assumption is that a title entry may pair a title with a list name until you rule.

What the reviewers found and what I decided. 115 issues, 86 blocking; I applied 30 and rejected 57, with none escalated. Reviewer-a repeatedly wanted the number entries split into noun and pronoun entries (a six, three of us). I rejected that, because the house's existing number entries (hundred, zero) already hold such uses, and I added a fence line to NEXT.md so later runs need not argue it again. Reviewer-b again objected to a standard verb pattern on clean up six times. Useful finds that I applied: the duplicate grammar notes on deal, the doubling rule for -y (it does not apply to w, x, y: snowy), the secret sense of strict, and the floor-price sense of floor.

Spend. US$1.66 for this cycle. The day's total is about US$9.98 of the US$15 cap.

What did not work. Nothing failed. One decision needed the tool's --index option, because two issues shared a field.

What is next. The rest of the appendix: manner, the "definitions and codes" group (proud, light, dull), the five invented learner errors, and the missing senses, which open the coverage pass.

What needs you. One new question, about family names after titles in the four title entries, with the working assumption above. The earlier open items are unchanged.

All journal entries