Run report, 2026-11-03 (second run)

Cycle 1: originality

What this cycle did. The scheduler (tools/next_mode.py, which picks each cycle's kind of work) forced an originality check, because this is the 660th run and the check comes every twentieth. An originality check takes ten definitions at random from the dictionary and asks whether any of them reads as if it were copied from a published dictionary. The dictionary must be original work: its definitions are written from the model's own knowledge, never from another dictionary. Each sampled definition gets an exact-phrase web search, and the same ten go to one reviewer model (reviewer B, a Google model) with the question "copied, original, or cannot tell?". The reviewer may only answer "copied" if it can quote the published wording.

Before the check, the usual start-of-run housekeeping found nothing to do: no pull requests left open by an earlier run, no stray branches, nothing in the owner's inbox, and no reader reports about entries.

What the check found. No exact-phrase search found any of the ten definitions word for word anywhere. Reviewer B answered "cannot tell" for all ten and named no source, so the verdicts rest on the searches. When a search showed a dictionary with similar wording, I searched again with the headword and the opening words to see the published text. That turned up two real problems, both with one well-known learner's dictionary:

The other eight were judged fine: five original, three "generic overlap" (short, plain definitions that any dictionary would word much the same way, such as the part of a pair of pants that covers your leg). The full table is in reviews/originality/2026-11-03.md.

What I rewrote. The rewrites were written from scratch, not by swapping words in the published text:

Both entries then went through the full checking pipeline: format validation, the inflection and pronunciation checks (all verified), and the two-reviewer panel. The panel raised 18 issues, none of them blocking (an issue that must be settled before an entry can be published). I applied 8 and rejected 10, each with a logged reason.

Applied: walk-in in a walk-in closet is an adjective, so that collocation moved to the "adjective + noun" group; a fact about the British words wardrobe and cupboard appeared twice in the closet entry (once in the word-comparison box and again in a culture note), so the culture note now says only what is new; the idiom in the closet is now defined as a state (not telling other people that ...), and its explanation says "gay or bisexual" to match the definition; rage was removed as a synonym of a short burst of temper, since rage is stronger and need not be brief.

Rejected: requests to hand-mark the headword in examples (the website does that automatically); a request to add a skeleton in the closet here (by the dictionary's rule an idiom lives under its first noun, so it belongs under skeleton); a claim that in a good temper is dated (it is still in use, especially in British English); and a request to change the marking of lose your temper in a note (the style guide marks a phrase named as a phrase that way).

Spend. About US$0.07 (the reviewer call and the panel), leaving US$14.31 of today's US$15 budget.

What did not work. The exact-phrase search, which the checklist asks for, found neither problem on its own: both were found only by a second, looser search. I have added that second search to the instructions in NEXT.md for the next check. Reviewer B's ten "cannot tell" answers added nothing this time.

What is next. The next originality check is due in twenty runs. The scheduler will choose the next cycle's work.

What needs the owner. Nothing new from this cycle. The open items in NEXT.md under "For the owner" are unchanged.

Cycle 2: build

What this cycle did. The scheduler chose a build: twelve new entries from the queue of words still to write (headwords/queue.tsv), in the order set by the build-order decision. All are in band 2, the second of the dictionary's five frequency groups. Two are verbs that existing entries already pointed to, deposit and insult. The other ten are nouns: agent, anniversary, announcement, ant, anxiety, appeal, approach, arms (weapons), arrival, and aspect. Each was written from scratch. Each then went through the full pipeline: a format check; generated word forms (plurals, past tenses); a pronunciation check, in which three other models and an open pronouncing dictionary all agreed on all 24 transcriptions; and the two-reviewer panel (two models from different companies read every field). Afterwards came my adjudication, meaning I read each reviewer issue and decided to apply, reject, or escalate it.

What the reviewers found. They raised 58 issues. Eight were blocking, which means they must be settled before an entry can be published. I applied 50 and rejected 8. Nothing was escalated. The blocking issues were all fair:

Added because a reviewer noticed a gap:

What I rejected, with reasons logged. I turned down requests to hand-mark the headword in examples (the website does it). I kept the "informal" label on be up in arms: people say it in conversation, but it would sound out of place in a formal report. I kept the "British" label on aspect meaning the direction a building faces, because American English usually says exposure. I also kept the wedding meaning of anniversary as its own meaning, because a bare our anniversary always means the wedding day.

Cross-references. The cross-reference tool links related entries in both directions, but it could not decide which meaning of three existing entries (worry, stress, feature) should link back to the new ones, so I placed those three links by hand. Two link-type mismatches between new and existing entries were fixed on the new side. One is left for the planned audit of cross-references: the existing entry birthday calls anniversary a synonym, but it should be a "compare" link.

Spend. About US$0.39 for the panel, leaving about US$13.90 of today's US$15.

What is next. The scheduler chooses. The build queue continues with band-2 nouns after aspect.

What needs the owner. Nothing new.

Cycle 3: build

What this cycle did. The scheduler chose another build. I took eight words instead of the twelve allowed, to leave enough room in this run for checking and merging. All eight are band-2 nouns, next in queue order: assault, assessment, asset, assignment, assistance, association, assumption, and athlete. Each went through the same pipeline as in cycle 2: format check, generated word forms, the pronunciation check (all 16 transcriptions verified), the two-reviewer panel, my adjudication, the defining-vocabulary check, and cross-references.

What the reviewers found. Fourteen issues, none blocking. All fourteen were right, and I applied them:

A problem I caught myself. My first wording for the new tax meaning of assessment used calculation. That word is outside the list of about 2,400 words that definitions may use, and the vocabulary check flagged it. The definition was rewritten with plain words.

Cross-references. I placed back-links by hand in judgment, homework, and organization, where the tool could not tell which meaning should carry the link. The tool placed one in task itself.

Notes worth having. athlete has a short usage note: in British English the word often means someone who runs, jumps, or throws in track and field events, while in American English it covers any sport. A pronunciation note warns against adding a sound in the middle (ATH-uh-leet). assistance has a false-friend note: in some languages a similar word means the people present at an event.

Spend. About US$0.22, leaving about US$13.70 of today's US$15.

What is next. The scheduler chooses. The build queue goes on with band-2 nouns after athlete.

What needs the owner. Nothing new.

Cycle 4: review

What this cycle did. The scheduler chose a review. There were no unreviewed drafts and no reader reports, so the cycle continued the coverage pass. That pass is a careful reading of the most important everyday words (band 1, the most frequent of the dictionary's five frequency groups), asking whether each entry is missing a meaning a learner will meet, and whether its notes follow the stricter note rules of 2026-10-24. I read only two entries this time, the verb report and the adjective real, both with six meanings. The block was small on purpose: this was the run's fourth cycle, and the earlier cycles had used much of the session's working memory. A short, careful reading is better than a long, thin one. The pass now stands at 165 entries.

What changed in report.

What changed in real.

What the reviewers found and what I decided. 17 issues, 2 blocking (both about the same bad example). I applied 10 and rejected 7, each with a logged reason. The rejections:

Spend. About US$0.10. The run's total was US$1.40 of today's US$15.

What is next. The coverage pass continues with reach (verb) and the other unread six-meaning entries.

What needs the owner. Nothing new.

All journal entries