Run report, 2026-10-27 (third scheduled run of the day)

Cycle 1: closure

What the cycle did. The scheduler chose closure: writing entries for words the dictionary already uses or points to but has no entry for. The closure list (words used inside definitions) is empty, so the words came from the family list (related forms that other entries name in their word-family boxes). Twelve were claimed. Two of them, run and set as nouns, are large, so I released four lighter adjectives (lost, middle, own, south) back to the queue for a later cycle and wrote eight entries in full:

Every entry went through the full pipeline: format check, the inflection tool, the pronunciation check (all sixteen transcriptions agreed with the pronunciation panel and with the open CMU pronouncing dictionary), the two-reviewer panel, my adjudication, and the vocabulary and cross-reference checks.

What the reviewers found and what I decided. The two reviewers raised 82 issues, 11 of them blocking (must be settled before an entry is published). I applied 19 and rejected 8 with logged reasons.

The vocabulary check refused seven words in definitions that have no entries (cricket, stockings, slope, sled, tennis, volleyball, scenery); each definition was reworded with plainer words.

Spend. US$0.30 for this cycle; the day stands at US$2.57 of the US$15 cap. Reviewer A runs on the owner's Claude subscription and costs nothing in the ledger.

What did not work. Nothing failed. Re-saving the exceptions table with a standard JSON writer reformatted one unrelated block; I restored the file and inserted the two lines by hand so the diff is ten lines.

What is next. The four released adjectives head the family list. A lint pass (the mechanical and judgment check of the whole dictionary, every ten runs) is due next.

What needs the owner. Nothing new. The open items in NEXT.md under "For the owner" stand.

Cycle 2: lint

What the cycle did. The scheduler forced a lint pass, the routine health check of the whole dictionary that runs every ten runs. It has a mechanical half (tools that check caps, links, cross-references, and the queue) and a judgment half (reading the tools' reports and the wiki's own pages for contradictions). It costs almost nothing: the only paid step is the band vote below.

Mechanical checks. Word caps on the rule files and the wiki all pass; the wiki stands at 29,981 of its 30,000 words, so I condensed two old paragraphs of the reviewer-precision note before adding to it. Every relative link resolves. The eight references to "ruled-out" targets (words the owner ruled should not have entries, such as look for as a phrasal verb) are unchanged; they belong to the cross-reference audit already on the list.

Cross-references. When entry A names entry B as a synonym, antonym, or "compare" word, B should name A back. The tool wrote these mirror links in seven entries on its own. For ten more it could not tell which sense of B the link belongs to, so I placed them by hand: for example, wonderful lists amazing as a synonym, and the link back goes on sense 2 of amazing (very good), not sense 1 (surprising); end lists start as an antonym, so start sense 1 (the beginning) now names end.

Reports. The "promises" report (a note that says a word "has its own entry" when no entry or queue row exists) shows the same seven false alarms as last time. The circular-definition report lists 320 definitions that use their own headword; most are harmless (all the people who live in a town), and they are judged one by one in review. The missing-senses report lists 724 pairs for the cross-reference audit.

Band vote. Thirty-three queue rows added by tools since the last vote had no frequency band (a band is the editorial estimate of how common a word is, 1 to 5). The three word-list models voted (US$0.006). Three wide disagreements were settled at the lower middle vote. Two rows were plainly wrong and marked duplicates: relieved queued as a verb (it is a form of relieve) and glad as a noun (it is an adjective, and the right row exists). Both came from the known weakness of the vocabulary checker in guessing a word's part of speech. A third row had a capital letter (Determinant); I lowercased it and wrote a short note about it in the wiki so that a later run can fix the cause.

Reviewer precision. Precision is the share of a reviewer's issues that I end up applying. A family of issues that falls under 30 percent over twenty or more decisions is switched off for that reviewer. Reviewer A (the Claude model) stands at 0.80 over 1,563 decisions, reviewer B at 0.79 over 526. Reviewer B's spelling-and-format family is at 0.28 but stays on by the owner's ruling; reviewer A's inflection family (0.25) and reviewer B's phrase family (0.18) are under the line but have fewer than twenty decisions each. No switch-offs.

Spend. US$0.006; the day stands at about US$2.58 of US$15.

What is next. The next lint is due in ten runs. The queue and the review list in NEXT.md are unchanged apart from the counts.

What needs the owner. Nothing new.

Cycle 3: build

What the cycle did. The scheduler chose build: new entries taken from the queue in the order the owner set. The next rows were twelve band-1 function words, all of them uses that other entries already point to as "a separate entry":

Function words need an explanation more than a definition, so most senses carry a short note on use, for example that as soon as takes a present tense even about the future, or that because of cannot introduce a clause. Each entry has at most two learner errors, chosen from mistakes learners really make (Even though it was late, but we stayed; She is taller then me).

Pronunciation. For five of the multi-word headwords the pronunciation panel (three models that transcribe independently) disagreed with my first transcriptions. I asked the panel for its own answers: all three used weak forms of the small words, as most speakers do (əz ˈɪf for as if, ˈnɛkst tə for next to, ˈi.vən ɪf for even if). That matches house practice for in spite of, so I changed them, and all 24 transcriptions are now verified. One re-check of next to was needed because a panel member's answer varied between calls.

What the reviewers found and what I decided. 48 issues, 8 of them blocking; 18 applied and 7 rejected, all logged.

The cross-reference tool found five pairs where one entry calls a word a synonym and the other calls it only "compare". I matched the new entries to the existing ones, so the audit list does not grow. A queue row for rather than as a preposition, created by my own first draft, was marked a duplicate.

Found along the way. The conjunction because points to an entry because-prep that should be because-of-prep. This is noted in NEXT.md for its next review, not fixed here, because the cross-reference is part of an entry's reviewed content.

Spend. About US$0.33 for this cycle; the day stands at US$2.91 of US$15.

What is next. The five hand-listed homographs (bark, ash, temple, bill, pulse) and the closure rows wait for the next build or closure cycle.

What needs the owner. Nothing new.

Cycle 4: review

What the cycle did. The scheduler chose review: rereading entries that have already been published, both to fix errors and to fill gaps. There were no drafts or reader reports waiting, so I took six entries. Three came from earlier cycles of this run, which had each found something missing. The other three were the next entries in the coverage pass, the owner's 2026-10-24 instruction to reread the most important everyday words, the ones with the most senses first.

What the reviewers found and what I decided. 92 issues, 11 of them blocking; 22 applied and 6 rejected. Both reviewers caught that have listed get as a synonym while also saying, in a note, that get means something different; get is now only under "compare". A definition of pick up used the word notice to define notice. Rejected: one reviewer called not used in continuous tenses "not a pattern", but it is on the closed list. Another wanted the hand marks removed from others in the pronoun's examples. Pronoun entries have no inflections table, though, so without the marks the site would not show others in bold. Most of the minor issues were notes repeating each other (the notes rules of 2026-10-24 say one fact, one field); I removed the copies.

Spend. About US$0.31; the day stands at US$3.21 of US$15.

What is next. The coverage pass continues with away, power, and quarter. The scheduler will pick the next cycle's mode.

What needs the owner. Nothing new this run. The standing items in NEXT.md under "For the owner" are unchanged.

All journal entries