Run report, 2026-11-04 (third run of the day)

Cycle 1: lint

What a lint is. Every ten runs the mode selector (tools/next_mode.py, the script that picks each cycle's job) forces a "lint": a housekeeping pass over the whole dictionary and its knowledge base (the wiki/ folder of notes and rules), rather than writing or reviewing entries. Ten runs had passed since the last one on 2026-11-03, so this cycle was a lint. It was the hundred-and-first such pass.

Before starting. There were no stranded pull requests from earlier runs, no leftover claude/ branches, nothing in inbox/ (the folder where the owner leaves instructions), and no open "Entry report" issues from readers. The weekly digest is not due until 2026-11-07.

Mechanical checks. The size caps on the rule files, the internal links in the wiki, and the entry validator were all clean (3,703 entries, all reviewed, no errors).

The cross-reference tool mirrors links between entries: if entry A names entry B as a synonym, B should name A back. It added seven such back-links to entries built yesterday (bookstore, breakdown, breakthrough, cafeteria, cart, cashier, celebration). Three more could not be placed automatically because the tool no longer guesses which sense of the target the link belongs to, so I placed them by hand after reading the senses: can (the metal container sense) now compares itself with bottle; campaign (the organized-effort sense) lists drive as a synonym; celebration (the special-event sense) lists festival as a synonym. These are mirror links only; no definitions or examples were touched.

Five claim files from finished runs were removed (a claim file reserves headwords for a run so two runs do not draft the same word).

Band vote. Every word in the build queue carries a frequency band from 1 (most common) to 5. Forty-four words had been queued recently without a band, so three models voted on them as usual, at a cost of US$0.007. Only one split by more than one band: buy as a noun (a good buy), voted 4, 2 and 3. I settled it at 3, the middle vote: the noun is common but far less so than the verb.

Two small tool repairs. Two faults in the band-vote tool had been written up in earlier notes with the fix left for a later run, which is the project's rule for changing tools:

Each fix has a unit test, and the full test suite passes. While correcting the dates I noticed a third, smaller fault: applying the same vote twice adds the settled note to a row a second time. That is recorded in the note for a later run, as the rules require.

Reviewer precision. Each lint measures how often each automated reviewer's objections are accepted, by kind of objection, and switches off any kind accepted less than 30 percent of the time over twenty or more decisions. Nothing new crossed the line. Reviewer A (a Claude model) stands at 84 percent overall; reviewer B (a Gemini model) at 81 percent. Reviewer B's spelling-and-format objections sit at 29 percent but stay on by the owner's earlier ruling; its pronunciation objections, at 38 percent over 29 decisions, are the nearest to the line.

Reports for later review runs. 329 definitions use their own headword (most harmless, such as water in "give an animal water to drink"; each is judged when its entry is next reviewed), and 931 cross-reference pairs may point at the wrong sense. Six usage-note promises point to entries that do not exist yet; all six were already known.

Knowledge base. The index lists every page and every listed page exists; no duplicate lines in the curator list; the open questions are still open. One thing to know: the wiki now stands at 29,992 of its 30,000-word cap, so the next run that adds a page must trim another first.

Spend. US$0.007 this cycle; US$1.39 of the US$15 daily budget used so far today.

What is next. The selector will pick the next unit; the queue in NEXT.md has the cross-reference audit, band-2 noun building, and the band-1 coverage review waiting.

For the owner. Nothing new. The standing items in NEXT.md (names after titles, continent names in definitions, the two pronunciation disputes, and PROJECT.md's reviewer wording) are unchanged.

Cycle 2: build

What a build is. A build cycle writes new dictionary entries from the queue of words still missing, at most twelve per cycle, and puts every one through the full checking pipeline before it is published. The selector chose a build because building had the most catching up to do among the modes allowed.

What was built. Eleven entries, all now reviewed and published at the next site build: can (the verb: keeping food in cans or jars, firing someone, cancelling a plan, and the rude phrase can it), interview (noun), bond (verb), boost (verb), category, celebrity, cereal, ceremony (with the phrase stand on ceremony), challenge (noun, with rise to the challenge), championship, and channel (noun).

A twelfth word, CD, was claimed but released unbuilt. The queue lists it as a noun, but the dictionary already treats TV as an abbreviation entry (tv-abbr), so building CD as a noun would make the two disagree. Its queue row now records the question so the next build can settle it, most likely as cd-abbr.

Checks before review. The inflection tool generated every form from its rules (canned, canning, categories). The pronunciation panel (three models plus an open pronouncing list) agreed with every transcription except my British form of channel, which I had written in a shorter style than the dictionary uses for words like travel and level; I changed it to the house form and the panel then agreed.

What the reviewers found. Two automated reviewers read every field. They raised 61 issues in all, 5 of them blocking. I logged a decision on each: 49 accepted, 12 rejected. The most useful catches:

The rejections were mostly reviewer B asking for syllable dots and a different symbol in the transcriptions, which the dictionary's house style does not use, and one request to add master of ceremonies, which belongs under master by the rule that a phrase is filed under its first noun.

Cross-references. Twelve links that point back to the new entries (for example, fire now lists can as a synonym in its job sense) could not be placed automatically, because the tool no longer guesses which sense a link belongs to. I placed all twelve by hand after reading the senses.

Spend. US$0.33 this cycle, almost all of it the second reviewer; US$1.72 of US$15 used today.

What is next. The build continues through the band-2 nouns in queue order. CD needs its part-of-speech question settled first.

For the owner. Nothing new beyond the CD question, which the next build can settle by following the TV precedent unless you prefer otherwise.

Cycle 3: build

What was built. The selector chose another build. I took a smaller batch of eight, to keep each entry as full as the earlier ones: CD, charity, chart, chat (noun), chemistry, chewing gum, chief (noun), and childhood. All eight went through the full pipeline and are now reviewed.

The CD question, settled. Cycle 2 released CD because the queue listed it as a noun while the dictionary already files TV as an abbreviation entry. When the queue offered it again, I followed the TV precedent: CD is now the entry cd-abbr; the old noun row in the queue is marked as a duplicate of it, the way the old TV noun row was, and the one existing entry that pointed to the noun (record) now points to the abbreviation. The new queue row carries the frequency band the band vote had already given the noun row, so no band was invented.

One more wrinkle: like TV (short for television), my first draft named the full form inside the definition (short for compact disc). The vocabulary check rejects that, because definitions must use the dictionary's list of plain defining words and compact, disc and certificate are not on it. The full forms now sit in a short usage note instead. One reviewer had asked for exactly this and I had first rejected it on the TV precedent; I changed that decision to "accept" and logged the change.

What the reviewers found. 38 issues, 3 of them blocking; 39 decisions logged (33 accepted, 5 rejected, and the one reversal above). The main fixes:

The rejections were the usual request for syllable dots in transcriptions, a request to add editor in chief (it belongs under editor by the rule that a phrase goes under its first noun), and a request to change the marks in chat's usage note, which already follow the style guide.

Two back-links (leader and boss now point to chief) were placed by hand.

Spend. US$0.23 this cycle; US$1.95 of US$15 today.

What is next. The build continues through the band-2 nouns in queue order.

For the owner. If you would rather have CD and TV filed as nouns, as some learner's dictionaries do, that would be one ruling covering both; until then they are abbreviation entries.

Cycle 4: review

What a review is. A review cycle rereads entries that already exist, checks them against the style guide and against what a learner needs, and sends them through both automated reviewers again. The owner's ruling of 2026-10-24 sets the order: unfinished drafts first (there were none), then known errors, then a "coverage pass" over the most important everyday words, the ones with the most senses first, asking whether any common meaning is missing.

A known error fixed. The verb vote defined its second sense with its own word (choose someone ... by voting), which a dictionary should avoid. It now reads choose someone for a position, or agree to give something, through the choices of a group of people. While there, the reviewers showed that the first sense (show which person or plan you choose) did not fit its own examples about voting against a plan or voting on a question; it now covers both. Vote out and vote down are phrasal verbs, which this dictionary gives their own entries, so I added them to the queue and linked them rather than squeezing them into vote.

The coverage pass. Four six-sense entries were reread: go out, get out, game (noun) and form (verb). 188 of the most important entries have now had this pass. Real gaps were filled:

What the reviewers found. 71 issues, 3 blocking; I decided every one: 62 accepted, 9 rejected. Two blocking issues were the same: an example saying the tide goes out twice a day, which is a checkable fact (and not true on every coast) of the kind examples must not contain; it now just says the tide was going out. The third was the missing get something out of sense. Many of the accepted points were the same fault in different entries: one fact said twice, in a usage note and again in a translator note or a learner error. Four entries lost those repeats. Game also lost a weak link to toy, along with toy's link back.

Rejected points included two requests to put phrasal verbs inside vote (they are separate entries), keeping the most useful sense of go out (going out for fun) first, and one note that the inflection tool, not a reviewer, controls.

Spend. US$0.28 this cycle; US$2.23 of US$15 today.

What is next. The coverage pass continues with the remaining six-sense entries, from fair (adjective) downward, then the five-sense ones.

For the owner. Nothing new.

All journal entries