Run report, 2026-10-28 (fourth run of the day)
Cycle 1: review
What this cycle did. The mode selector (the script that picks each cycle's job from what is most overdue) chose review: a second reading of entries that already exist. There were no unreviewed drafts and no reader reports open on GitHub, so the cycle continued the coverage pass. That pass, which you asked for on 2026-10-24, rereads the most important everyday words (band 1, the most frequent of the five frequency levels), starting with those that have the most senses, and asks one question of each: does the entry cover every meaning a learner will actually meet? Four entries were read: position (noun), part (noun), out (adverb), and open (adjective). Seventy-seven entries have now had this reading.
What was added.
- position: a new sense for a person's importance or rank in a society or a group (a position of power, a special position in the family); a military sub-sense (the enemy positions); and the common negative be in no position to do something.
- part: a sub-sense for a numbered section of a book or series (the story continues in part two); do your part; two phrases, the better part of something (the better part of an hour) and the informal in these parts.
- out: a sub-sense for reaching away from the body (hold out your hand), and three new senses: giving or sending to many people (hand out the tests), removing something (cross out the wrong words), and doing something completely (clean out the garage, the concert sold out).
- open: sub-senses for "not covered" (an open fire, an open wound), "spread out" (the book lay open), and "shown without hiding it" (open anger); and two new senses, "available" (Is the job still open?) and "at risk" (open to criticism).
What the reviewers found and what I decided. Each changed entry went to the two automated reviewers (reviewer A, a Claude model; reviewer B, a Google model). They raised 50 points, 2 of them marked blocking, both about the same example in position: She was in the lead position for most of the race, which is not how people talk. It now reads in first position. Of the 50 decisions, 38 applied the reviewer's point and 12 rejected it with a written reason. Applied changes included removing the "formal" label from the job and opinion senses of position (both are ordinary language), tightening the opinion definition to a view stated in public, a clearer definition of an acting part, and removing notes that repeated another field in all four entries (the house rule is one fact, one field). In out, the note about out of before a noun had been said two different ways; it now lives once, in the learner error, which also names out the door and out the window as the normal exceptions. The American label on the "outdoors" sense of out was removed, since out in the garden is used everywhere. Rejected points included requests to add idioms whose keyword is another noun (with open arms and keep an open mind are already under arm and mind), to move jockey for position to a verb entry (its first noun is position), and to mark part in part of the money as uncountable (it is a use without an article, which the usage note already explains).
Three cross-references that a script could not place automatically were placed by hand: open now appears as the opposite of private (sense 1) and as a synonym of available (sense 1) and free (the "not used" sense).
Spend. The reviewers cost US$0.27. The day's total is US$3.80 of the US$15 daily cap.
What did not work. Nothing failed. One small mismatch: the coverage list in NEXT.md had been ordered by number of senses only; this cycle's count of the next candidates also includes sub-senses, so the next four are head (noun), blow (verb), say (verb), and thing (noun).
What is next. More coverage-pass reads, building, or closure as the selector decides. Lint is due in about ten runs.
What needs you. Nothing new. The open items listed under "For the owner" in NEXT.md still stand.
Cycle 2: originality
What this cycle did. The selector forced an originality check, which runs every twentieth run. Its purpose is to make sure our wording has not drifted toward a published dictionary's. A script picked ten fields at random from across the dictionary (mostly definitions, one usage note), and each was searched on the web as an exact phrase. Reviewer B was also asked whether each one reads as copied; it is told to answer "copied" only if it can quote the matching published text.
What was found. Six fields matched nothing and are original. Two overlapped with published wording only in the way any plain definition of the same thing would: the phrase global village and the sense of feel meaning to search with your hands in the dark. Two were close enough that I rewrote them:
- spell (verb), the sense in spell trouble and spell disaster. Ours said be a sign that something, usually something bad, will happen; a learner's dictionary defines a different word, portend, with the same frame. It now reads mean that a particular result, usually a bad one such as trouble or failure, is likely to follow. That also fits the examples better, since spell the end is about a result, not a warning.
- make a practice of something, a phrase in practice (noun). Ours said do something regularly, as a habit, which a learner-dictionary site paraphrases almost word for word. It now reads do something often because you have decided it is a good thing to do.
Reviewer B answered "original" six times and "cannot tell" four times and quoted nothing, so, as in earlier checks, the web searches decided.
What the reviewers found in the rewritten entries. Both entries went through the full two-reviewer panel, which read every field, not only the rewritten ones. There were 19 minor points; 18 were applied and 1 rejected. In spell, the fact that the British past form is spelt was stated twice in different words and now lives only in the inflection note; the second sense ("the letters spell a name") got a clearer definition with its subject named in the explanation, and the summary at the top of the entry now says that the result is usually bad. In practice, habit was moved from "synonyms" to "compare", with a one-line difference, and the matching link in habit was changed too; an example about tipping in restaurants (a custom that differs between countries) was replaced; the doctor-and-lawyer sense now covers the work itself (the practice of medicine), not only the business; the "formal" label on make a practice of was removed; and My French is out of practice, which people do not say, became The team is out of practice after the long break. I rejected the claim that spell in this sense is often used in continuous forms: it is rare, and the "not continuous" code helps learners.
Spend. US$0.09 (reviewer B's answer and the panel). The day's total is US$3.89 of US$15.
What did not work. Nothing failed.
What is next. The next originality check is due in twenty runs.
What needs you. Nothing new.
Cycle 3: build
What this cycle did. The selector chose build: writing new entries from the queue (the list of words waiting for an entry). The next rows were eight adverbs whose other parts of speech already have entries. I wrote six: a little, about, all right, along, as, and long, all as adverbs. Each entry stays with its own uses and points to its neighbors; for example, about the adverb covers about ten dollars and running about, and points to about the preposition for a book about birds.
Two rows were not written. A bit and as well already exist as phrases inside bit (noun) and well (adverb), following the house rule that an idiom lives under its keyword. A separate entry would say the same things twice, so I marked both rows as duplicates, the way of course was handled earlier. Four entries (slightly, also, too, and as well as) still link to the unbuilt slugs; a later review should repoint those links to the phrases.
What is in the new entries. A little: "slightly" (a little tired) and "for a short time or distance" (rest a little). About: not exact (about six), almost (just about finished), and the mainly British "here and there" and "nearby" uses (Is Sam about?), with a note that American English says around there. All right: "well enough" (Did the interview go all right?) and the informal "certainly" (It's him all right). Along: forward, together with someone, arriving (a bus came along), and progress (How's the project coming along?, as you go along). As: the first as in as tall as, and the use that marks a surprising number (as many as fifty people). Long: for a long time (Have you been waiting long?), with how long as a sub-sense; much earlier or later (long ago); and all through a period (all night long); plus the phrases no longer, before long, and the old-fashioned so long.
What the reviewers found and what I decided. The two reviewers raised 43 points, 8 of them blocking. I applied 38 and rejected 5. Most of the blocking points were examples under the wrong sense: A good bike can cost as much as a small car is an ordinary comparison, not a surprising amount (now as little as five dollars); Not long after, it started to rain means soon after, the opposite of the "much later" sense; How long have you lived here? asks about length rather than meaning "for a long time", so it moved to its own sub-sense; and my whole life long is old-fashioned. Reviewer B also asked that all along (I knew it all along) be placed under along; it already lives under all (adverb), so the usage note now points there. Rejected points: calling the progress sense of along gradable (well along has no comparative form), adding out and about (it is under out), and asking for hand marks on longer, which the site marks automatically.
Five links back to the new entries could not be placed by the script and were placed by hand, in a lot, exactly, well, certainly, and on.
Spend. US$0.20 for the reviewers. The day's total is US$4.09 of US$15.
What did not work. Nothing failed. The two duplicate rows show that the tool that queues missing cross-reference targets cannot tell when a target already exists as a phrase; this has happened before (of course) and costs little, so it is noted here rather than in a tooling note.
What is next. The build queue moves on to phr and verb-plus-preposition rows, which need the same duplicate check.
What needs you. Nothing new.
Cycle 4: lint
What this cycle did. Ten runs had passed since the last lint, the housekeeping pass that runs the dictionary-wide checks and fixes what they find, so the selector forced one. This was the ninety-second.
The routine checks. File-size caps, page links, and the vocabulary scan came back clean. The cross-reference tool (which keeps two-way links such as "synonym of" in step) added seven missing return links. Six old claim files, from cycles whose entries are now built, were removed.
Links to entries that will never exist. Fifteen links pointed at headwords that have been ruled out, usually because the word already lives somewhere else. Four of them were new from this run's build cycle, where a bit and as well were found to be phrases inside bit and well. I repointed all fifteen by hand, in twelve entries: the as well and a bit links now go to well and bit with a short note naming the phrase; links to a non-existent adjective other now go to the determiner other; links to ok go to okay; and links to verb-plus-preposition pairs that are not separate entries (look for, care for, think of, look at) go to the verb itself. Four of the return links that the script could not place were also placed by hand (in about, okay, other, and open). The links that now point to a phrase inside another entry were not mirrored back, since a phrase is not a sense.
Frequency bands. Fifty-four words had been added to the queue since the morning's vote without a frequency band (a level from 1, most common, to 5). Three outside models voted on them for under one cent, with no disagreement of more than one band, and all 54 are now ordinary pending rows. One problem recurred: a second vote on the same day overwrites the first day's vote file. I restored the earlier file from the repository's history and kept the new votes in a second file. This has now happened twice, so the tooling note says the fix is due in a later run.
Reviewer precision. Each lint measures how often each reviewer's points are accepted, by kind of point. A kind accepted less than 30 percent of the time over at least twenty decisions is switched off for that reviewer. Reviewer A's points about inflections (word forms such as longer) were accepted 5 times out of 24 (21 percent), so they are now switched off. Nothing is lost: the forms come from rules and an exceptions table, not from the reviewers. Reviewer B's spelling-and-format points are also under the line (29 percent) but stay on by your ruling of 2026-10-05.
Reports for later reviews. 328 definitions use their own headword (most are harmless, such as water in "give an animal water to drink"); 6 notes promise entries that do not exist; and 758 pairs of linked senses could use a hand check. These are listed by the tools for later review cycles.
Housekeeping on the wiki. My update to the reviewer-precision page pushed the wiki four words over its 30,000-word cap, so I condensed five older measurement sections into one, and corrected that page's line in the index, which still listed switch-offs reset by your 2026-10-25 rulings.
Spend. Under one cent. The day's total is US$4.09 of US$15.
What did not work. The band-vote overwrite above. Also, I gave the metrics script a count of 31 changed entries; the true count is 24. The script can be run only once per cycle, so the error is recorded in the log instead.
What is next. Review, build, or closure as the selector decides; the next lint is due in ten runs.
What needs you. Nothing new.