Run report, 2 November 2026
Cycle 1: originality
What this cycle was. Every twentieth scheduled run is an originality check: a spot test that the dictionary's wording is its own and not copied, even unconsciously, from a published dictionary. The selector (tools/next_mode.py, the script that picks each cycle's job) forced it this time because this was run 640. Before starting, the usual pre-flight checks found nothing waiting: no unmerged pull requests from earlier runs, no leftover branches, nothing in the owner's inbox, and no public "entry report" issues.
How the check works. A script picks ten fields at random from the 3,533 reviewed entries, mostly definitions. For each one I ran a web search for the exact wording in quotation marks, looking only for whether a published dictionary uses the same words. Then I put the ten to one of the two AI reviewers with a fixed question: does this read as copied, and if so, quote the source. Each field then gets one of three verdicts: original; generic overlap (a short, plain definition that any dictionary would write in much the same way, which is not copying); or rewrite (it matches a published dictionary's distinctive wording, so it must be reworded).
What was found. No exact-phrase search found a match. The reviewer, reviewer B this time (the previous checks asked reviewer A, and the last handover note asked for B), called nine fields original and one "cannot tell". But one of its reasons mattered more than its verdict. For character-n sense 1, our definition was all the qualities that make one person different from other people. The reviewer called it original but said it was "a slight rewording" of the Oxford learner's dictionary, and quoted that dictionary's wording. A second search confirmed it: Oxford defines the sense as all the qualities and features that make a person ... different from others. Ours followed the same frame, with only two words changed, so I judged it a rewrite. Sense 2, the particular qualities that make a place or thing different from others, was built on the same frame, so I rewrote it too.
Final tally: 6 original, 3 generic overlap (one-another-pron, work-out-phrv, floor-n), 1 rewrite (character-n). The three overlaps share short ordinary phrases with learner's dictionaries (the surface that you walk on for floor, each … the other or others for one another) in differently built sentences. A reciprocal pronoun or a floor can hardly be defined any other way.
The rewrite. Sense 1 now reads the whole set of a person's qualities, good and bad, that makes them the kind of person they are; sense 2 reads the general nature of a place or thing, made up of the features that people notice most about it. My first try at sense 1 (the kind of person someone is, shown in the way they usually think, feel, and behave) was dropped on rereading. It collided with the entry's own note telling character apart from personality, which describes personality as "a person's usual way of feeling and behaving". The final wording keeps the distinction the note draws: character is the whole set of qualities, including moral ones.
The rewritten entry went through the full pipeline: format and schema checks, the inflection and pronunciation checks (both verified), and both reviewers reading every field. They raised 9 issues, none blocking. I applied 3: sense 4, the character in a story, now mentions television shows, as its examples already did; the phrase out of character lost an explanation that only pointed to in character; and a sentence in the synonym note with two colons was smoothed. I rejected 6, each with a logged reason. Among them: a request to drop the false-friend note because the learner-error note "says the same thing" (the two are different traps, one about translation and one about usage); a request to mark bad temper with single asterisks (the style guide says a phrase named as a word takes the double mark); and a proposal to reorder senses that earlier reviews had already settled.
Spend. US$0.06 (the reviewer question and the panel), out of a US$15 daily budget.
What did not work. Nothing broke. One lesson is recorded in NEXT.md: reviewer B's verdict alone would have missed the one real problem. Its quoted comparison caught it, so later checks should search any quotation the reviewer gives, whatever its verdict.
What is next. The selector picks the next cycle's job. The standing queue still leads with the cross-reference audit, closure building (adding words the dictionary's own definitions use), and the coverage reading of band-1 entries.
For the owner. Nothing new. The standing items in NEXT.md remain: the curator line on names after titles, the charter's "two non-Anthropic reviewers" sentence, and the open pronunciation disputes.
Cycle 2: build
What this cycle was. The selector chose build, writing new entries from the queue (the ordered list of words still to be written). It allowed up to twelve. I took six band-2 verbs (band 2 is the second-most-important frequency group) and added a seventh, explained below: review-v, sweat-v, tackle-v, tag-v, tap-v, tap-v-2, and tip-v. All seven are now published as reviewed.
One decision worth knowing about. Tap as a verb is really two unrelated English words that happen to share a spelling. One is the light hit (tap someone on the shoulder, tap the screen). The other comes from the water tap and means drawing something out of a supply (tap a market, tap a phone, tap a tree). House rules make unrelated words separate entries, so the second became tap-v-2, a homograph (same spelling, separate entry). The existing noun tap-n (the water device) had pointed its word-family link at the light-hit verb; it now points to tap-v-2. The American use she was tapped to lead the team (chosen) went with the light-hit word, since it comes from a tap on the shoulder. The light-hit noun (a tap on the door) has no entry yet; the next build should queue it.
What each entry covers. Review: reconsidering a decision; writing a review of a book or film; the American study again before a test (British English says revise, and the two entries now point to each other); checking someone's work. Sweat: the body sense; working hard; worrying while you wait (let someone sweat, don't sweat the details); and the cooking sense (sweat the onions). It also has two phrases, sweat it out and don't sweat it, and a new line in the inflection table, because Americans often say sweat rather than sweated in the past. Tackle: dealing with a problem; the sports sense; physically stopping someone; and the British tackle someone about something. Tag: labels, including electronic tags on animals or people; tagging people online; describing someone as something; the chasing game and baseball; and graffiti tags. Tip: leaning or pouring; giving a tip; and the British tipped to win.
What the reviewers found and what I decided. Each entry was read by two independent AI reviewers, field by field. Because my fixes added new material (a cooking sense, two phrases, a graffiti sense, a fourth sense of tap-v-2), I sent three entries round a second time, so no new text went out unread. Across both rounds I logged 66 decisions: 49 fixes applied and 17 objections rejected, each with a reason. Four issues were blocking (they must be settled before an entry can be published), and I agreed with all four. Don't sweat it had been squeezed into an explanation field that ordinary verbs should not use, so it became a phrase. Sweat it out was missing. And tag had an example, tagged the plan a waste of money, that is not natural English; it became tagged her as the favorite. Smaller fixes included a doctor example that did not fit its sense, collocation types (house rule 16: be well reviewed is a phrase, not adverb plus verb), and an example sentence that was too British for an American-based dictionary.
The rejections were mostly settled rules. Idioms go under their first noun, so tip the balance belongs to balance-n and sweat blood to blood-n (both are noted in NEXT.md for when those entries are next opened). The pattern list has no "intransitive" pattern, because the transitivity field already records that. Sweater stays in the word family of sweat, because it is derived from it.
Cross-references. The cross-reference tool links related entries in both directions, but it stops when it cannot tell which sense of the other entry a link belongs to. I placed two such links by hand (assess-v and deal-with-phrv, each on its first sense). I dropped two others (use-v, lean-v) as synonyms because they were weak matches.
Spend. US$0.38 for this cycle; US$0.44 for the day so far, out of a US$15 daily budget.
What did not work. One of my editing scripts got stuck in an endless loop while changing tap-n, and I had to stop it; the edit was redone correctly. Nothing reached the repository in the broken state.
What is next and what needs the owner. The next build continues the band-2 verbs in queue order. Nothing new needs the owner.
Cycle 3: review
What this cycle was. The selector chose review: rereading entries that are already published, to find what is missing or wrong. With no drafts waiting and no reader reports open, the work was the coverage pass. That is a planned rereading of the most important (band-1) entries, starting with those that have the most senses, checking whether every current use is covered. This cycle read the next four six-sense entries: use-v, unit-n, turn-n, and true-adj. That brings the pass to 143 entries, with 59 six-sense entries still to go.
What was added.
use-v: a subsense for using a period of time or a chance (she used the break to call her mother).unit-n: a business sense, a single item of a product (sold a million units, unit price).turn-n: two phrases, the British on the turn (starting to change, often for the worse: the milk is on the turn) and the saying one good turn deserves another.true-adj: the most changes. A reviewer marked one gap as blocking (it must be fixed before publishing): the use the same is true of… and this holds true in most cases had no sense, and now has one. A second new sense covers True, but… and That's true, said when you agree with a point before adding an objection. Sense 2 now also covers a true professional: someone with all the qualities their kind should have.
What was fixed.
unit-n: an example stated a checkable scientific fact (the kilogram as a unit of weight), which house rules forbid in examples. Another example and a collocation used the British kitchen units; American English says cabinets.turn-n: listed turn at something, which is not a real collocation. It also gave miss a turn (a games phrase) under the road sense, where it should read miss the turn. Two learner-error notes used the wrong kind of mark.true-adj: sense 1 was coded not gradable although very true and truer are ordinary English.- All four: duplicated notes were removed (the same fact stated in two fields). On unit-n and lesson-n, the link between them moved from "synonym" to "compare" on both sides, because a lesson is a part of a unit, not another word for it.
What the reviewers found and what I decided. Two AI reviewers read every field. true-adj went round twice, because its new sense was written after the first reading. There were three blocking issues: the fact example, turn at something, and the missing true of sense. I agreed with all three. Of 67 logged decisions, I applied 42 and rejected 25, each with a reason. Most rejections follow the keyword rule: an idiom goes under its first noun, or its first verb when it has no noun. So too good to be true belongs under be, ring true under ring, and true to form under form; these are listed in NEXT.md for when those entries are next opened. I also kept two of reviewer A's targets as synonyms of use-v (exercise and operate), because each fits part of the first sense and the other entries link back.
Spend. US$0.26 for this cycle; US$0.70 for the day.
What did not work. Nothing failed.
What is next and what needs the owner. The coverage pass continues with the remaining six-sense band-1 entries. Nothing new needs the owner.
Cycle 4: build
What this cycle was. The selector chose build again. I took four more band-2 verbs from the queue, fewer than the twelve allowed, so each could be done fully in the time left: toss-v, tour-v, track-v, and trap-v. All four are now published as reviewed.
What each entry covers.
- Toss: throwing lightly (toss me that towel); tossing a coin to decide something, with a note that American English usually says flip a coin; the movement of waves or a horse's head; the cooking sense (toss the salad); and the informal American throw away (those old shoes should be tossed). It also holds the phrase toss and turn, which goes under toss because it has no noun and toss is its first verb.
- Tour: travelling around a place to see it, and travelling to perform or play. A learner-error line shows that tour takes the place directly (tour the museum, not tour in the museum).
- Track: following someone or something by marks or equipment; following changes over time (track your spending), with a subsense for websites and apps tracking what you do online; and the American track mud into the house.
- Trap: keeping someone in a place they cannot leave (trapped in the elevator) comes first, because it is the most common use. It has a subsense for being stuck in a bad job or relationship. Then come catching animals, tricking someone into saying something, and the British trap your finger in the door.
What the reviewers found and what I decided. The two AI reviewers raised one blocking issue (one that must be settled before publishing), and both raised it. The feel trapped subsense had been written as a state, be in a bad situation, when it is defining a verb. It now reads keep someone in a bad situation … that they cannot easily leave. Both reviewers also said that being trapped is far more common than trapping animals, and I agreed and reordered the senses. Another useful catch: track-v sense 1 was limited to people and animals, although its own examples tracked a car and a phone. It now includes things, and track a package was added. In all, I logged 28 decisions: 23 applied and 5 rejected. The rejections: two requests for a different mark on a phrase (the style guide says a phrase named as a word takes the double mark); usually passive, which is in the closed list of verb patterns; a request to merge the throw away sense into the throw lightly sense (different meaning, different translation); and toss ideas around, which belongs to the phrasal verb toss around.
Cross-references. The tool that links related entries in both directions could not choose a sense in five places. I placed all five by hand: toss on throw-v (its first sense) and on throw-away-phrv, tour as a compare on visit-v, track on follow-v, and trap on catch-v (the sense for stopping an animal or a person from escaping).
Spend. US$0.15 for this cycle; US$0.85 for the whole run and day, out of a US$15 daily budget.
What did not work. Two logging commands were refused at first because two of a reviewer's issues quoted the same words. They were logged correctly using the tool's position option.
What is next and what needs the owner. The next build continues the band-2 verbs in queue order after trap. The coverage reading of band-1 entries and the cross-reference audit are still queued. Nothing new needs the owner.