Run report, 2026-10-26
Cycle 1: site
What this cycle did. The scheduler (tools/next_mode.py, the script that picks each cycle's kind of work) chose a site check. A site check rebuilds the public website from the dictionary's entry files and then opens the pages in a real browser, at a phone width and a desktop width, to make sure nothing is broken. The scheduler gave this reason: of all the kinds of work that were allowed today, the site check had gone longest without being done. Writing new entries from the words the dictionary already uses ("closure") is still blocked, because your ruling of 2026-10-24 says milestone 1, the list of about 2,400 basic defining words, must be finished first; 120 of those words are still unwritten.
Before the site check, the cycle did its usual safety checks. There were no open pull requests left behind by an earlier run, and no "Entry report" issues from readers on GitHub. One leftover branch, claude/gallant-carson-ax7wls, belongs to yesterday's run; its work is all on main already, so nothing was lost. This run's tools cannot delete branches, so I added a line to reviews/needs_curator.txt asking for it to be pruned.
The result. The build tool (tools/build_site.py) turned 2,924 entries into 2,538 headword pages. A headword page can hold more than one entry, for example a noun and a verb with the same spelling, which is why there are fewer pages than entries. No entries were held back as unfinished drafts, and there were no renamed entries needing redirect pages.
The browser check (tools/site_check.py) ran 79 tests, and all 79 passed. Among other things, it confirmed that:
- no page scrolls sideways on a 360-pixel phone screen, across all 2,544 pages;
- searching for an inflected form such as ran finds run first, and pressing Enter opens it;
- a phrase can be found by its words (a cut above finds a cut above someone or something), and a misspelling such as abiilty offers "Did you mean ability";
- hovering on a linked word shows a short preview of its entry;
- the translator's view can be turned on and off, and the setting survives a reload;
- dark mode works, the menu behaves at both widths, and the sense menu on long entries such as act jumps to the right sense;
- a cross-reference to a word with no entry yet is shown in plain grey, as your ruling C2 asks;
- an address for a page that does not exist gives a "not found" page that searches for the word.
The twelve adjectives written in yesterday's last cycle (bent to lucky) appear and render normally.
What changed. Nothing in the build tool or its page templates needed fixing, so no code changed. The website files themselves are never stored in the repository; GitHub builds and publishes them after each merge.
Spend. Nothing. A site check uses no paid model calls.
What did not work. Nothing failed.
What is next. If the run clock allows, the next cycle will take whatever the scheduler picks: most likely building more milestone-1 words, or continuing the coverage review of the most important common words (next on that list are fall, come, catch, ground, fresh, and over). The routine tidy-up pass ("lint") is due in about three runs.
What needs you. Nothing new from this cycle. The standing items are listed under "For the owner" in NEXT.md: the question about names such as Ms. Kim in a few entries, the charter line in PROJECT.md that still says both reviewers are non-Anthropic, and whether you want your old switch-off of reviewer A's pronunciation flags lifted now that reviewer A is a different model.
Cycle 2: build
What this cycle did. The scheduler chose a build: writing new entries from the queue (headwords/queue.tsv, the ordered list of words still to be written). Under your ruling A2 the queue serves milestone 1 first, the basic defining words, so this cycle took the next twelve: the adjectives ordinary, ripe, salty, selfish, sour, spoken, steep, sticky, temporary, unfriendly, and worn, and the adverb greatly. Milestone 1 now has 108 words left.
Each entry went through the full pipeline: the format check, the inflection tool (which writes forms such as riper and stickier from rules), the pronunciation panel (several models transcribe each word and their votes are compared, along with the open CMU pronouncing dictionary for American English), both outside reviewers, my decision on every point they raised, and the checks again.
A few things worth knowing about the entries:
- ordinary has a second, slightly negative sense (the food was very ordinary), the phrase out of the ordinary, and a translator's note warning that in some languages a look-alike word means rude.
- salty covers taste, lively but slightly rude language (salty language), and the informal annoyed (he's still salty about losing).
- sour covers taste, milk that has gone bad, a bad mood, and things that go sour.
- steep covers a slope, a sudden large change (a steep rise), and, informally, a price that is too high.
- temporary has a learner error for a temporal job, a real mistake for speakers of languages where the look-alike word means temporary.
- greatly has a learner error for using it before plain adjectives (greatly interesting).
Pronunciation. All 24 transcriptions (American and British for each word) ended up verified. The panel did not accept my first British form of salty (ˈsɔːl.ti). I changed it to ˈsɒl.ti, the form the dictionary's entry for salt already uses, and kept the other as a variant. It then passed.
What the reviewers found. They raised 45 issues; 8 were blocking (meaning the entry should not stand as it is). I accepted 34 and rejected 11.
The blocking ones were real. In worn, one example (the steps were worn smooth by many feet) and two collocations were really the verb wear in the passive, not the adjective. I replaced them (my favorite jeans are worn at the knees). In sour, one sense was coded "comparative with more" while the entry gives sourer; I removed the code. In unfriendly, an example about plastic bags made a claim about the world, which the house rules forbid; I replaced it with a neutral sentence, and I reworded a confusing explanation.
Among the minor points, I accepted that unfriendlier is standard, so the entry now shows it. steep's summary sentence (the "core idea") now covers the price sense. greatly lost a claim that it is "a little formal", which no label backed up. sticky gained a sticky note.
The rejections were mostly reviewer A calling a gradability code a duplicate of the inflection note. That note is written by a script from the code, so there is no real duplication; the house rules already settle this. Two more asked for idioms whose keyword is another word (a ripe old age belongs under age, a steep learning curve under curve).
Cross-references. The linking tool adds a matching link in the other entry when it can tell which sense is meant. Seven times it could not tell. For three I placed the link by hand: fresh now lists sour as an opposite, sharp lists steep as a synonym, and sour points to salty. The other four synonyms were loose, so I removed them from the new entries instead.
Spend. About US$0.33, for pronunciation and reviewer B. Reviewer A runs on your subscription at no charge to the budget.
What did not work. Nothing failed.
What is next. More milestone-1 words, the coverage review of the most important words (next: fall, come, catch), and lint in about two runs.
Cycle 3: review
What this cycle did. The scheduler chose a review: rereading existing entries and improving them, rather than writing new ones. Under your ruling B1 the review mode is working through a "coverage pass" over the most important common words, those with the most senses first. It asks of each entry: is any common use missing, and does every example fit the sense it sits under? This cycle read six entries: fall, come, catch, ground, fresh, and over (the adverb). Each went through both outside reviewers and my decision on every point they raised.
What was missing, and is now there.
- fall: fall into a group (most students fall into this age group); fall to someone, meaning to become their duty; fall in battle (labeled literary); a face falling in disappointment; and eyes, blame or stress falling on something.
- catch: being caught in the rain or in traffic; hitting someone (the ball caught her on the head); a fire starting to burn; and the informal goodbye catch you later.
- come: one reviewer pointed out that a single sense mixed two meanings, happening at a time and having a place in an order. I split it. The new "order" sense holds B comes before C, family comes first, and race positions (our team came in second), which had been filed under the wrong sense. Also new: come to an end or a decision; come as a surprise; come into view or into effect; music comes naturally to her; the phrase to come (for years to come); and the sexual sense, labeled informal, which the taboo-vocabulary rules say to include and define plainly.
- ground: the phrases middle ground, give ground, run something into the ground, break ground (American, starting to build), and burn to the ground.
- fresh: the old-fashioned American don't get fresh with me.
- over: bending (lean over, bend over), handing something over, staying the night (stay over), and liquid boiling over.
What the reviewers found. They raised 149 issues, 11 of them blocking (meaning the entry should not stand as it is). I accepted 120 and rejected 29.
The blocking ones were real:
- In over, lean over so I can see was filed under "through the air above a place". It is bending, so it went to the new bending sense. We painted over the old color was the preposition, not the adverb, so I replaced it with the cut soon healed over.
- In catch, you won't catch me eating there again is an idiom meaning "I would never". It did not fit the sense "see someone doing something wrong", so I replaced it.
- Two senses of fresh had muddled definitions: air from outside, and a "clean, cool" sense that did not fit clean sheets. I rewrote both.
Many minor points were also right:
- Several "learner errors" were regularized forms (falled, catched). Your notes rules ask for those to be removed, so they are gone.
- Several entries said the same fact in three or four places, such as "fall never takes an object" and "catch, caught, caught". Each fact is now in one place.
- Some examples made claims about the world (fresh vegetables are better for you, milk comes from cows). I replaced them.
- Two translator's notes were generic. They now name a specific trap: many languages call fresh water sweet water, and the "reason" sense of ground is unrelated to soil.
I rejected three kinds of point:
- Idioms whose keyword is another word: fall into place, catch your breath, come to mind.
- Requests to mark broke and stood in the ground phrases. The site marks the headword, and here that is ground, which appears unchanged.
- Claims that a verb pattern such as usually passive is not on the closed list. It is.
I also turned down moving the "reason" sense of ground ahead of the land senses.
Other changes. When a review removes a cross-reference, its mirror in the other entry goes too. drop lost a "compare" link to fall, and jump lost an "opposite" link to fall. Ground rules was added to the queue as a compound noun of its own.
Spend. About US$0.51.
What did not work. Nothing failed.
What is next. The coverage pass continues with hand (one of the three model entries in the style guide), then wild, weak, turn, start, side, and hold.
Cycle 4: lint
What this cycle did. The scheduler forced a lint pass, because ten runs had passed since the last one. A lint pass checks the whole dictionary, not single entries. It has two parts. The mechanical part is a set of tools run over every file. The judgment part means reading the wiki (the project's rule book and notes) for gaps and contradictions.
The mechanical checks.
- Caps and links. The word limits on the rule files all pass, and every link between pages resolves. The wiki stands at 29,872 of its 30,000-word limit, after I shortened the oldest part of the reviewer-precision note to make room for today's figures.
- Cross-references. The linking tool adds the matching link in the other entry whenever one entry points to another. It added six such links. Ten times it could not tell which sense was meant, so I placed those by hand. Most came from this run's new adjectives. ordinary now lists regular as a synonym and odd and strange as opposites. ripe lists green (not ripe) as an opposite. unfriendly lists cool, and cruel lists sick (as in a sick joke). The rest went into makeup, mess, relative, and laundry. One new mismatch appeared: cut calls chop a word to compare, while chop calls cut a synonym. I left it for the cross-reference audit your ruling C1 asked for, with the two earlier mismatches.
- Promised entries. A note sometimes says "this has its own entry". This report lists notes where that entry does not exist and is not queued. It found the same seven as last time, all false alarms: the entries exist under other names, or are homographs.
- Circular definitions. 322 definitions use their own headword, nearly all harmlessly: a second sense that names the first (all the people who live in a village). Reviews judge them one by one.
- Missing senses. 707 pairs where a link may point to the wrong sense. These are the audit's work list.
- Band vote. A band is our measure of how important a word is, from 1 to 5. 64 queue rows that tools had added since yesterday had no band, so the three word-list models voted on them, for about US$0.01. Two votes were spread widely (shampoo and thunder as verbs, votes 4, 3 and 2). I kept the middle vote, 3, for both. The record is in
experiments/band-vote-2026-10-26/. - Claims. Five old claim files (the record that a run had reserved some words) were pruned. Their words are all built.
Reviewer precision. Precision is the share of a reviewer's points that I accepted. Under the house rule, any kind of issue where a reviewer falls under 30 percent, over twenty or more decisions, is switched off for that reviewer. Reviewer A, the Claude model, stands at 0.83 over 805 decisions. Its weakest kinds are "coverage" at 0.54 (mostly idioms whose keyword is another word) and "grammar code" at 0.64 (mostly gradability codes it calls duplicates). Reviewer B, the new Gemini model, stands at 0.88 over 152 decisions, with no single kind yet over twenty decisions. Nothing is under the line, so nothing was switched off.
The judgment checks. Every wiki page is listed in the index, and every index line points to a page that exists. The two open questions for you (continent names in definitions, and whether the former and the following are nouns or pronouns) are still open and still shown in NEXT.md. In the curator queue (reviews/needs_curator.txt), three old requests to delete leftover branches are now closed, because those branches are gone. I added one new request this morning for yesterday's branch.
Spend. About US$0.01 for the band vote. The day's total for the run is about US$0.85 of US$15.
What did not work. Nothing failed. The tool that refreshes the band stored in each entry raised an error after the vote. That is expected here: all 64 voted words are still in the queue, with no entries yet. Their bands will be stamped into the entries when they are built.
What is next. The next lint is due in ten runs. The next cycles will continue milestone 1 (108 words left) and the coverage pass (next: hand, then wild, weak, turn, start, side, hold).
What needs you. Nothing new. The standing items are under "For the owner" in NEXT.md.