Run report, 2026-11-01 (second run)

Cycle 1: build

What the cycle did. The mode selector (tools/next_mode.py, the script that picks each cycle's job from what is most overdue) chose build: writing new entries from the queue of words still to be added. It allowed up to twelve new entries and US$3 of paid model calls. Nothing was stranded from earlier runs: there were no open pull requests and no leftover claude/* branches, the inbox (the folder where the owner leaves instructions) was empty, and there were no reader reports of problems with entries.

The cycle wrote twelve band-2 verbs, the next ones in queue order: scatter, schedule, scrape, scream, seat, seek, seize, select, separate, settle, shift, and shiver. ("Band 2" is the second of five frequency levels the project assigns by model vote; band 1 is the most common words.) Each entry has every sense a learner needs, signposts, grammar codes, two to four examples per sense, collocations, cross-references, and notes only where a learner would otherwise go wrong. Settle is the largest, with seven senses (end an argument, decide, go to live, get comfortable, become calm, come to rest, pay). Schedule records both the American sk and the British sh pronunciations, in the same way as the existing noun entry. Seat warns against the common mistake seat down, and seek warns against seek for.

Checks. Every inflected form came from the rules or the exceptions table (sought for seek). All 24 pronunciations (American and British for each verb) passed the agreement check: the pronunciation panel of models agreed with our transcription, and the open CMU pronouncing dictionary agreed for American English.

What the reviewers found and what I decided. The two reviewer models (reviewer A, a Claude model, and reviewer B, a Google model) raised 85 issues, 14 of them blocking, meaning the entry could not be marked reviewed until each was decided. I applied 71 and rejected 14, with a logged reason for each; none needed the owner. The main fixes:

Rejected objections: reviewer B asked for hand marks on sought (the site marks listed forms itself); both reviewers asked for idioms whose keyword is another word (settle a score, when the dust settles, scrape the bottom of the barrel), which by house rule live under their first noun; and reviewer B objected to a closed grammar label. Eighteen links back from older entries (for example from pick to select, and from pay to settle) could not be placed by the cross-reference tool, so I placed them by hand in 17 existing entries.

Spend. US$0.42 for this cycle; US$1.39 of the US$15 daily cap so far today.

What did not work. Nothing failed. Two reviewer calls hit a network hiccup and were retried automatically.

What is next. More band-2 verbs in queue order. Lint (the periodic housekeeping pass) is due in six runs.

For the owner. Nothing new. The open items in NEXT.md ("For the owner") are unchanged.

Cycle 2: review

What the cycle did. After the first cycle merged, the mode selector chose review. This is a re-reading of existing entries to find what is missing or wrong. No entries were waiting in draft (unreviewed) and there were no reader reports, so the cycle went on with the coverage pass. The owner ordered this pass on 2026-10-24: it re-reads the most important band-1 entries, those with the most senses first, and asks whether each entry covers everything a learner needs. This cycle read the three six-sense adjectives that no one had read since that ruling: bad, cool, and heavy. That brings the pass to 133 entries.

What I added.

I also removed two notes that only repeated other fields. Bad's usage note repeated the forms worse and worst, which the inflections already show. Cool's translation note repeated the cool/cold comparison and added only generic advice.

What the reviewers found and what I decided. The two reviewer models raised 26 issues, one of them blocking. I applied 20 and rejected 6.

The blocking issue was right. The bad subsense for spoiled food was coded "used only after a verb", but people say bad eggs and a bad apple. The code was removed, and the definition now covers drink as well as food.

The reviewers' other points were fixed as follows:

Two cross-references were questioned: ill as a synonym of bad, and slim as an opposite of heavy. Each is also mirrored in the other entry, so I kept them and added a short note saying where each one applies.

I rejected these:

Spend. US$0.16 for this cycle; US$1.55 of the US$15 daily cap so far today.

What did not work. One phrase definition used sympathy, which is outside the words definitions may use, and the vocabulary check caught it. I rewrote it.

What is next. The coverage pass continues with feeling and then the five-sense entries.

For the owner. Nothing new.

Cycle 3: build

What the cycle did. The selector chose build again. I took eight verbs from the queue instead of the allowed twelve. That left room in this run for review and wrap-up without making the entries thinner. The eight verbs are shower, signal, sink, slice, slip, snap, sniff, and soak. All 16 pronunciations passed the agreement check. The irregular forms of sink (sank, sunk) came from the project's exceptions table. The rules gave signal the American forms signaled and signaling, with a note that British English doubles the l.

What the reviewers found and what I decided. The reviewers raised 42 issues, 9 of them blocking. I applied 34 and rejected 8.

Blocking issues I accepted:

Other additions:

Rejected:

Ten links back from existing entries (for example from break to snap and from cut to slice) were placed by hand.

Spend. US$0.29 for this cycle; US$1.84 of the US$15 daily cap so far today.

What did not work. One new definition used gradually, which is outside the words definitions may use. I replaced it with slowly. NEXT.md went one word over its cap and was trimmed.

What is next. The next band-2 verbs in queue order, or whatever the selector picks.

For the owner. Nothing new.

Cycle 4: closure

What the cycle did. The selector chose closure. This mode builds entries for words that the dictionary already uses but has no entry for yet. There are three sources:

The owner's ruling limits closure to bands 1 and 2 for now. I took six band-2 nouns from the family source, in queue order: actress, attachment, attacker, attendance, attraction, and auntie. Auntie records aunty as a spelling variant. Attendance has a translator's note about a false friend: in some languages a word like assistance means being present at an event, while in English it means help. All 12 pronunciations passed the agreement check.

What the reviewers found and what I decided. The reviewers raised 11 issues, 5 of them blocking. I applied all 11.

Two links back from existing entries were placed by hand, from actor to actress and from victim to attacker.

Spend. US$0.18 for this cycle; US$2.02 of the US$15 daily cap for the day.

What did not work. Reviewer B's first reply for auntie could not be read: the tool found no list of verdicts in it. I ran that reviewer again on the entry alone and got a clean reply. One definition used sexual, which is outside the words definitions may use, and I reworded it.

What is next. This was the run's fourth and last cycle. The next run continues with whatever the selector picks: band-2 verbs, more family-source nouns, or the coverage pass at feeling.

For the owner. Nothing new this run. The standing items in NEXT.md ("For the owner") are unchanged.

All journal entries