Run report, 2026-10-30 (third run of the day, started 12:43 UTC)

Cycle 1: build

What this cycle did. The selector is the script that picks each cycle's kind of work from the project's state. It chose build, which means writing new entries from the queue (the ordered list of words still to be written). It said this kind of work had fallen furthest behind, and it allowed up to 12 entries and US$3 of paid model calls. I wrote the next twelve common ("band 2") verbs in queue order: assure, exchange, exclude, exhaust, exhibit, exit, experience, face, fade, file, film, and fish. Every one went through the full pipeline:

The dictionary now has 3,353 entries, all reviewed.

Before this cycle I checked for work an earlier run might have left behind. There were no open pull requests and no reader reports filed as issues. The five leftover claude/ branches hold nothing that is not already on main, and each one is already listed for pruning in the curator file (the list of questions and chores for you).

Some choices in these entries:

What the reviewers found and what I decided. The two reviewers raised 68 issues, 5 of them blocking (a blocking issue must be settled before the entry can be published). I applied 56 and rejected 13. One decision was logged twice, because I first applied a change and then reversed it.

Back-links placed by hand. The cross-reference tool links entries both ways, but it no longer guesses which sense of the other entry to link from. Two new synonyms needed that: ensure now names assure, and tire now names exhaust, each in its first sense.

Spend. This cycle cost US$0.39 (pronunciation panel and reviewer B; reviewer A costs nothing). The day's total is US$2.10 of the US$15 daily cap.

What did not work. Nothing failed. The reviewer tool needs an index number when one reviewer raises two issues on the same field, so a few decisions were logged a second time with it.

What is next. More band-2 verbs in queue order, then the review, closure, and lint work listed in NEXT.md.

What needs you. Nothing new from this cycle. The open items under "For the owner" in NEXT.md have not changed.

Cycle 2: build

What this cycle did. The selector chose build again, with the same limits: 12 entries and US$3. I wrote the next twelve common verbs in queue order: flee, flood, forecast, function, gain, glance, glue, grab, grade, graduate, grin, and grind. Each went through the same full pipeline as in cycle 1. The panel verified all 24 pronunciations. Three verbs have irregular forms, and those come from the project's exceptions table: fled, forecast (with forecasted noted as also used), and ground. The dictionary now has 3,365 entries, all reviewed.

Some choices in these entries:

What the reviewers found and what I decided. 56 issues, 16 of them blocking. I applied 34 and rejected 22, which is more rejections than usual.

Reviewer noise. Reviewer B's precision this cycle was low (0.48, meaning about half its issues were accepted), mostly because of the hand-mark and usually passive objections. I recorded the pattern in wiki/notes/reviewer-noise.md. If it recurs, a later run can add a line to the reviewers' house rules.

Back-link placed by hand: serve sense 6 ("be useful") now names function.

Spend. US$0.31 this cycle; US$2.42 so far today.

What is next. Band-2 verbs continue in queue order.

What needs you. Nothing new.

Cycle 3: review

What this cycle did. The selector chose review: a second reading of entries that already exist. There were no unreviewed drafts and no reader reports, so the cycle continued the coverage pass. That pass is your ruling B1 of 2026-10-24: each important band-1 word is reread to check that it has every common sense, starting with the words that have the most senses. I read four entries: fill (verb), fix (verb), heat (noun), and high (adjective). That makes 109 entries read in the pass so far.

What I added.

Before the reviewers saw them, I also removed facts stated twice. Our style rule is "one fact, one field": each point lives in one place in the entry. For example, the British word heating appeared both in a sense explanation and in the usage note of heat. The fact that tall, not high, is used for people appeared in four places in high.

What the reviewers found and what I decided. 51 issues, 9 of them blocking. I applied 41 and rejected 10.

Spend. US$0.25 this cycle; US$2.67 so far today.

What did not work. Adding a sense in the middle of heat broke a sense number in a note. The reviewers caught it. Notes that cite sense numbers break whenever senses move, so I now avoid them.

What is next. The coverage pass continues with level and life and the other seven-sense entries not yet read.

What needs you. Nothing new.

Cycle 4: build

What this cycle did. The selector chose build again and allowed 12 entries. I wrote eight. A cycle late in a long run risks thinner entries, and the cap is a ceiling, not a target. The eight are the next common verbs in queue order: grip, group, guarantee, guide, hesitate, highlight, hop, and hum. Each went through the full pipeline. The panel verified all 16 pronunciations, and every verb form comes from the regular rules. The dictionary now has 3,373 entries, all reviewed.

Some choices in these entries:

What the reviewers found and what I decided. 40 issues, 3 of them blocking. I applied 35 and rejected 5. Both reviewers had high precision this cycle (about 0.86 to 0.89), meaning most of what they raised was right.

Back-link placed by hand. jump sense 1 now names hop as a word to compare.

Spend. US$0.32 this cycle. The whole run cost about US$1.27, and the day's total is US$2.98 of the US$15 cap.

What did not work. I started one reviewer call in the background of a shell command. It stopped when that command ended, after reviewing only one entry. I ran the rest in the foreground; nothing was lost.

What is next. The next runs continue with band-2 verbs in queue order and the coverage pass (level, life, and the other seven-sense entries). A lint pass is due in six runs.

What needs you. Nothing new from this run. The open items are unchanged in NEXT.md under "For the owner". They are: the name-after-a-title question, the PROJECT.md line about two non-Anthropic reviewers, the reviewer-A pronunciation switch, and the open questions on continent names and the former.

All journal entries