Run report, 2026-10-31

Cycle 1: build

What the cycle did. The selector (the tool that picks each cycle's kind of work) chose build: writing new entries from the queue, the ordered list of words still to write. It claimed twelve band-2 verbs (band 2 is the second most important frequency group) in queue order: interpret, interrupt, investigate, joke, justify, knit, label, leap, line, link, load, and loan. All twelve were drafted from scratch, checked, reviewed by both reviewers, corrected, and marked reviewed, which means they will appear on the public site. The weekly digest was also due, so the cycle wrote journal/digest-2026-10-31.md first: the week's numbers, ten entries for you to sample, and the questions that wait on you.

Checks before review. The inflection tool generated every verb form from its rules and the exceptions table (knitted, also knit; leaped, also leapt; American labeled, British labelled). The pronunciation check, a panel of three models plus the open CMU pronouncing dictionary consulted at run time, agreed with every transcription but one: the American form of interrupt was first written with a doubled r sound; it was corrected to ˌɪn.təˈrʌpt, which then had every vote and the CMU dictionary on its side.

What the reviewers found. The two reviewers (reviewer A, a Claude model; reviewer B, a Google model) raised 84 issues, 10 of them blocking (an issue that must be settled before an entry can be marked reviewed). Every one was decided by hand: 66 applied, 18 rejected, none escalated to you. The main changes:

What was rejected, and why. Three requests to add idioms were refused under the keyword rule (an idiom goes under its first noun, else its first verb): knit your brows belongs under brow, line your pockets under pocket, look before you leap under look. Reviewer B's requests to hand-mark forms like joked in examples were refused: the site marks listed forms itself. Its objection to the corrected American interrupt transcription was refused against a unanimous panel and the CMU dictionary.

Cross-references. The cross-reference tool adds back-links automatically when it can tell which sense of the other entry to attach them to. In five cases it could not, so they were placed by hand: join, disturb, research, and examine now point back to the new verbs, and loan now treats borrow as a word to compare rather than an opposite, matching how lend already does it. One new queue row was added, the phrasal verb load up, which a reviewer asked for.

Spend. US$0.38 of the day's US$15 budget (reviewer B and the pronunciation panel; reviewer A runs at no charge).

What did not work. Nothing failed. Three definitions first used words outside the defining vocabulary (complaint, needles, gallery); the final check caught them and they were reworded.

What is next. More band-2 verbs in queue order; the next whole-dictionary lint pass is due in one run.

What needs you. Nothing new from this cycle. The digest lists six standing items, among them the two open questions and the pronunciation disputes.

Cycle 2: lint

What the cycle did. The selector forced a lint pass because ten runs had passed since the last one. A lint pass is the periodic whole-dictionary check. It runs the mechanical checks over every entry and every wiki page, and then a reading for things a script cannot see. This was the ninety-fifth pass. It made no new entries and changed no definitions.

The mechanical checks. The word caps on the project's own files, the links between wiki pages, the validator, the unit tests, and the definition-vocabulary gate all passed. One cap was briefly breached: the wiki as a whole came to 26 words over its 30,000-word limit once this pass's precision note was added. The fix was to condense an older section of wiki/notes/reviewer-precision.md, the page that tracks how often each reviewer is right. Six claim files left by finished runs were pruned. A claim file is the record of which words a run had reserved.

Cross-references. Every cross-reference in the dictionary should have a matching link back. crossref.py adds those back-links when it can tell which sense of the other entry to attach them to. Across the whole dictionary it wrote back-links to four entries and left 16 pairs it could not place. Eleven of these came from older entries that already pointed at the verbs written this morning. Examples are border, code, decode, read, take, sew, and tie. These were placed by hand on the right sense of each new verb. Two other pairs from earlier runs were placed the same way: exemplify to illustrate, and fall over to trip. The remaining five were left alone. They are forward references that point at an idiom instead of a sense, such as also pointing at well for as well. They belong to the cross-reference audit you asked for in ruling C1, and nothing about them has changed.

The band vote. Since the last vote, tools had added 77 new rows to the queue with no frequency band. The three word-list models voted on them for US$0.012. Each row takes the lower median of the three votes. Two rows were split by more than one band, and both were settled at the middle vote: kid (verb) in band 2 and knitwear in band 4. The record is experiments/band-vote-2026-10-31/README.md.

Reviewer precision. A reviewer's precision is the share of its issues that turn out to be right. Any issue family that falls under 30 percent over twenty or more decisions gets switched off. Nothing new crossed that line. Reviewer A (the Claude model) is right 83 percent of the time overall. Reviewer B (the Google model) is right 81 percent of the time. Reviewer B's spelling-and-format flags are right only 24 percent of the time, but they stay on under your ruling. Its pronunciation flags (6 of 16 right) and inflection flags (3 of 9) are the closest to being switched off, but they do not yet have twenty decisions.

Reports. These reports are queued for review runs. None needs action in a lint pass:

The reading. Every wiki page is listed in the index. The open questions are still current. The curator file (reviews/needs_curator.txt) has no duplicate lines.

Spend. US$0.012 this cycle; US$0.39 for the day.

What is next. Build again: band-2 verbs in queue order. The next lint pass is due in ten runs.

What needs you. Nothing new.

Cycle 3: build

What the cycle did. The selector chose build again. This cycle claimed ten band-2 verbs instead of the allowed twelve, to keep each entry full: exceed, kid, lower, maintain, march, matter, mistake, murder, nod, and note. All ten were drafted, checked, reviewed, corrected, and marked reviewed. Every verb form came from the rules or the exceptions table; mistake takes mistook and mistaken. The pronunciation panel and the CMU dictionary agreed on all twenty transcriptions on the first pass.

What the reviewers found. The two reviewers raised 44 issues, 12 of them blocking (an issue that must be settled before an entry can be marked reviewed). Of the 44, 33 were applied and 11 rejected; none was escalated to you. The real problems were:

What was rejected, and why. Reviewer B raised four blocking issues asking for mistook and mistaking to be marked by hand in the examples. All four were rejected: these forms are in the entry's own inflections table, so the site marks them automatically. Two idioms went to other entries under the keyword rule (an idiom goes under its first noun): lower your guard belongs under guard, and time marches on under time. I could murder a pizza was left out as mainly British and not yet a need for intermediate readers.

Cross-references. In six places the cross-reference tool could not tell which sense to link back to, so the links were placed by hand. Fool, keep, support, kill, shake, and write down now point back to the new verbs.

A tool problem noted. The band-vote tool writes the date "2026-10-24" on every row it settles, whatever the real date. The kid row settled this morning shows the wrong date, though its band is right. This is recorded in wiki/notes/band-vote-settled-date.md for a later run to fix, as the rule against changing tools on the spot requires.

Spend. US$0.30 this cycle; US$0.69 for the day.

What is next. More band-2 verbs in queue order.

What needs you. Nothing new.

Cycle 4: review

What the cycle did. The selector chose review: a fresh reading of entries that already exist. There were no draft entries and no open entry reports. The work therefore went to the coverage pass under your ruling B1. The coverage pass rereads the most important band-1 entries, the ones with the most senses first, and asks whether any common use is missing. This cycle read four seven-sense content words that had not been read since 2026-10-24: deep (adjective), card (noun), drive (verb), and control (noun). That brings the coverage pass to 121 entries.

What was added. These are the uses an advanced learner would expect to find and did not:

What the reviewers found. The two reviewers raised 60 issues, 7 of them blocking. Of the 60, 53 were applied and 7 rejected; none was escalated to you. The blocking ones:

Several notes repeated themselves, which the "one fact, one field" rule forbids:

What was rejected, and why.

Spend. US$0.27 this cycle; US$0.96 for the day, against the US$15 daily budget.

What is next. The coverage pass continues with the unread seven-sense entries course and do, then the six-sense ones. Building band-2 verbs continues in queue order. The next lint pass is due in eight runs.

What needs you. Nothing new from this cycle. This week's digest, journal/digest-2026-10-31.md, has ten entries to sample and six standing items: two open questions, the names-after-titles line, the charter line on reviewers, reviewer A's pronunciation switch-off, and the pronunciation disputes.

All journal entries