Run report, 2026-10-27 (scheduled run 20261027T004409Z-n9o1cm)

Cycle 1: build

What the cycle did. The mode selector (tools/next_mode.py, the script that picks each cycle's kind of work) chose build: writing new entries from the queue of words still to be written. It gave a ceiling of twelve entries and a paid-call budget of US$3. The words came in the order the owner set for milestone 1 (the roughly 2,400 words the dictionary uses to write its own definitions, which must all have entries before other words). The twelve were: thankful, yearly, scarcely, seldom, wrongly, the phrasal verb take apart, and six prefixes: anti-, de-, mis-, non-, post-, pre-. Milestone 1 now has 18 words left.

Before the build, the pre-flight checks found no stranded pull requests, no open entry reports from readers, and nothing in the owner's inbox. Three old work branches remain on GitHub with nothing unmerged on them; each is already listed in reviews/needs_curator.txt for deletion, since the session's tools cannot delete branches.

Each entry was drafted from scratch and then went through the full pipeline: the schema check; the inflection tool (which writes forms like took apart); the pronunciation check, where three outside models vote on each IPA transcription and the open CMU pronouncing dictionary is consulted where it has the word; and the two-reviewer panel. Reviewer A is a Claude model and reviewer B is a Google model. They read every field and return a closed verdict on each.

What the reviewers found and what I decided. The panel raised 106 issues, only one of them blocking. In the grateful/thankful comparison box for thankful, the example (I'm thankful the train was late too) made no sense on its own. I replaced it with one that does. Of the issues, 79 were applied and 14 rejected, each with a logged reason in reviews/decisions.jsonl. None needed escalation. The main changes:

Rejections included a reviewer calling comparative with more "not a grammar code" (it is one; this is a settled house rule), a request to mark took apart in an example (the site marks listed inflected forms itself), and two proposals to merge senses of non- and pre- that differ in grammar and in how they are usually translated.

The pronunciation panel agreed on every transcription except American anti-. I gave AN-tee first and AN-tie as a variant, and the panel leaned to the other form. Both are standard, so the entry carries a pronunciation-disputed flag and a line for the owner in reviews/needs_curator.txt. This is the same situation as semi- yesterday.

After review, the cross-reference tool listed nine places where a new entry points to an older one but the tool could not tell which sense of the older entry should point back. I placed all nine back-links by hand (for example, beat sense 2, "defeat", now lists take apart as a synonym). The compare links from mis- to un- and dis- were dropped rather than mirrored, because mis- ("badly, wrongly") is not a negative prefix.

Spend. US$0.36 of the day's US$15 cap, all of it on reviewer B and the pronunciation panel. Reviewer A runs on the owner's subscription and costs nothing in the ledger.

What did not work. The definition-vocabulary check caught seven definition words outside the defining list (reverse, undo, advance, mild, criticize, glad, deserved). Each was rewritten in plainer words before the entries were marked reviewed.

What is next. The selector now points to a review cycle. Milestone 1 has 18 words left. New queue rows from this build include yearly (adverb), wrong (adverb), rightly, incorrectly, pro-, and in-.

For the owner. One new pronunciation line (anti-, American). Nothing else new.

Cycle 2: review

What the cycle did. The selector chose review: a second reading of existing entries. This one continued the coverage pass the owner ordered on 2026-10-24, which reads the most important everyday words, those with the most senses first, and adds whatever a learner would expect to find but cannot. The four entries were drive (noun), full (adjective), front (noun), and place (noun), each with eight senses. I read each against my own knowledge of current English. I made the additions below and then sent each entry to both reviewers.

What I added before review:

What the reviewers found and what I decided. The panel raised 63 issues, five of them blocking, all on full. Of these, 58 were applied and 5 rejected. Every decision is logged in reviews/decisions.jsonl.

The blocking issues on full were all real. Sense 2 (full of) was coded "predicative only", meaning it is used only after a verb, but a box full of books shows it right after a noun. Sense 4 was coded "attributive only", meaning it is used only before a noun, while its moon subsense had the moon is full tonight. And an example said Since he got the promotion, he's very full of himself, where since needs he's been. All three are fixed. The reviewers also caught several real weaknesses I had missed:

On drive, reviewer A pointed out that the British "road to a house" sense came second, ahead of the far more common computer drive. The senses were reordered. The usage note had repeated facts found elsewhere in the entry; it now says only that drive appears in street names. I rejected a request for an American-football "drive" sense as too narrow for this dictionary.

On front, I had added up front meaning "honest" to the phrase's definition, but the phrase's explanation already said the same thing with the one-word spelling upfront. The reviewers caught the repetition, and I reverted my addition. The point that English says opposite, not in front of, for something across a street was stated in three fields; it now lives only in the learner-error box. Tail was removed as an antonym of front, together with the matching link in the tail entry.

On place, the new decimal sense now speaks of one figure in a number, not "a number". The signpost "chance to join" became "student or team", which fits he lost his place on the team.

Spend. US$0.27; US$0.63 for the day so far.

What did not work. The defining-vocabulary check caught three words in new definitions (rank, decimal, membership). All were rewritten in plainer words.

What is next. The coverage pass continues with the band-1 words that have the most senses and were read least recently: pick up, come out, field, away, power, quarter.

For the owner. Nothing new from this cycle.

Cycle 3: build

What the cycle did. The selector chose build again. The next twelve milestone-1 words in the queue were six prefixes (sub-, super-, under-, en-, in-, inter-), three combining forms (bio-, mid-, multi-), and three grammar terms (auxiliary, conjunction, determiner). A combining form is a word part like bio- that comes from another language and joins with other parts, some of them not English words on their own (biology). Milestone 1 now has only six words left.

The entries went through the same pipeline as in cycle 1: schema check, inflection tool, pronunciation panel, both reviewers, my adjudication, the definition-vocabulary check, and the cross-reference tool.

What the reviewers found and what I decided. The panel raised 80 issues; 18 decision lines were for blocking issues. Of these, 68 were applied and 12 rejected, each logged in reviews/decisions.jsonl. This batch drew more blocking issues than cycle 1. Prefixes are hard to get exactly right, and several of the reviewers' catches were real mistakes of mine:

Other coverage added after review:

The spelling forms em- (before b, p, m) and il-, im-, ir- are now recorded in each entry's variants field instead of only in prose.

I rejected three kinds of request:

Pronunciation. Two British transcriptions are disputed: auxiliary and en-. For en-, my first transcription (in) failed in both varieties, so I changed the main form to en with in as a variant. American en then passed. Both disputed entries carry the pronunciation-disputed flag and have a line in reviews/needs_curator.txt stating my working assumption.

Spend. US$0.34; US$0.97 for the day.

What did not work. The defining-vocabulary check caught rank (three times) and volunteers in new definitions; they now read "position or importance" and "working without pay".

What is next. Six milestone-1 words remain. When they are built, the owner's ruling E5 asks the run to report that milestone 1 is finished.

For the owner. Two new pronunciation lines (auxiliary and en-, both British).

Cycle 4: site

What the cycle did. The selector chose site: rebuilding the public website from the entries and testing it. The build script (tools/build_site.py) wrote the 3,032 reviewed entries onto 2,642 headword pages. Words with more than one part of speech, such as drive (noun and verb), share a page. In all it wrote 5,903 files into docs/, a folder that is never committed; GitHub Actions builds the public copy from the entries.

Then the checking script (tools/site_check.py) opened the built site in a real browser, at a phone width of 360 pixels and at a desktop width, and ran its 79 checks. These cover:

All 79 checks passed, and nothing needed fixing. The pages for this run's new entries, for example the prefixes sub- and in- and the grammar term conjunction, were built and are covered by the sideways-scroll check.

Spend. None; this mode makes no paid calls. The day stands at US$0.97 of US$15.

What is next. The selector will choose among build (six milestone-1 words left), review (the coverage pass continues with pick up, come out, field, away, power, quarter), and the periodic lint, due in about five runs.

For the owner. Nothing from this cycle.

All journal entries