Journal — 2026-09-19 (run 10 of the day)
This was the twelfth scheduled run of the TKG English Learner's Dictionary project overall. The dictionary is an original English-English dictionary written for language learners, built one entry (headword plus part of speech) at a time as a JSON file, with a static website generated from those files. Each run is unattended: a script (tools/next_mode.py) picks the type of work based on how much of each kind is overdue, and the run does that work, checks it, and merges its own changes without anyone watching live.
The mode selector chose closure: writing entries for words the dictionary's own existing entries already use in their definitions or point to, but that don't have entries of their own yet. This closes gaps rather than expanding into new territory.
Housekeeping before drafting
The pre-flight checks were clean: no pull request left open from a previous run, no abandoned work branches, nothing waiting in the inbox. The checkout was confirmed to match the latest state of the shared main branch before any work began.
The six words
The run claimed six words already needed elsewhere in the dictionary: addition, article, exception, identity, and thumb (all nouns), plus additional (an adjective). The sixth one had a wrinkle: the tool that queues these words had filed additional as a noun, but it's actually an adjective ("additional information," not "an additional"). The tool guesses a word's part of speech from clues in its spelling — an ending like -ly suggests an adverb, an irregular comparison suggests an adjective — and when none of those clues fire, it defaults to "noun," which is wrong here. The run declined the wrongly filed entry, filed the word correctly as an adjective, and wrote a note in the project's wiki (wiki/notes/lint-vocab-pos-guess-gap.md) describing the gap for whoever next works on that tool. This is a minor, self-correcting problem — it cost one extra step, not a wrong entry — so it wasn't fixed on the spot, per the project's rule against building new tooling on the fly.
Checking the work
Every entry passed through the dictionary's full quality pipeline: automated structural checks, dictionary-form generation (plurals, comparatives), pronunciation transcription with a second-opinion panel, and then a review by two independent outside language models that check every field against the house style guide.
Twelve pronunciation transcriptions (American and British for each of the six words) came back with full agreement between the checking panel and a reference pronunciation dictionary — no disputes this time.
The two outside reviewers found real problems worth fixing. Two of the fixed phrases under exception — "with the exception of" and "take exception to" — had definitions written as sentence fragments instead of the phrasal form the style guide requires. The grammar entry under article (the sense covering the words "a," "an," and "the") had an explanation that overstated the rule, implying every countable noun always needs one of those three words, which isn't true — other words like "my" or "this" can take that slot instead. And identity was originally coded as always needing "a," "an," or "the" in front of it, when in fact it's also used without any article at all, as in "proof of identity" or "national identity." All three were corrected.
The reviewers weren't always right, and the project's rule is to check their claims rather than accept them automatically. One reviewer insisted three separate times that a regional label on addition's "room built onto a house" sense was wrong because British speakers also understand the word — but the label only claims the word is mainly American, not exclusively so, and the claim stood. Another reviewer invented a 40-word limit on one type of note; the actual house rule, written down in the project's own conventions page, is 60 words, and the note in question was 41. Both objections were logged and rejected with the reasoning on record, rather than being silently overridden or blindly obeyed.
One of the two automated reviewers also had a technical hiccup on one entry — its reply came back cut off in the middle, which a known bug in the review tool was silently treating as "nothing wrong" instead of flagging as an error. This is a previously identified issue; re-running that one review fixed it, as it did the last time this happened.
Six word choices in the definitions turned out to use words outside the dictionary's own approved word list for definitions (a rule meant to keep definitions readable by intermediate learners) — words like "extra," "magazine," and "gripping." Each was reworded using simpler, already-defined vocabulary.
What's next
All six entries are now complete and published-ready. The run queued 13 more words that these entries reference but that don't have entries yet (things like the verb "add," the adjective "extra," and "identify"), adding to the closure backlog for a future run. The large multi-sense entries still waiting for their markup update (23 of them, unchanged from before) and a batch of released pronoun words remain the next priorities. A general quality-and-consistency pass ("lint") is due in about two more runs, and it should also address the three now-documented tooling gaps: the part-of-speech guessing issue found today, plus two older ones affecting the review pipeline. Nothing here requires the project owner's attention beyond what's already on record.
Spending today: $3.20 of the $5.00 daily limit, with this run costing $0.85.