2026-09-21 (session 8) — Build run: eleven new entries
This was the eighth automated session of the day on the TKG English Learner's Dictionary, an original English-English dictionary for language learners built as one JSON file per word ("entry"), published as a website. Each session is unattended and picks its own task; this one was assigned "build" — drafting new entries from the word queue.
What the session did
Pre-flight was clean: no leftover pull request, no abandoned work branches, nothing waiting in the owner's inbox, and the local copy was already up to date. The automatic word-claiming tool was asked for eleven new words and, by the ordering it uses (most common words first, then by part of speech), it grabbed the entire remaining set of common prepositions — including on and over, the two heaviest words the project's own baton file had specifically flagged as too big to tackle yet. Rather than draft twelve heavy, closely related function words in one session, I discarded that claim and hand-picked a more varied batch instead, following the baton file's own suggestion: act (as both a noun and a verb), activity, address, adjective, advertise, affect, aim, air, airport, and alarm.
Each entry went through the full required process: writing definitions, examples, and grammar notes from my own knowledge (never from a published dictionary); generating the correct word forms (plurals, past tenses) by rule; checking pronunciation against a panel of other AI models plus a public pronunciation dictionary (all twenty-two transcriptions across the eleven words came back agreed, with no disputes); and a second review by two more AI models, reading every field of every entry against a checklist. That review raised 38 issues, 18 serious enough to block publication. I read each against the entry and the house style guide myself, fixed 35 as genuine problems (mislabeled collocation types, an ungrammatical example, a definition that was too narrow, a usage note that wrongly implied a word is never used as a noun in a specialized field), and rejected 3 where the reviewer had misread the style guide's own rules — one wanted a grammar-topic label repeated on an entry's single sense when the guide says a whole-entry label already covers it. Two AI reviewer replies came back unparseable the first time; re-running just the broken half fixed both without redoing the whole review.
The interesting finding
Seven of the eighteen serious issues were the same underlying problem: a definition had used an everyday-sounding word — opera, behave, aircraft, broadcast, identifies — that turns out not to be one of the roughly 2,400 words the dictionary allows itself to define other words with, and that word doesn't have its own entry yet either. The fix in every case was to reword the definition with a word that's already allowed, not to write a new entry on the spot, which would have gone beyond this session's scope. Separately, while patching one entry's word-pairing examples by script, a small programming mistake (changing a list while stepping through it) sent one script into a runaway loop; I caught and killed it within a couple of seconds, confirmed the target file hadn't actually been written to yet, and redid the fix correctly the second time. No data was lost, but it's a reminder to be more careful with that kind of in-place edit.
What this means and what's next
The dictionary grew from 262 to 273 finished, published-quality entries. Both AI reviewers scored unusually well this session — right about 90 to 94 percent of the time they flagged something — well above their running averages, though this is too small a sample to read much into. Twenty-eight related words that these new entries mention but don't yet define (like law, behave, speech, and airline) were automatically added to the to-do queue. The next build session should keep mixing word types rather than claiming an entire batch of one kind at a time; on, over, up, to, and under are all still unbuilt and still heavy. The scheduler's own next pick, based on where effort is most overdue, is a "site" session — rebuilding and spot-checking the published website.
What needs the owner
Nothing new to decide immediately. The three previously open items are unchanged: turning on the public website (GitHub Pages), and two small style questions already answered with a working assumption. The ten pronunciation and word-classification questions already waiting for the owner are unchanged, plus one new entry (the specialized psychology sense of affect as a noun) was added to the to-do queue rather than the owner's queue — no decision needed there, just future work. Total spending today across all eight sessions: $3.93 of the $5.00 daily budget; this session spent $0.35 on the pronunciation and review panels.