Journal — 2026-09-19 (run 3 of the day)
This was the fifth scheduled run of the "Routine," the unattended process that drafts, checks, and publishes entries for the TKG English Learner's Dictionary (a project made of one small JSON file per dictionary entry, published as a website). Nobody reads this in real time; it is the record for the owner (Tom) to catch up on later. The Routine's mode selector chose "build": write new dictionary entries.
Start-of-run housekeeping
No pull request was left open from a previous run, so there was nothing stranded to merge or fix. No orphan branches needed sweeping. The inbox/ folder — where the owner can drop instructions that jump the queue — was empty except for its archive of already-handled items. The checkout was confirmed to hold the latest work from main.
The main work: ten determiners and four pronouns
I claimed twenty headwords from the queue but decided to draft fewer — fourteen — following the project's own guidance that "fewer, better entries beat the cap." I chose ten determiners (words placed before a noun, like that in that book): that, their, these, this, those, what, whatever, which, whose, your. And four pronouns (words that stand in for a noun phrase on their own): all, another, any, anybody.
This dictionary has a strict rule: a word's use as one part of speech (say, a determiner, directly before a noun) and its use as another part of speech (say, a pronoun, standing alone) get separate dictionary entries, cross-linked to each other. So that book (determiner) and I don't like that (pronoun, standing alone) are two different entries. This makes the words in this batch genuinely tricky to draft correctly, because English speakers don't consciously track this distinction — it's easy for an example sentence or an explanation to slip from one use into the other without anybody noticing.
For every entry, I wrote the definitions, example sentences, and grammar notes, then ran the project's full quality pipeline: a form checker, a pronunciation checker (which compares my proposed pronunciations against three other AI models and an open pronunciation dictionary), and a two-reviewer panel — two more AI models, from different companies, that read every field of every entry and flag anything they think is wrong.
The panel flagged 38 issues, 28 serious enough to block publication until resolved. I read every one myself against the entry and my own knowledge of English, rather than applying or rejecting them automatically. I applied 28 fixes and rejected 10 objections as mistaken.
The two most interesting disagreements:
- One reviewer insisted that in which-det (the entry for "which" used before a noun), two example uses — "which one" and "in which case" / "by which time" — were secretly pronoun uses, not determiner uses, and should be deleted from this entry. I disagreed: in each case "which" sits directly in front of a noun ("one", "case", "time"), which is exactly this entry's own definition of determiner use. I checked this against an outside grammar reference rather than just trusting my instinct, found support for my reading (grammarians call this pattern a "relative determiner"), and kept the entries as drafted, logging my reasoning.
- The same reviewer, though, caught two genuine mistakes: my whatever-det entry had let in a use where "whatever" was followed by a whole clause rather than a noun (that's the pronoun entry's territory, not this one's), and my all-pron entry had two example sentences where "all" was directly followed by a noun phrase like "the tickets" — also not pronoun use, but the determiner entry's territory. Both were genuine drafting slips, and I rewrote the affected parts.
While drafting, I also noticed that two of these words — "these" and "those" — can stand alone as pronouns too ("These are mine"), the same way "that" and "this" can, but nobody had ever added those two pronoun entries to the project's to-do queue. I added them, so a future run can pick them up.
Checks, cost, and what's next
Every automated check passed: the schema validator, the word-count limits on key project files, the internal-link checker, 292 automated tests, and two "closed vocabulary" checks that make sure no field uses a value outside its approved list. Today's spending is $0.65 of a $5.00 daily budget; this run's share was $0.33, almost all of it the two AI reviewers' fees.
Nothing needs the owner's attention beyond what was already waiting: a few old branches this tool cannot delete, and one disputed British pronunciation. The next run is due to pick either another batch of reviews (108 entries have only been checked by the panel once so far) or another round of new entries, whichever the scheduler judges more overdue.