2026-09-21 (session 5) — Review run
This was the fifth 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 "review" — a second read-through of entries that a computer-generated review panel had only checked once before, to catch mistakes the first pass missed.
What the session did
The project has 250 finished ("reviewed") entries and no unfinished ones, so the automatic task-picker chose to re-review, rather than write, new entries. It picked the 20 oldest entries that had been checked only once: some of English's most common words and word-parts — the, an, a, hand, money, have, do, used to, must, un- (the prefix meaning "not," as in unhappy), however, -ness (the ending that turns adjectives into nouns, as in happiness), very, o'clock, TV, anyway, ASAP, both, enough, and every.
Two AI models (not the one running this session) reviewed all 20 entries field by field and raised 74 concerns, 53 of them serious enough to block the entry from being marked reviewed again until settled. I read every one against the entry and my own knowledge, and where the reviewers' claims were about real-world facts I wasn't certain of, I checked them against web sources rather than guessing. I fixed 58 of the flagged issues and disagreed with 16, always with a written reason (logged in a permanent decision file).
The interesting findings
The two most notable fixes were structural, not just wording touch-ups. Both review models agreed that the entry for the wrongly included the "the more, the better" pattern as a use of the determiner the (as in "the cat sat on the mat"). That pattern is actually a different word — the same spelling, but a leftover of an old, separate word that happens to look identical today. I split it out of the the entry and queued a new entry for it to be written later. I made the same kind of fix to very, removing "my very own" and "the very same" (which are actually a different, adjective use of very) from the entry for the adverb very (as in "very big").
I also caught three cases where an AI reviewer had said a word was flatly wrong when it was actually a rare-but-real dictionary word (curiousness, strongness, difficultness — all listed in Merriam-Webster, though far less common than curiosity, strength, and difficulty). I verified this by searching the web before removing those examples, rather than trusting either the entry or the reviewer blindly. On the other side, I rejected two claims from one reviewer about IPA pronunciation symbols (the technical alphabet used to show how a word sounds) that were simply misreadings of symbols that were already correct — this is the second time in recent sessions that this particular reviewer has misread stress-mark placement, so I flagged it in the project's ongoing notes as something to watch.
Along the way I also found and documented a small bug in the project's own software: the automatic tool that generates plural forms cannot represent a word like money, which normally has no plural but does have a rare one (monies) used only in legal documents — the tool's shortcut for "no plural" fires before it would even look up an exception. I wrote this up for a future session to fix; today's entries were only mildly affected, since the fact is already correctly explained in the entries' prose.
What this means and what's next
The dictionary's oldest entries — several of which serve as the project's own model examples for how a good entry should look — turned out to have real, fixable errors, which is exactly what a second review pass is for. Nothing here changes the project's overall direction. 184 entries still await a second review round; the next review session should continue working through them, oldest first.
What needs the owner
Nothing new. The three previously open items are unchanged: turning on the public website (GitHub Pages), and two small style questions recorded in the project's open-questions file. Total spending today across all five sessions: $3.20 of the $5.00 daily budget.