Journal — 2026-09-19 (run 8, review)
This was the eighth scheduled run of the day for the TKG English Learner's Dictionary, an automated project that writes an original English-English dictionary for language learners, one entry (one word or phrase) per file. Each run is unattended: a scheduler decides what kind of work to do, and a fresh instance of the model does it, checks its own work, and merges its own change before stopping.
What kind of run this was
The scheduler picked "review" — a second, deeper look at entries that have only ever had one round of AI review, or that still carry a leftover "markup-pending" flag from a formatting rule the owner introduced on 2026-09-18 (a rule that words discussed as words, or phrases shown as examples inside explanatory text, need to be marked in bold or italics so the published site can style them). With no unfinished drafts waiting, the review queue was next in line, and 33 entries still carried that flag.
What I did
I picked the 10 flagged entries with the fewest word senses — alone, quickly, sorry, borrow, lend, speak, happy, tell, talk, and say — so that a thorough pass on each would fit in one run rather than rushing through all 33. Each entry went through the project's full check: two independent AI reviewers read every field and flagged anything wrong, I read every one of their 96 flags myself and decided whether to fix the entry, disagree with the reviewer, or (never needed this run) ask the owner; then I fixed what needed fixing, added the required bold/italic marks by hand throughout every sentence, and re-ran every automated check.
Several of the fixes were real content errors, not just formatting. The adjective entry for "alone" used two examples ("Kim lives alone," "walking home alone") that are actually adverb uses, not adjective ones — and the entry's own usage note already said as much, so it was quietly contradicting itself; I replaced those examples and tightened the related grammar notes. "Quickly" had a sense defined more like "soon" (a point in time) than what "quickly" actually means (how little time something takes); I reworded it. "Say" had a use — the kind in "meet me at, say, six o'clock" — that didn't fit the sense it was filed under, so I split it into its own, more accurately defined sub-entry. Several entries also had small overclaims ("glad is used only after a verb," "talk never reports what was said") that a reviewer correctly caught as too absolute; I softened those to match how the words actually behave, per the project's own rule against unhedged claims a reviewer can always poke a hole in.
I also found and reported, without trying to fix it myself, a real bug in one of the project's own checking tools: for one entry ("talk"), one of the two AI reviewers' replies got cut off partway through, and the tool quietly recorded that as a clean, zero-problem review instead of flagging the failure — which could let a genuinely unreviewed entry slip through looking reviewed. I caught it by noticing the reviewer's reported count didn't match the size of the entry, re-ran that one reviewer, and got a normal result. I wrote the problem up for a future run to actually fix the tool; the project's rule is that observed problems get written down first and fixed later, deliberately, rather than patched on the spot.
What this means, and what's next
All 10 entries are now marked reviewed, correctly formatted, and pass every automated check (schema validity, banned-word rules, cross-references between entries, 292 automated tests). 23 flagged entries remain, all larger and more complex than this batch, so expect a future review run to get through fewer of them at a time. Spend on paid AI reviewer calls today is about $1.71 of the $5.00 daily limit; this run used about $0.61 of that.
What needs the owner
Nothing urgent from this run. The tool bug I found (a silently empty review result) is worth a real fix in a future session but isn't blocking anything right now, since I caught and worked around it. The same small backlog as before is still waiting on you: turning on the public website (GitHub Pages), a couple of open style questions, and a short list of branch-cleanup notes this tool set can't act on itself.