2026-09-22 (session 5) — Quality-check session: nothing new to fix, but two things worth knowing
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") and published as a website. The automatic scheduler chose "lint": a periodic housekeeping pass that runs every five sessions, checking things a computer script cannot catch on its own — contradictions between pages, drift between the documentation and the actual tools, and how well the two AI reviewers who check every new entry are performing.
What the session did
Pre-flight was clean: no leftover pull request, no abandoned work branches, nothing new in the owner's inbox. All the mechanical checks passed: word and length limits on every rules file, every internal link in the project's notes, 305 automated tests, and a check that every word one entry refers to (synonyms, opposites, related words) either has its own entry or is correctly queued to get one. That last check also filled in four "back-links" automatically — for example, since "arrive" already listed "leave" as its opposite, the script added "arrive" as "leave"'s opposite too, a mechanical, symmetric fix the rules explicitly allow a script to make.
The more judgment-based part measured how often each of the two AI reviewers' complaints about an entry turn out to be right, broken down by the type of complaint (grammar, pronunciation, wording, and so on). Five combinations of reviewer and complaint-type have been switched off for months because they were wrong far more often than right (for example, one reviewer's opinion on how many examples a phrase should have was right only 16 percent of the time). None of those five changed enough to need revisiting, and no new combination crossed the line into "switch this off" territory either. But one number stood out: the second, cheaper AI reviewer's overall accuracy (69 percent) has now caught up to, and very slightly passed, the first, more expensive reviewer's (68 percent) for the first time since tracking began. I updated the relevant question-for-the-owner to note that the case for upgrading the cheaper reviewer to a pricier model has weakened, not strengthened.
What I found
While reading through the housekeeping checks, I noticed a small but real bug: when the script that finds words used in definitions but missing their own entry queues the same word for later a second time, it sometimes appends an overlapping, half-repeated note to the word's queue entry instead of noticing the information was already recorded. I confirmed this by tracing the exact code path and wrote it up as a new page in the project's notes, since two of the four affected words had already picked up duplicate text. It is purely cosmetic for now (the affected text is bookkeeping data, not any of the dictionary's actual content), so I left the fix for a future session, as the project's own rule requires — write up what's wrong first, fix it later, deliberately, rather than patching things on the fly mid-session.
What this means and what's next
No dictionary entries were changed in content this session, which is expected for this kind of pass. The next session will most likely resume writing new entries, unless the scheduler again forces a housekeeping or quality-sample pass.
What needs the owner
Nothing urgent. The eleven items on the curator's question list are unchanged. The public website remains live at https://tkgally.github.io/eex-dict/. Total spending today: $0.90 of the $5.00 daily budget; this session spent nothing, since housekeeping passes make no paid calls to the AI reviewers.