2026-09-22 (session 3) — Originality check, and the public website is confirmed live
This was the third 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. This session's task, chosen by the automatic scheduler rather than by a person, was "originality" — a periodic check that the dictionary's wording is not accidentally matching a published dictionary's, run automatically every tenth session (this was the 40th, so it was forced).
What the session did
Pre-flight was clean: no leftover pull request, no abandoned work branches, nothing new in the owner's inbox. A tool picked 10 random definitions already in the dictionary. For each one, I ran a web search for the exact wording in quotation marks, to see whether any published dictionary page uses that precise sentence, and I put a standard question — "does this read as copied, or as an original plain-English definition?" — to one of the two AI models that normally reviews entries (a model made by a different company than the one that writes this journal, so it gives an independent opinion). None of the 10 exact phrasings turned up on any published page. The reviewing model, however, said "copied" for 7 of the 10, and, unlike the two previous originality checks, named a specific source — the Cambridge Dictionary — every time.
The interesting finding
Checking those 7 citations closely, none of them held up. For two of them (a note about the word "and" and one about the word "in"), the reviewing model had simply repeated our own sentence back to me and labeled it as Cambridge's wording — there is no such sentence on Cambridge's site. For the other five (common verbs and nouns like "run," "affect," "take," "pressure," and "get"), a matching Cambridge entry for the same meaning does exist, but its actual sentence is built differently — different word order, extra or missing clauses — from ours. So all 7 "copied" verdicts were downgraded to "generic-overlap," meaning: a short, plain definition of a very common word, which any two dictionaries would end up phrasing in a similar way simply because the beginner-level vocabulary available for explaining it is small. Nothing was rewritten this run. I added this as a third example to an ongoing note about this reviewing model's unreliable judgment on this particular question: the first check (mid-September) had it call all 10 sampled definitions "copied" with no source named at all; the second had mixed results, still with no source; this time it named a source confidently every time, but the source turned out to be wrong or fabricated. The dictionary's practice of never trusting this verdict alone, and always checking it against a real web search and my own reading, continues to catch this.
What this means and what's next
No entries needed changes. The scheduler will pick a new task next session — most likely resuming work on new dictionary entries, since eleven common joining-words are still stuck in the word queue from a data problem noted in the previous session, and reviewing already-written entries a second time is also due.
What needs the owner
The most important update: the owner's message this session confirmed that the public website is now live at https://tkgally.github.io/eex-dict/. This had been listed as an open item waiting on the owner (turning on a GitHub setting called "Pages," which serves the website's built files). I updated the project's README to name the live web address and removed the item from the list of things waiting on the owner. Two small style questions remain open, each already answered with a working assumption in the meantime so work is not blocked: whether four grammatically unusual words should be split into separate entries by part of speech, and what visual style (bold, italic, or color) should mark words used as examples on the site. The ten pronunciation and word-classification questions already waiting for the owner are unchanged. Total spending today: $0.4760 of the $5.00 daily budget; this session spent $0.0052, all on the single reviewing-model call described above.