Journal — 2026-09-19 (run 7, originality check)
This was the seventh 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 forces an "originality check" every tenth run, regardless of what else is due. This project's rule is that entries must be written from the model's own knowledge, never copied or lightly reworded from a real published dictionary (even though no copyrighted text is ever fed in during drafting, a model can still land on wording close enough to a specific dictionary's phrasing to be a problem). The originality check is the project's way of testing that, after the fact, on a small random sample.
What I did
The check picked 10 definitions at random from the 132 entries already finished and reviewed. For each one, I searched the web for the exact phrase, to see whether any indexed page — a published dictionary's website, for instance — used that same wording. None did; all 10 searches came back empty. I also asked one of the project's two AI reviewers (a model from a different company than the one drafting, used specifically so it has no stake in defending its own work) a plain question about each definition: does this read as copied from a real dictionary, or as an original plain-English definition? That reviewer answered "copied" for all 10, every time citing only that the writing sounded like polished, standard dictionary phrasing — never naming an actual source or quoting a matching sentence.
That combination — no hard evidence from search, but a confident "copied" verdict from the reviewer on every single item — meant the call was mine to make by reading each definition myself against dictionary wording I remember from training. Seven of the ten were, in my judgment, the kind of short, plain sentence that any two dictionaries would independently write almost the same way for a common, simple idea (for example, defining "and" as showing that one action follows another) — not evidence of copying, just the shared limits of how few ways there are to say something that simple. Three, though, looked like real paraphrases of a specific published dictionary's distinctive sentence shape, not just its idea: the phrase "be that as it may," the pushy, urgent sense of "please," and the phrase "in person." I rewrote all three definitions in new wording of my own, then sent each changed entry back through the project's full review pipeline (two more AI reviewers checking every field, my own read of every flagged issue, and the automated checks) before marking them finished again.
Two of my first attempts at the rewrites accidentally used words outside the small "defining vocabulary" this dictionary limits itself to when writing definitions (for example, "admit" and "mail"), which an automated check catches; I rewrote those again with words already inside the allowed list. While one of the three entries ("person") was open for editing anyway, I also finished a small formatting job it had been waiting on — adding required emphasis marks around words used as examples — since the rule is to do that by hand only when the entry is already open, never as a separate automated pass.
What this means, and what's next
The reviewer's 100 percent "copied" rate, with no source ever named and no supporting search evidence in most cases, looks like that particular reviewer over-triggering on any dictionary-style sentence rather than genuinely spotting copying. I wrote this up as a note for future runs to watch, since it's only been measured once so far. No entries were removed from the dictionary; three got better, more clearly original wording, and the rest of the sample was judged fine as is.
Everything the project checks automatically — formatting rules, internal consistency, word-count limits, automated tests, cross-references between entries — passed. Total spend on paid AI calls today is about $1.09 of the $5.00 daily limit; this run used about $0.16 of that.
What needs the owner
Nothing urgent from this run. The usual small backlog is still waiting: turning on the public website (GitHub Pages), two open style questions, and a short list of branch-cleanup notes — none of these can be resolved by an unattended run, so they sit in the project's files until you have a moment to look. The next run is likely to go back to drafting or reviewing new entries.