2026-10-05 (14): the eighteenth originality check, no rewrites
Second cycle of the run that started at 12:44 UTC. Mode: originality. Run id 20261005T130614Z-wa1mcj.
What this cycle did
The selector (tools/next_mode.py, the script that picks each cycle's job) chose originality. It forces this check every tenth run. The check asks whether any of our wording matches a published dictionary. Every entry must be written from the drafter's own knowledge, never copied, and this check is how we catch a slip.
A script picked ten fields at random from the whole dictionary. Most were definitions of senses and phrases; one was a usage note. For each, I ran an exact-phrase web search and looked at whether a published dictionary shows the same wording. I also asked one outside reviewer model (reviewer-a) whether each "reads as copied". It may answer "copied" only if it can quote the published wording. Each field then gets one of three verdicts:
- original: no match;
- generic-overlap: a short, plain definition that any dictionary would write the same way, which is not copying;
- rewrite: a match with a dictionary's distinctive wording, which must then be rewritten.
What was found
- Eight original. No match for child ("a person or thing whose character has been strongly shaped by a particular time, place, or influence") or setting (the metal part of a ring that holds the stone). There was also no match for set ("decide on a time, a price, a rule, a goal, or a limit"), any old under old, open to discussion under discussion, throughout, or hang, or for the usage note on passed and past in pass.
- Two generic overlaps.
- the other day (under day) is defined as "a few days ago". Collins explains the phrase with the same three words, and no plainer way to say it exists.
- drive, sense 2, is "travel somewhere in a car". Cambridge's definition opens with the same five words and then goes on. Short, plain glosses like this resemble every dictionary by necessity. Earlier checks treated girl, "a female child", the same way.
- No rewrites. No entry was changed.
The reviewer
My first question to reviewer-a went out with its hidden reasoning switched on. It spent its whole 3,000-token allowance thinking and returned an empty reply. That cost US$0.04 for nothing. I asked again with reasoning off, and it answered "original" on all ten, with one short reason each and no quotations.
Over recent checks this reviewer has swung from calling everything "copied" (with quotations search could not find), to "cannot tell" on everything, and now to "original" on everything. So, as for the last several checks, the verdicts come from the web searches. Its answers are recorded beside them. The note on reviewer noise (wiki/notes/reviewer-noise.md) now says to ask with reasoning off from the start, and NEXT.md repeats it.
Spend
US$0.04 this cycle, nearly all of it the wasted first call. The day's total is US$6.00 of the US$15 daily cap.
What did not work
Only the empty first reply. It cost a few cents and is now noted so that the next check avoids it.
What is next
The selector picks the next cycle's job: probably a build or a review. A lint pass is due in about three cycles. The auxiliary verbs be, do, and have are still waiting for review cycles, under the owner's ruling.
What needs the owner
Nothing new. The full record is reviews/originality/2026-10-05.md.