2026-10-11: originality — thirty-first check, one rewrite
Cycle 4 of the run that started at 12:43 UTC, and the last one (a run has at most four cycles). Run id 20261011T131823Z-7cgg25.
What this cycle did
An originality check tests whether any of our definitions copy a published dictionary. The scheduler forces one every tenth run, and this was run 380. The tool tools/originality_check.py picked ten fields at random from the published entries: five sense definitions, two phrase definitions (idioms), and three usage notes. For each one I ran an exact-phrase web search. I looked only at whether a published dictionary showed the same wording, and I copied nothing.
The results:
- Eight original. like (press a button online), think outside the box (under box), feeling, the usage notes on nobody, explanation, and arrive, spell (letters forming a word), and have a thing about something (under thing). For the three usage notes, learner websites state the same rules in their own words. That is expected for a grammar rule, and the wording is ours.
- One generic overlap. hall sense 4, a building at a college or university where students live, is close to one learner's dictionary definition of hall of residence, but the words come in a different order. A short plain definition like this will look like others by necessity, so it stays.
- One rewrite. common sense 1 read happening often, or existing in large numbers or in many places. The search found a learner's dictionary whose first sense of common matches it almost word for word. That is a distinctive phrase, not an accident of plain wording, so I rewrote it from scratch: seen, done, or found a lot, by many people or in many places. The new wording is inside the defining vocabulary, and the entry's examples (a common mistake, Snow is common here in winter) still fit.
The checklist also puts one standard question to a reviewer model: does each field read as copied? As in recent checks, I used reviewer-a with its reasoning switched off. It answered "original" for all ten fields, so it missed the one match the searches found. Its answer is weak evidence. As before, the searches decide the verdicts. The completed checklist is in reviews/originality/2026-10-11-2.md.
Review of the rewritten entry
A rewritten field gets a full review, like any change to an entry, so both reviewers read all of common again. Reviewer-b found nothing. Reviewer-a raised one blocking issue, about the note that compares common (shared) with mutual. The note said mutual is for feelings and friends shared by two people. That is too narrow, since agreements and interests can be mutual too, and among more than two. I applied it. The note now reads mutual is used for feelings, agreements, and friends that people share with each other. That is one decision, logged in reviews/decisions.jsonl.
Spend
US$0.08 this cycle: US$0.078 for the review panel and less than one cent for the reviewer question. The day's total is US$9.76 of the US$15 cap.
What did not work
Nothing failed. The reviewer model's blind spot for copied wording is unchanged.
What is next
This run ends here, because a run has at most four cycles. The next run's scheduler picks its first cycle. Band-1 adjectives continue after funny, and the oldest one-round entries for review start with game.
For the owner
Nothing new needs you beyond the question recorded in the previous cycle: should the former and the following, used alone, be nouns or pronouns? It is in wiki/open-questions.md.