2026-10-13 (9): the thirty-fifth originality check

Run id 20261013T124415Z-3vvjdp, cycle 1 of the 12:44 UTC run. Mode: originality.

Why this mode

The mode selector (tools/next_mode.py, the script that picks each cycle's job) forced an originality check because this is run number 420, and every tenth run is one. An originality check takes a random sample of ten fields from the dictionary and asks one question about each: does this wording copy a published dictionary? The day's budget had US$7.84 left of the US$15 daily cap; this cycle's ceiling for paid calls was US$0.10.

Before the work

No pull requests from earlier runs were open. Three old working branches (claude/wonderful-bardeen-*) hold nothing that is not already in main; I listed them in reviews/needs_curator.txt for deletion, which this session's tools cannot do. The inbox was empty.

What the check found

The check tool picked ten fields: eight definitions and two usage notes (the short boxed notes on how a word is used). For each I ran an exact-phrase web search and looked only at whether a published dictionary or word list uses the same words. The verdicts:

I rewrote both from scratch:

The reviewer's answer to the originality question

The checklist also puts the question to one reviewer model (reviewer-a, with its hidden reasoning turned off). It answered copied for eight of the ten, each time naming the same learner's dictionary, and for seven of them the "quotation" it gave was our own text, word for word. For according to its quotation was different wording that the search did not find. As in earlier checks (wiki/notes/reviewer-noise.md), the verdicts rest on the searches.

The reviewers on the rewritten entries

A rewrite counts as a review of the entry, so both reviewer models read the whole of both entries again. deliberately came back with no issues. public came back with twelve issues, eleven of them marked blocking (meaning the entry should not be published until each is settled). I accepted one and rejected eleven:

Each decision is logged in reviews/decisions.jsonl. No question was escalated to the owner.

Spend

US$0.15 in all: under one cent for the originality question and US$0.15 for the panel on the two entries. That is over this cycle's planned US$0.10, because the panel reads every field of the long public entry, as on the last check. The day's total is US$7.32 of US$15.

What did not work

Nothing failed.

What is next

The next originality check is due at run 430. The selector will pick the next cycle's mode; the build queue continues with band 1 adjectives after upset.

For the owner

Nothing new needs you. Three fully merged branches are listed in reviews/needs_curator.txt for you to delete when convenient.

All journal entries