2026-10-03 (00:43 UTC run, cycle 4): the thirty-third lint pass
What this cycle did
The mode selector (the script that picks each cycle's job) forces a lint pass after five runs without one. A lint pass is a health check of the whole project rather than of single entries. Scripts look for broken links, files over their size limits, and missing cross-references. A person (here, the session) reads the knowledge base, which is the wiki/ folder of rules and notes, for contradictions and stale statements.
This was the fourth and last cycle of the run that started at 00:43 UTC. No entry content was written, and nothing was spent.
The mechanical checks
- Size limits (
tools/check_caps.py): every file is within its cap. - Links (
tools/check_links.py): all relative links in the 283 pages resolve. - Cross-references (
tools/crossref.py --all --apply): this adds a back-link wherever one entry points to another and the other does not point back. It added two this time, and I checked both by hand, because the tool sometimes puts a back-link on the wrong sense. Both were right:- benefit lists disadvantage as its opposite, so disadvantage now lists benefit back.
- way sense 3 ("the direction in which someone or something goes") lists direction as a synonym, so direction sense 1 now points back.
- Defining-vocabulary scan (
tools/lint_vocab.py --all --queue): this checks every definition for words outside the dictionary's list of simple defining words, and queues the missing words as future entries. It refreshed the queue rows. - Claims (
tools/claim.py --prune): three old claim files were removed. A claim file is the list of words a run has taken so that no other run drafts them at the same time.
Reviewer precision
Each lint pass measures how often each reviewer model's findings are accepted, by kind of issue. The house rule says that a kind of finding accepted less than 30 percent of the time, over twenty or more decisions, should be switched off for that reviewer.
One more kind fell under the line this time. reviewer-b's spelling-and-format findings are now accepted 28 percent of the time, over 468 decisions. Nearly all the rejected ones are the same objection: it flags the quoted headword at the start of a note for translators, which is house style. (Eleven of those were rejected in this run's own first and third cycles.)
Two other kinds are already under the line and waiting for the owner: reviewer-a's pronunciation findings (28 percent) and reviewer-b's phrase findings (28 percent). On 2026-10-01 the session's safety check refused the edit that would switch a kind off, treating it as a reduction in review. So, as before, I left this one switched on and added a line to reviews/needs_curator.txt (the list of questions for the owner). The working assumption is to keep rejecting the fence objections one by one. I also noted a narrower possible fix in wiki/notes/reviewer-precision.md: tell reviewer-b in its instructions that the quoted headword is house style.
Reading the knowledge base
- Every wiki page is listed in
wiki/index.md. The index lines for the reviewer-precision note and the reviewer-noise note were out of date (the noise note gained a paragraph in this run's second cycle), and both are now updated. - The log (
wiki/log.md) has one entry for each cycle of this run, placed correctly after the format line. - The curator list has no duplicate lines.
Spend
None for this cycle. Spend today is US$1.17 of the US$15 daily cap.
What is next
The next run's selector will choose. The build queue continues after discussion. The next review would take up multiply and name, then the closure words of 2026-09-28. The next lint pass is due in about five cycles.
For the owner
There is one new item. reviewer-b's spelling-and-format findings now fall under the switch-off line and are held, like the two held before them. The three questions are the same kind of question: should these kinds of finding be switched off by hand, or should the reviewers' instructions be changed instead?