2026-10-05 (17): the forty-third lint pass
First cycle of the run that started at 16:43 UTC. Mode: lint. Run id 20261005T164354Z-kb063q.
What this cycle did
The scheduler (tools/next_mode.py, the script that picks each cycle's kind of work) forces a lint pass after five runs without one, and this was the fifth. A lint pass is the regular health check of the whole project. Its mechanical part runs the checking tools over every dictionary entry and every page of the knowledge base. Its judgmental part reads the knowledge base (the wiki/ pages that hold the rules and the record of decisions) and the owner-facing files for contradictions, gaps, and out-of-date claims. A lint pass writes no new entries and makes no paid model calls.
Before the lint, the usual start-of-run checks found nothing stranded: no open pull requests from earlier runs, and the one leftover branch from the morning run (claude/wonderful-bardeen-cp7jp7) holds nothing that is not already on main. It is already listed in the curator file for deletion, because the tools available to a run cannot delete branches.
Mechanical checks.
- The size limits on the project's own files are all respected, and every link between wiki pages resolves.
- The cross-reference tool (
tools/crossref.py) makes sure that when entry A names entry B as a synonym, an opposite, or a word to compare, B names A back. It added one missing back-link. The noun air, in its sense "the general appearance or manner that someone or something has" (an air of confidence), lists manner as a synonym, so the tool added air to the synonyms of manner. The tool always attaches a new back-link to the first sense of the target, which here was "the way in which something is done or happens". That is the wrong sense: an air of confidence is about how a person comes across, so I moved the back-link by hand to the second sense, "the way someone behaves toward other people". No pairs named each other under different headings. - The vocabulary check (
tools/lint_vocab.py), which looks for definition words that are outside the defining vocabulary and have no entry, queued nothing new. - The claims tool cleared four old claim files. A claim file records which words a run is working on, so that other runs leave them alone; once the run's work is merged, the file is no longer needed.
Reviewer precision. Each entry is read by two reviewer models, and every issue they raise is applied or rejected. A reviewer's precision in an issue family (pronunciation, labels, examples, and so on) is the share of its issues that were applied. A family below 30 percent over twenty or more decisions is normally switched off for that reviewer.
- No family crossed the line this time; the seven switched-off pairs are unchanged.
- Reviewer A overall: 0.71. Reviewer B overall: 0.54.
- Reviewer B's spelling-and-format family is still at 0.27. It has made no new decisions in that family since the last pass, so the clearer instruction the owner asked for this morning cannot be judged yet.
- Reviewer B moved to a larger model this morning. Since then, 8 of its 24 decisions were applied (0.33). That is still a small sample.
Judgmental checks.
- Every wiki page is listed in the index (
wiki/index.md), and every cycle today has its log entry. - No open questions are left in
wiki/open-questions.md. - The curator file (
reviews/needs_curator.txt, the list of questions only the owner can settle) had two old lines from September recording that a tool had queued additional and upward as nouns. Both words were drafted long ago with the right part of speech (additional-adj,upward-adv), so nothing is left for the owner; I moved them to the Resolved section. Three open lines remain: the disputed American pronunciation of several, the disputed British pronunciation of medicine, and the old branch to delete. - The count of one-sense entries that still carry a signpost (a short guide word meant to help readers find a sense among several, which a one-sense entry does not need) is 36, matching what
NEXT.mdsays.
What did not work
Nothing failed. The one recurring weakness is the cross-reference tool's habit of putting every back-link on the target's first sense; it is noted in wiki/notes/ and each lint pass checks the new back-links by hand.
Spend
None this cycle. The day's spend stays at US$7.95 of the US$15 daily cap.
What is next
The scheduler picks the next cycle. The queue in NEXT.md is unchanged: building nouns from message, a second reading of the oldest one-round entries from bicycle, the remaining owner ruling (moving the auxiliary uses of have and do into their own entries, one per review cycle), and the tool fix for reviewer B's partial answers.
For the owner
Nothing new. The two pronunciation disputes (several, American; medicine, British) and the old branch to delete are still in reviews/needs_curator.txt.