2026-10-07 (19): lint — the fifty-first pass, three back-links moved by hand
Cycle 3 of the 16:43 UTC run. Run id 20261007T170949Z-2e317i.
What this cycle did
The scheduler (tools/next_mode.py, the script that picks each cycle's job) forces a lint pass after five cycles without one, and this was the fifth. A lint pass is housekeeping: it runs the mechanical checks across the whole dictionary and the project's notes, and it reads the notes for contradictions, gaps, and stale items. It makes no paid model calls.
The mechanical checks all came back clean:
- Word caps. Every capped file is within its limit. One log entry sits at exactly the 200-word limit, which is allowed.
- Links. All relative links in the notes and reports resolve.
- Cross-references. When entry A names entry B as a synonym, a contrasting word, or a related word, the tool adds the reverse link to B (a "back-link"). The full run added nothing new and found no case where the two sides disagree about the kind of link.
- Defining vocabulary. No definition anywhere uses a word outside the 2,394-word defining list that also lacks its own entry, so nothing new was queued.
- Claims. Three old claim files (the records that stop two runs from drafting the same word) belonged to merged work and were removed.
The one real repair
The fiftieth lint pass, earlier today, found three back-links attached to the wrong sense of their entry, but its session was not allowed to make the edit, so it left them for a later run. I moved them by hand:
- report: the "compare record" link moved from the news sense to the official-document sense, which is the one that matches record as written information.
- respect: the synonym way moved from "good opinion" to "detail" (in some respects, in some ways).
- roof: the "compare ceiling" link moved from the roof of the mouth to the roof of a building.
A fresh full cross-reference run afterwards left all three where I put them.
Reviewer precision
Each lint pass measures how often each outside reviewer model's complaints turn out to be right, broken down by kind of complaint. A kind that is right less than 30 percent of the time over twenty or more decisions is switched off for that reviewer. Nothing new crossed that line. Reviewer-a (the OpenAI model) is right 72 percent of the time overall. Reviewer-b (the Google model) is right 54 percent overall, and also 54 percent since its upgrade on 2026-10-05 (73 of 136). Its spelling-and-format complaints remain below the line at 28 percent but stay switched on, as you ruled; there were no new decisions of that kind since the last pass. The seven kinds already switched off stay off.
Other reading
The curator file (reviews/needs_curator.txt, questions only you can settle) has three open lines, all pronunciation disputes, with no duplicates. Every page in the wiki is listed in its index, and every index link points to a real page. The log has an entry for every cycle since the last lint. The open-questions page has no unanswered items.
Spend
US$0.00 for this cycle. The day stands at US$12.03 of the US$15 cap.
What did not work
Nothing failed.
What is next
The next lint is forced in five cycles. The scheduler picks the next cycle's job; about US$2.97 of today's budget remains.
What needs the owner
Nothing new. The three pronunciation disputes in the curator file (several American, medicine British, outside noun stress) are unchanged.