2026-10-09 (18): lint — the fifty-eighth pass
Cycle 2 of the run that started at 16:44 UTC. Run id 20261009T170140Z-xcqumb.
What this cycle did
A lint cycle checks the whole dictionary and its knowledge base (the wiki/ folder of notes and decisions that later runs read) for things that have drifted out of line: broken links, files over their size caps, links between entries that do not match, and contradictions between pages. The scheduler forces a lint pass after five cycles without one, and this was the fifth. No paid calls are made in a lint cycle.
The mechanical checks
- Caps and links. Every file is under its word or line cap, and every link inside the knowledge base points to a file that exists.
- Defining vocabulary. The vocabulary checker read all 1,908 entries. No definition uses a word outside the defining vocabulary (the list of about 2,400 plain words that definitions should use) without an entry of its own.
- Claims. Three old claim files were removed. A claim file reserves words for one cycle so that two runs do not draft the same word.
- Cross-references. When entry A names entry B as a synonym, an opposite, or a word to compare, the cross-reference tool adds the matching link in B (a back-link). Over the whole dictionary it added eleven, all into entries written today, from older entries that already pointed to them. For example, tour now has its match in trip, lie in truth, and hollow in valley. I checked each one. Nine were on the right sense. Two were not, and I moved them by hand: form ("a type or kind") had been attached to type in the sense "a kind of person", and market ("buying and selling") to trade in the sense "a kind of business". Both now sit on sense 1. This is the tool's known weakness: it cannot always tell which sense a link belongs to.
- A mismatch. University named college as a word to compare, while college named university as a synonym. The two must agree, so university now lists college as a synonym too, keeping the note on how the words differ.
Reviewer precision
At each lint pass I measure how often each reviewer's objections turn out to be right (the share I applied rather than rejected), by type of issue. A type of issue whose precision falls under 30 percent over twenty or more decisions is switched off for that reviewer: its objections are no longer shown to the adjudicator.
This time one pair crossed the line. Reviewer-b's objections about inflections (the plural, past tense, and other forms of a word) had been right only 2 times in 21. Almost all of the rejected ones asked for plurals that the inflection tool owns and cannot yet write, or for forms on auxiliary verbs, which by the schema carry none. I switched the pair off in tools/review_panel.py; the same family was already off for reviewer-a, so neither reviewer now checks inflections. That is safe because the forms come from a rule-based tool with its own exceptions table, not from the drafter.
Overall, reviewer-a's objections are applied 72 percent of the time and reviewer-b's 53 percent. Reviewer-b raised only two issues on the twelve entries of the previous cycle. Its yield was low once before and then recovered, so I noted it in the precision page to watch, without changing anything.
The judgment checks
- The wiki index lists every page, and every index line points to a page that exists.
- The log, the journals, and the metrics file each have seventeen records for today before this cycle, so no cycle is missing.
- The open questions and the curator file (the list of points a person should look at) have no stale or duplicated lines. The two old branches listed for deletion are still there.
Spend
US$0 this cycle. The day stands at US$11.93 of the US$15 daily cap.
What is next
The next lint pass is due in about five cycles. The scheduler will choose the next unit; the noun build continues after video.
What needs the owner
Nothing new. Three pronunciation questions (several, medicine, outside) are still open in reviews/needs_curator.txt, and two old branches can be deleted.