2026-10-04 (10) — lint: the thirty-eighth lint pass
Second cycle of the run that started at 08:44 UTC. Run id 20261004T090107Z-vka7g0.
What this cycle did
The mode selector (tools/next_mode.py, the script that picks each cycle's job) chose lint. A lint pass is a housekeeping check of the whole dictionary and its knowledge base (the wiki/ folder of rules, decisions, and notes). It is forced every five cycles, and the previous one was five cycles ago. A lint pass makes no paid model calls.
It runs two kinds of check.
Mechanical checks, by script.
- Word and line caps on the rule files and the wiki: all within limits.
- Links inside the knowledge base: all resolve.
- The defining-vocabulary check: clean. It confirms that every definition uses only the plain words on the dictionary's defining list, or words that have their own entry. Words still missing are kept in the queue.
- Cross-reference symmetry (
tools/crossref.py --all --apply): when entry A names entry B as a synonym, a word to compare, or a family member, B should name A back, and the script adds the missing half. This time it added two back-links, and I checked both by hand:- good (the adjective) lists goods as a family word. The back-link from goods to good is right.
- fuel names gas as a word to compare, with the note "American, for cars". The script attached the back-link to the second sense of gas, the gas burned for cooking and heating. It belongs on the third sense, gasoline for cars, so I moved it there. The script picks the sense whose definition shares the most words with the other entry, and it is still wrong about a quarter of the time. That is why every lint pass checks these lines by hand.
- Old claim files (the lists of words a run has taken, so two runs do not write the same entry): two were removed.
Judgmental checks, by reading.
- Reviewer precision. Two outside models review every entry, and for each kind of issue they raise we track how often their flags are accepted. The project rule is that a kind of issue accepted less than 30 percent of the time, over twenty or more decisions, is switched off for that reviewer. Six pairs were already off. No new pair fell under the line. The two pairs that are under it but held for your ruling are unchanged:
- reviewer-a on pronunciation, 26 percent: it keeps objecting to the house style for American long vowels.
- reviewer-b on spelling and format, 29 percent: it keeps objecting to the quoted headword at the start of adaptation notes.
- The wiki index (
wiki/index.md) against the pages: every page is listed and every listed page exists. - The open questions (
wiki/open-questions.md) and the curator file (reviews/needs_curator.txt, questions only the owner can settle): nothing stale and no duplicates. - Tool documentation.
tools/README.mddid not listtools/eexlib.py, the shared library the other tools import. One line was added. - Claims in
NEXT.mdand the style guide. The count of 36 one-sense entries that still wrongly carry a signpost (the short sense label used only when an entry has several senses) is still correct. The adverb the (the more, the better), which the style guide says was split off from the determiner, is still waiting in the queue.
Spend
None this cycle. The day's total stays at US$3.50 of the US$15 daily cap.
What did not work
Nothing failed. The cross-reference script again misplaced one back-link out of two. This is a known weakness, recorded in wiki/notes/crossref-first-sense-backlinks.md.
What is next
The next lint pass is due in about five cycles. The selector will choose between building new entries and a second review of older entries. About 480 entries have had only one review round so far.
For the owner
Nothing new. The two held reviewer switch-offs and the seven open questions still wait on you (NEXT.md, "For the owner"). One curator line from the first cycle of this run is also open: the British pronunciation of grandfather and grandmother.