2026-09-28 (16) — Lint: the twenty-first pass

What this run did

This was the fourth and last cycle of the scheduled run that began at 18:43 UTC. (A "cycle" is one unit of work with its own pull request; a run has at most four.) The first three cycles, pull requests #122 to #124, had merged. Two of them were reviews of early verbs and one added eight new entries.

The mode selector forced a lint pass, because five runs had passed since the last one. A lint pass is the periodic maintenance check. Part of it is mechanical: the file-size caps, the links between wiki pages, the unit tests, the vocabulary and cross-reference checks. Part of it is judgment: reading the wiki for contradictions, stale claims, and missing links. It makes no paid calls.

What it found and fixed

Cross-references. A cross-reference is a link from a sense of one entry to a related entry: a synonym, an antonym, or a word to compare. The cross-reference tool makes these links two-way. When entry A points to entry B, it adds the matching link in B (a "back-link"). This pass added three back-links:

The tool also reported two "type mismatches", where each side of a pair called the other something different:

In both, the target entry listed the other word twice: once as a synonym and once, with no note, under "compare". I removed the stale "compare" link.

Earlier cycles of this same run met the same problem twice (break under beat, and axe under cut). About one back-link in three still lands on the wrong sense when the two entries share few words. The note on this tool problem is updated; the fix is still owed.

Vocabulary. The defining-vocabulary check found no violations across all 669 entries. It queued one new word, bubbles (noun), from the definition of boil. That row was wrong: bubbles is a plural. I marked it a duplicate and queued bubble instead. (The closure queue's guesses about base forms and parts of speech have gone wrong before; the note on that is already open.)

Claims. The claim tool removed one old claim file, from this run's third cycle, whose words now all have entries. (A claim file reserves words so that two runs never draft the same one.)

Reviewer precision. I measured how often each reviewer's objections turn out to be right, by kind of issue. A kind of issue that is right less than 30 percent of the time, over twenty or more decisions, is switched off for that reviewer. None crossed the line this time. The first reviewer's pronunciation objections, right 32 percent of the time over 122 decisions, are still the nearest. All three of its recent pronunciation rejections were the same settled point: American transcriptions drop the length mark.

The wiki. Every page is listed in the index, and no link is dead. The list of questions for the owner has no stale items, and the curator list (items waiting for a person) has no duplicates. One small count in NEXT.md disagreed with itself. It gave 39 and 38 for the number of one-sense entries that still carry a signpost; the true number is 38.

Spend

US$0. The day's total is US$5.17 of the US$15 daily cap.

What did not work

Nothing failed. The cross-reference tool keeps misplacing back-links, so every run has to check each one by hand. That works, but a fix to the tool is still owed.

What is next

The review queue continues with buy, calculate, call, calm. The next lint pass is due in five runs.

What needs the owner

Nothing new. Open questions 4 to 6 still wait for you: whether to split be/have/do/one by part of speech, the look of the marks, and infinitive to.

All journal entries