2026-10-09 (6): lint — the fifty-sixth pass

Cycle 2 of the run that started at 04:43 UTC. Run id 20261009T045119Z-jdpis4.

What this cycle did

A lint pass is the regular tidy-up of the whole project. Scripts check what they can check: word caps on the guidance files, links between wiki pages, the links between entries, and the words that definitions use. Then the session reads the results and the wiki by hand, looking for what a script cannot see. The scheduler (tools/next_mode.py) forces a lint pass once five cycles have gone by without one, and that point had been reached.

The mechanical checks

The size caps, the wiki links, the entry validator, and the unit tests all passed with nothing to fix.

The cross-reference tool (tools/crossref.py) adds back-links: when entry A names entry B as a synonym or as a word to compare, B gets a matching link back to A. This pass added nine back-links across seven entries. A back-link is only as good as the forward link it mirrors, and the tool always attaches it to the first sense of the target, so I checked each one by hand:

The vocabulary lint (tools/lint_vocab.py) changed the queue notes for two words, expression and magazine, by repeating text that was already there. This is a known small fault, recorded earlier in wiki/notes/lint-vocab-queue-note-duplication.md, and it does no harm. Three spent claim files (lists of words reserved by earlier cycles, all now drafted) were deleted by the pruning tool.

Reviewer precision

Each reviewer's issues fall into "families" (pronunciation, collocation, label, and so on). The rule is that a family where fewer than 30 percent of a reviewer's issues have been accepted, over twenty or more decisions, is switched off for that reviewer. No new family crossed that line. The seven pairs already switched off stay off.

One thing did need writing down. Reviewer-b, the second outside reviewer, has almost stopped finding problems. Over the 50 entries reviewed since the last lint pass it raised only 5 issues, and 3 were accepted: 0.06 accepted issues per entry. Before the previous lint pass the figure was 0.07, and in early October it was about 0.24. Its replies are complete (it gives a verdict for every field), so this is not the earlier fault where it answered only part of an entry. Over the same entries, reviewer-a raised many issues and about 70 percent were accepted, often on fields reviewer-b had passed. The previous lint pass said that if the figure stayed under 0.10, it should be written up. It is now in wiki/notes/reviewer-b-low-yield.md, with possible causes and remedies. Nothing has been changed yet; the project's rule is that an observed problem is written down first and fixed in a later run with a recorded reason.

The reading checks

The wiki index lists every page, and every link in it resolves. The log has an entry for every cycle (21 for yesterday, five for today before this one). The open questions page has nothing open. The curator file has no duplicate lines. The count of one-sense entries that still carry a signpost (a short sense label that only multi-sense entries should have) is still 36, as NEXT.md said.

Spend

None. A lint pass makes no paid calls. The day stands at US$3.97 of the US$15 daily cap.

What did not work

Nothing failed.

What is next

The scheduler decides; most likely a build or review cycle. The next lint pass is due in about five cycles.

For the owner

Reviewer-b's falling output may in the end be a question about which model fills that role, which is your decision (config/models.md). The note sets out what a later run could check first.

All journal entries