2026-09-25 (second run): lint, the eleventh pass
What this run was
This was the sixty-sixth scheduled run. The run-selector script (tools/next_mode.py) chose lint, a housekeeping pass due every five runs. Lint writes no new entries. It runs the checking tools over the whole dictionary and reads the project's notes for problems that a script cannot see.
There was no unfinished work from earlier runs: no open pull request (a proposed batch of changes waiting to be merged), no leftover branch, and nothing in inbox/.
The mechanical checks
- Size limits and links. Every rule file and wiki page is within its word limit, and every link between the project's pages works.
- Cross-references. When one entry lists another as a synonym, an antonym, or a word to compare, the second entry should point back. The tool
crossref.pyadds these missing "back-links". It added 11 this time, to 11 entries. - Defining vocabulary. The check for definition words outside the controlled vocabulary found nothing new to add to the queue of words still to be written.
- Old claims. Four old "claim" files were removed. (A claim file reserves words for a run that is drafting; these runs had finished.)
- Reviewer accuracy. Every entry is checked by two outside AI reviewers. The project keeps count of how often each reviewer's objections turn out to be right. A type of objection that is right less than 30 percent of the time, over twenty or more decisions, gets switched off. No type crossed that line this time. Both reviewers are right about 70 percent of the time overall. The type closest to the line is reviewer A's pronunciation objections, at 34 percent.
The real work: back-links on the wrong sense
A back-link has to go on one particular sense (meaning) of the target word. Until now, when nothing in the target entry pointed to that sense, the tool simply put the back-link on sense 1. The last lint pass found five of thirteen back-links in the wrong place this way. This pass found two of eleven:
- get ("become ill") had been added to develop sense 1 ("grow or change slowly"). The right place is sense 4 ("start to have an illness, a problem, or a feeling"), and it was moved there by hand.
- run ("take someone somewhere quickly, especially in a car") had been added to drive sense 1 ("control a vehicle"). It was moved by hand to sense 3 ("take someone somewhere in a car").
The other nine were checked and are in the right place.
The problem has now happened twice, so the tool was changed, and the reason is recorded in the log. Now, when no sense already points to the source word, the tool counts the ordinary words that the source sense and each target sense share, leaving out very common words like the, something, and make. The target sense with the most words in common wins. It needs at least two shared words, or else the back-link goes on sense 1 as before. The run tested this on the dictionary's 585 existing pairs of senses that already point to each other. Always choosing sense 1 would have been right 68 percent of the time; the new rule is right 77 percent of the time. That is an improvement, not a cure. The new rule would still have missed the develop case, because those two definitions share only one word. So lint runs still check each new back-link by hand. Four unit tests (small automatic checks of the tool) now cover the new behaviour. The tool's own description had also said that it writes the note "added as a back-link", but it never did. The description now says what the tool actually does.
The judgment checks
Every wiki page is in the index; the curator file (the list of questions for you) has no duplicates; no open question is stale; no pages contradict each other.
Spend
No paid model calls. Today's total stays at US$0.44 of the US$5 daily cap.
What is next
The next runs return to building new entries: band 1 verbs from exercise onward (exist, expand, expect, express, fall, feel...). There will also be second review rounds on older entries, about 312 of which have been reviewed only once. The next lint pass is due five runs from now, and an originality check (a search for any wording too close to published dictionaries) is due soon.
What needs you
Nothing new. The three open questions are unchanged: 4 (whether be, have, do, and one should be split by part of speech), 5 (how the bold and italic marks look), and 6 (whether infinitive to gets an entry). See wiki/open-questions.md.