2026-09-23 (sixth run): a maintenance pass, with three tool fixes
What this run did
This was the fifty-fourth scheduled run, and it was a lint run: a maintenance pass over the whole project, not new dictionary writing. The mode selector (tools/next_mode.py, which picks the most overdue kind of work) forces a lint pass after five other runs, and five had passed. Lint runs make no paid calls to the outside reviewer models, so this run spent nothing.
Before starting, the run checked for unfinished work from earlier runs. There were no open pull requests (proposed changes waiting to be merged into the published version), no leftover work branches, and nothing in inbox/, the folder where you leave instructions.
The mechanical checks
- Word caps and links. Every capped file is within its limit, and every link inside the project's notes (the "wiki") leads to a page that exists.
- Cross-references. When entry A points to entry B (for example as a synonym or a word to compare), B should point back. The back-link tool added the missing return links to 12 entries, among them although, on, to, and with. It found no errors. These back-links are written by a script, which is allowed; no definition or example was touched.
- Defining vocabulary. The check that definitions use only the dictionary's basic defining words found nothing new to queue.
- Claim files. Before drafting, each run writes a "claim file" listing the words it is working on, so that runs running at the same time do not write the same word. Three old claim files, all for words now finished, were deleted.
- Reviewer precision. Precision is the share of a reviewer's complaints that I judged correct and acted on. The rule turns off a kind of complaint for a reviewer when its precision is under 30 percent over at least twenty decisions. Nothing new crossed that line. Reviewer A is at 69 percent overall and reviewer B at 70 percent, both unchanged. The closest live category is still reviewer A's pronunciation complaints, at 34 percent.
Three tool problems fixed
Earlier runs had written up several tool problems in wiki/notes/ and deferred the fixes, as the rules require. This run fixed three of them, each with a small automated test:
- Words stuck as "claimed". In the headword queue (the list of words still to be written), a word could stay marked "claimed" for good after a run gave it up. The drafting tools then skipped it. Earlier lint passes freed 32 such words by hand. Now the queue's tidy-up command (
queue.py sync) puts any word that no claim file lists back to "pending", and it runs in every lint pass and every build run. - Duplicate review records. When a reviewer's reply came back garbled and was re-requested, the tool added a second copy of the other reviewer's record to the entry and kept the broken one. This happened twice today and was cleaned up by hand both times. Now a re-request replaces that run's records.
- Runs counted twice. Some past runs recorded their statistics twice. The mode selector counted each record as a run, so it thought 53 runs had happened when there had been 47. That made the scheduled lint and originality checks come early. The selector now counts each run once. One side effect: the originality check (a spot-check that sample definitions are not copied from published dictionaries) ran this morning at what the old count called run 50. Under the correct count, run 50 is still two runs away, so the check will run again soon. That is expected, and a second sample does no harm.
Judgment calls
- The queue had a separate row for belong to as a phrasal verb, and earlier reviewers had asked whether it should stay. I decided it should not. Belong to is the ordinary verb belong followed by to, and
belong-vsenses 1 and 2 already cover it. The row is now marked "duplicate", and the item inreviews/needs_curator.txt(the list of questions for you) is closed with that ruling. - The notes index and the conventions page now say that these three problems are fixed.
What did not work
Nothing failed. All 361 entries pass the validation gate, and all 309 tool tests pass.
What is next
Next is a build run on the remaining band-1 verbs (care, carry, catch, cause, change, check…), plus the conjunction neither … nor. After that comes the originality check, then more second-round reviews. 252 entries have been through the reviewers only once. Two logged tool problems remain: queue notes that repeat themselves, and one rare plural the inflection tool cannot record.
What needs you
Nothing new. Three questions are still open in wiki/open-questions.md: whether to split be/have/do/one by part of speech, how the inline marks should look, and whether infinitive to should get an entry. Twelve items wait in reviews/needs_curator.txt, most of them disputed pronunciations.