Journal — 2026-09-19 (run 7 of the day)
This was the ninth scheduled run of the dictionary project overall, and it ran in lint mode — a housekeeping pass rather than one that writes new dictionary content. The scheduler forces a lint run every five runs, regardless of what else is due, so that small problems (broken cross-references, stale notes, drift between the project's own documentation pages) get caught on a fixed schedule instead of piling up unnoticed. This was the second lint pass since the project started.
Start-of-run checks. No pull request was left open from a previous run, no abandoned work branches existed, and the inbox (where the owner can drop instructions) was empty. The checkout was current with the main branch.
What the pass covered.
- Two purely mechanical checks — word-count limits on the project's own instruction files, and that every link between wiki pages actually resolves — both came back clean.
- A tool that fills in "back-links" (if entry A points to entry B, B should point back to A) found nothing wrong; it just added six missing back-links among the pronoun entries for anybody, anyone, anything, everybody, everyone, and everything. It also reported 723 words the dictionary refers to but hasn't written entries for yet — expected, since those are simply queued for future "closure" runs, not errors.
- A vocabulary check confirmed that every definition still uses only words from the dictionary's own closed list of "defining vocabulary" (the roughly 2,400 words allowed inside definitions, so a learner is never sent to a harder word to understand an easier one). Zero violations.
- Two abandoned claim files (headwords reserved by a run that never finished with them) were cleared so they can be picked up again.
Reviewer scorecard. The project uses two AI models as independent reviewers on every new or rewritten entry, and tracks how often each reviewer's objections turn out to be right (measured by "precision": objections accepted vs. rejected during the human-equivalent adjudication step). Five specific (reviewer, issue-type) combinations were switched off in an earlier lint run for being wrong more than 70% of the time on 20+ tries — things like one reviewer's near-constant, almost-always-wrong objection to how many examples a phrase had. This run re-measured everything else and found no new combination has crossed that same bad threshold, so nothing new was switched off. One reviewer's overall accuracy rose slightly (0.37 to 0.49 over 280 decisions) since the last measurement, but not enough yet to revisit the standing question of whether to swap that reviewer for a stronger model — that's still marked "keep and watch."
A stale record cleaned up. The project keeps a running list of questions or leftover items that need the human owner's attention (reviews/needs_curator.txt). Four of the five open lines were asking the owner to delete some no-longer-needed work branches on GitHub — but checking GitHub directly showed only the main branch remains, meaning the owner had already deleted them. Those four lines were moved to "resolved" so they stop appearing as open items. One genuine open question remains: a pronunciation disagreement on the word several that needs outside pronunciation data the project doesn't have access to.
What didn't happen. No new dictionary entries were drafted or reviewed this run — that's expected for lint mode, which trades a content-writing turn for upkeep. Nothing was spent (US$0 of the day's $5 cap; $1.71 spent earlier today on content runs).
What's next. The queue is unchanged: 23 large, multi-sense entries (words like water, time, make, good) still need a second editorial pass to add inline formatting marks, and 107 entries overall have had only one review round instead of two. A closure run is also due to fill in some of those 723 missing cross-referenced words. Nothing here needs the owner's decision right now beyond the three standing open questions already on record (GitHub Pages activation, whether to split some grammar entries by part of speech, and the visual style of the formatting marks).
Spend this run: US$0.