2026-09-22 — lint run (eighth pass)
This was an unattended scheduled run of the TKG English Learner's Dictionary project, an original English-English dictionary for learners, written one entry at a time and published as a static website. "The Routine" is the name for one of these scheduled runs; each does one unit of work, checks it, records it, and merges its own work before handing off to the next run. Nobody watches a scheduled run happen, so it has to make its own calls and write down what it decided.
What happened
The automatic mode selector forced a "lint" pass: a periodic housekeeping unit that runs every five scheduled runs, checking things a script can catch (broken links, word-count limits, unit tests) plus things only a careful read can catch (contradictions between two pages, a note that says a problem is unfixed when the fix already shipped, stale bookkeeping). Pre-flight was clean: no leftover pull request from a previous run, no abandoned branch, nothing waiting in the owner's inbox.
Every mechanical check passed with zero errors: word-count and line-count limits on every governing document, all internal links resolving, all 305 automated tests, and the vocabulary and cross-reference checkers. I measured how often each of the two independent AI reviewers' suggestions turn out to be right ("precision"). Both reviewers are trending up slightly (69% and 70% of their suggestions get accepted, up from 68% and 69% last measurement); no new category of feedback needs to be switched off for either of them.
The careful-read part of the job found two real problems. First, the project's own index page still described two known tooling bugs as "not yet fixed," when both fixes had actually landed in the code back on 2026-09-20 — I corrected the descriptions so a future run doesn't waste time re-investigating a solved problem. Second, and more consequential: a bug first noticed on 11 conjunctions (queue rows that get stuck marked "claimed" forever, with nothing actually holding them, usually because a branch that claimed them was abandoned) turned out to have spread to every single "claimed" row in the whole queue — 32 words in total, 16 conjunctions and 16 common prepositions like on, over, and through. None of them had anything backing the claim. I freed all 32 back to draftable status by hand, which is the documented workaround; the actual tool fix (making the system reset these automatically) is still not written and will need to happen eventually, or this will keep recurring.
I also confirmed a previously flagged content bug is still there — one sense of the entry for break has a definition written in a house-style-violating way ("of weather, or of something hidden: begin suddenly" instead of stating the restriction separately) — but deliberately left it alone, since fixing dictionary content properly requires sending it back through the paid two-reviewer panel, and this run's budget for that was zero.
What didn't work
Nothing failed outright this run. The two stale-documentation and stale-queue problems above are both examples of small drift that a periodic careful-read pass exists specifically to catch; nothing else surfaced.
Spending and what's next
This run spent nothing (lint passes don't need paid model calls); the project has spent $3.55 of today's $5 cap. The next scheduled run is due for "build," now the highest-priority mode, with 20 new entries and a $1.25 budget — and it has 32 newly freed, high-value words to choose from, including core conjunctions and prepositions the dictionary badly needs. Two open questions from earlier runs still await a human decision; nothing else needs owner action right now.