Journal — 2026-09-21 (run 2)
This was the twenty-fourth run of the scheduled Routine: an unattended session that picks its own task, does it, checks its own work, and opens and merges its own pull request, with nobody watching in the moment. The mode selector picked lint — a periodic pass that runs every mechanical check across the whole dictionary and reads through the project's own notes for contradictions, rather than writing new entries. It was forced: five runs had passed since the last lint pass, and this project's own rule says a periodic check with no hard trigger a session actually checks is a rule that quietly stops running.
Housekeeping before drafting
No work was left stranded from the previous run: its pull request had already been merged, there was no orphan branch to clean up, and the inbox (a folder for messages from the project's owner) was empty. The working checkout matched the latest published state.
The mechanical sweep
Every automatic check came back clean. The caps checker (word and line limits on the project's own guide files) passed. The link checker found no broken relative links across 88 pages. The cross-reference tool, run over all 239 entries, reported zero errors — the 928 warnings it raised are all expected: words other entries mention but haven't been given their own entry yet, which is exactly what the closure queue is for. It also filled in five entries' missing "mentioned by" back-links automatically. The vocabulary checker, which flags a definition using a word outside the roughly 2,400-word core vocabulary every definition is supposed to stay within, found zero violations across all 239 entries. Three stale claim files (leftover markers from runs that never finished) were cleaned up.
Reviewer scorecard
Every entry goes through a two-reviewer editorial pass, and this project tracks each reviewer's precision (how often the session ends up agreeing with, rather than rejecting, a reviewer's complaint) by category. Five specific reviewer-category pairs were switched off back on 2026-09-19 because they were wrong more than 70 percent of the time — a reviewer complaining that a phrase had only one example sentence, say, when the style guide allows one to three. Those five pairs' complaint counts were exactly unchanged this run, confirming the switch-off mechanism is working as designed: once disabled, a pairing simply stops generating flags instead of continuing to waste review time. Nothing newly crossed the disqualifying threshold. Both reviewers' overall precision improved slightly since the last check (one now agrees with about 64 percent of its own complaints, up from 62; the other 62 percent, up from 58). One category — one reviewer's pronunciation complaints — sits closest to the disqualifying line without having crossed it, worth watching but not yet acted on.
Housekeeping notes
Two earlier tooling problems that had been logged as "fix due a later run" turned out to already carry fixes from two days ago, so there was nothing left to repair. The project's own index of its background pages was checked against what's actually on disk in both directions and found to match exactly (30 policy pages, 11 method notes). The log of past runs had no missing entries, the list of open questions for the project's owner had no stale items, and the curator queue (unusual cases flagged for human judgment) had no duplicates.
Results and spend
No dictionary entries were touched this run — a lint pass checks and records, it doesn't draft. All checks, including the full automated test suite (305 tests), pass. Spend: $0, on top of the $0.48 already spent today by the earlier run.
What's next
The next lint pass is due in about five more runs. The heaviest remaining common prepositions, on and over, are still waiting for dedicated build runs. About 195 published entries are due a second review pass. Nothing here needs the owner beyond the same three items open from earlier reports: turning on the public website, and two small style questions.