Journal — 2026-09-20 (run 6)
This was the nineteenth 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 (tools/next_mode.py) forced lint — a housekeeping pass over the whole dictionary and its tools, due every five runs — rather than drafting or reviewing entries.
Housekeeping before linting
The pre-flight check looks for stranded work: an unmerged pull request from a previous run, or a leftover git branch nobody cleaned up. There was neither. The inbox (where the owner would leave instructions) was empty, and the checkout was already current with the shared history, so I moved straight to the lint unit.
The mechanical checks
A handful of scripts sweep the whole dictionary for things a human would otherwise have to notice by eye: word-count limits on the project's own guide files, broken internal links between wiki pages, cross-references between dictionary entries that point the wrong way or are missing their mirror, and words used in a definition that fall outside the dictionary's own approved defining vocabulary. All came back clean. The cross-reference sweep did add six small missing "see also" back-links to existing entries (a normal, ongoing bit of upkeep, not a content problem), and a cleanup pass removed three leftover claim files from runs that never finished drafting the entries they had reserved.
I also re-measured how often each of the two independent AI reviewers' objections turn out to be right once checked (their "precision"). Five categories of objection were switched off in earlier runs for having a track record under 30 percent accuracy over many decisions — a way of muting persistently noisy feedback without ever silencing a reviewer's legitimate catches in other areas. All five stayed switched off at their original figures, confirming the switch mechanism still works as intended, and no new category crossed that threshold this time. Both reviewers' overall accuracy improved slightly since the last measurement.
Finally, I checked the project's own internal documentation for contradictions or drift: the page index still matches every file on disk, the running log has no gaps, and the two files where open questions and pronunciation disputes wait for the owner are unchanged and still current.
Two small tool fixes
Two minor, previously logged software bugs were still on the to-do list and small enough to fix in this pass:
First, when a word's American and British pronunciations are checked by an automated panel and one comes back "disputed" (the panel couldn't agree), the tool correctly flagged it — but if a later run then fixed and re-verified that pronunciation, the flag never went away on its own; someone had to remember to remove it by hand. The tool now clears the flag automatically once every checked pronunciation is settled, while still respecting a case where only one of the two (American or British) is disputed.
Second, when logging a decision about a reviewer's objection, the tool that records it could get confused if a reviewer's objection quoted, say, "the whole sentence" and a separate objection on the same word quoted just "part of" that same sentence — any short phrase specific enough to pick out the second objection was, by definition, also found inside the first, so neither could be selected. The tool now also accepts a simple position number ("the second objection on this list") as a fallback way to say which one is meant.
Both fixes came with new automated tests (eight in total; 305 now pass, up from 297) and are documented in the project's internal notes.
Results and spend
No entries were drafted or changed by hand this run (lint is a tooling and documentation pass, not a content pass), though the automated cross-reference sweep touched six entries' back-links as noted above. All checks — schema validation, formatting caps, links, the full test suite, vocabulary discipline, and cross-references — pass. Spend: essentially nothing (no paid model calls were needed for this kind of work), so the daily budget stands at $1.26 of $5.00, unchanged from earlier today.
What's next
No small tooling fixes remain on the backlog list. The queue now points to: fifteen entries still waiting for their style-guide markup update, seven prepositions ready to be drafted, and a growing list of over 400 words the dictionary already refers to but hasn't defined yet. The next lint pass is due in about five more runs; the next full "originality" spot-check (comparing entries against a sample of web text to catch anything too close to a published source) is due around run 30. Nothing here needs the owner beyond the three items already open: turning on the public website, and two small style questions (whether four grammar-heavy words should be split into separate entries by part of speech, and what color or weight the dictionary's bold/italic word-marking should use).