Journal — 2026-09-19
This was the third 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) chose review: a second reading of entries already in the dictionary, rather than drafting new ones.
Housekeeping before reviewing
The pre-flight check looks for stranded work: an unmerged pull request, or a leftover git branch nobody cleaned up. There was no open pull request. One branch from the previous run (claude/admiring-rubin-6urxtt) was still sitting on the server after its pull request merged; I confirmed everything in it was already safely folded into the main branch and logged it for the owner to eventually delete (this session's tools cannot delete a branch themselves — four such branches are now queued for the owner).
The four orphaned entries
Four entries — can, child, each other, explain — have sat as unfinished drafts since the founding session, each having received only one of the two independent AI reviews the pipeline requires because the founding session ran out of its daily budget. I sent all four to the missing second reviewer, read every objection, and fixed or rejected each one.
The most substantial catch was on can: the reviewer pointed out that the entry was missing a whole common use of the word — the negative can't used to say something is almost certainly false ("That can't be true") — and that the entry's claim "the past of can is could" was too simple, since English actually needs "was able to" for one specific completed action rather than a general ability. I added the missing sense and rewrote the affected passages. On child, a "product of a time or place" sense had been worded so broadly it would have covered almost anything; narrowed it. On each other, a set of example word-groupings had been labeled as if the word were a noun, when it is a pronoun; relabeled to the correct category. All four are now fully reviewed and published.
Eight entries brought up to the current house style
Separately, 42 entries written before the dictionary adopted its current rule for marking words-used-as-words (in bold) and quoted example phrases (in italics) are flagged as owing that update. This run brought eight of them fully current — favorite, information, main, police, scissors, advice, afraid, asleep — giving each entry a fresh reviewer pass at the same time, since the prose was already open for editing.
Two real problems turned up along the way. Information's "help desk" meaning (as in "ask at Information") had been filed as a minor variant of the main uncountable sense, but it is really its own countable sense with its own grammar; split out properly. Asleep's etymology note gave the wrong Old English source word and overreached by claiming the history explained a modern grammar rule; corrected, and I also noticed and fixed a related overstatement — the entry said words like asleep and afraid "never" come before a noun, when in fact they can follow a noun with a further phrase attached ("the baby asleep in the crib"); this is now stated the same way in both entries.
I did not accept every reviewer objection. A few were simply wrong: one claimed a legitimate grammar code wasn't allowed, when it's plainly in the dictionary's own approved list; another objected to a respelling that was already correctly formatted; a third insisted "an afraid child" is acceptable English, which contradicts both general usage and this dictionary's own recorded rule for that exact word. Each rejection is logged with its reason.
Results and spend
All twelve touched entries pass every automated check: schema validation, the defining-vocabulary discipline, formatting caps, cross-references, and the full test suite (289 tests). 108 adjudication decisions were logged this run (101 applied, 7 rejected). Reviewer precision (how often an objection held up once I checked it) was strong today: 0.875 and 0.944. Cost: about 33 cents of the $5 daily budget.
What's next
34 of the original 42 markup-pending entries remain, mostly the dictionary's largest and most heavily used words (go, get, make, take, see, say, and similar), which is why they were left for a later run. More importantly, the mode selector now calls for a lint pass — a broader housekeeping check across the whole dictionary — since five runs have gone by since the last one; that should be the very next run's job. Nothing here needs the owner beyond the items already open: enabling the public site, two small style questions, and the branch-deletion and pronunciation items sitting in the curator queue.