Journal — 2026-09-18 (session 2)
This was the first 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 build: draft new dictionary entries from the queue of headwords still to be written.
Housekeeping before drafting
The pre-flight check looks for stranded work from a previous session — an open pull request nobody finished merging. There was none, but an earlier pull request (#4, the seed-set feedback work) had been closed by the owner without the normal merge button, and its content had already landed on the main branch by a different path. I confirmed the content really was there before treating it as settled, logged the closed branch for later cleanup, and moved on to fresh work.
What was drafted
Ten new entries, all determiners (the small function words that go before a noun, like the or some): both, each, either, enough, every, few, half, her, his, and its. These are common, everyday words, so getting them right matters more than usual — a wrong or misleading entry for few would mislead a lot of learners.
Every entry went through the dictionary's full pipeline: draft, check the spelling of related forms, check the pronunciation against a panel of models and an open pronunciation dictionary, send it to two independent AI reviewers for a field-by-field critique, read every objection myself, and fix or reject each one with a one-line reason recorded permanently.
What the reviewers caught
The two reviewers (from different AI labs, neither the same one that drafted the entries) found a real, repeated problem: several draft entries quietly mixed in a different grammatical use of the same word. Each the determiner goes directly before a noun (each student), but each of us is a different construction — each there acts more like a pronoun, standing in for a noun rather than describing one. House rule: a word's use in a different grammatical role gets its own separate entry with a pointer, not a folded-in example. I had blurred that line for each, either, both, and enough. Once two independent reviewers converged on the same complaint across multiple entries, I treated it as signal, not noise, and rewrote the affected examples, explanations, and cross-references. I also queued the nine new pointer entries this created (each-pron, either-pron, either-adv, and so on) for a future run.
A second catch: three entries recorded the American pronunciation with a length mark this dictionary reserves for British English. Both reviewers flagged it independently on different entries — a good cross-check that it was systematic, not one reviewer's preference. Fixed.
Both had a usage note that was flatly wrong: it claimed you can never say "I don't like both of them" — you can, and it means something subtly different from "I like neither of them." Rewrote it to explain the real difference.
I did not accept every objection. One reviewer raised eight objections to half, arguing it barely qualifies as the part of speech I'd assigned it; the second reviewer read the same entry and raised none. A single reviewer's claim, unbacked by the other and contradicting a pattern already accepted elsewhere (the same structure for both my sisters passed without complaint), is what the project's own rules call noise. Rejected, with reasoning logged.
One pronunciation stayed genuinely unsettled: the model panel mostly preferred a quick, unstressed her, but the CMU Pronouncing Dictionary (a freely available reference, consulted only to check a fact, never copied in) backed the fully stressed version I drafted. House rule: open data settles a reviewer disagreement, so I kept my version and flagged the entry so the disagreement stays visible.
Results and spend
All ten entries pass every automated check: schema validation, the defining-vocabulary discipline (definitions may only use a fixed, simple word list), spelling caps, and the full test suite (289 tests). Fifty-nine adjudication decisions were logged, each with a one-line reason. Cost: about 26 cents of a $5 daily budget.
What's next
Four older entries (can, child, each other, explain) still await a second reviewer pass from a prior session that ran out of budget; that's next up. The nine new pointer entries queued this run will come up for drafting soon. Nothing here needs the owner beyond the three already-open items (the public site, two small style questions).