Weekly digest, 2026-10-24 to 2026-10-31
This is the second weekly digest under your ruling E3 of 2026-10-24: a short summary of what the dictionary did over the past week, ten entries worth your looking at, and the things that still wait on you. It is written by the scheduled run of 2026-10-31.
The week in numbers
- Entries built: 632 new entries. Entries re-reviewed (existing entries read again by both reviewers): 206. The dictionary now has 3,389 entries, every one marked
reviewed(it has been through the full checking pipeline). The queue of words still to write holds 5,383 rows. - Milestone 1 is finished (2026-10-27): every word in the defining vocabulary (the roughly 2,300 words definitions are written in) now has its entry. Building has moved on to band-2 words and to closure (words the dictionary already uses but has no entry for), limited to bands 1 and 2 under your ruling A3.
- Reviewer decisions: every issue a reviewer raises is decided as applied (the reviewer was right and the entry was fixed), rejected (wrong, with a reason), or escalated (left for you). This week: 4,544 applied, 1,192 rejected, 2 escalated; 5.42 real fixes per entry opened, against 1.64 the week before.
- The two reviewers changed this week under your rulings of 2026-10-25.
reviewer-ais now a Claude model run through the command line at no charge to the budget; it was right 79 percent of the time and produced 3.83 accepted fixes per entry.reviewer-bis now Gemini 4 Argon; it was also right 79 percent of the time, with 1.59 accepted fixes per entry (up from 0.13). Two kinds of flag are switched off because they were mostly wrong: reviewer A on inflections (2026-10-28) and reviewer B on phrases (2026-10-30). - Spend: about US$4 to US$5 a day since 2026-10-26, under the US$15 daily cap; 2026-10-24 and 2026-10-25 were higher (US$11.36 and US$14.55) because of the model trials.
- Ruling E5: the first coverage block (band-1 entries read again specifically for missing senses) is done: 117 entries by 2026-10-30, against the roughly 100 you asked for.
Ten entries for you to sample (ruling E4)
Six are picked at random; four are entries where the drafting session overruled both reviewers on a question of meaning or usage.
- bruise-v (random)
- eighth-num (random)
- fifth-num (random)
- frost-n (random)
- head-n (random)
- luck-n (random)
- america-n (both overruled)
- bit-n (both overruled)
- catch-v (both overruled)
- clean-up-phrv (both overruled)
The calls made against both reviewers (41)
Most of the 41 fall into four patterns, each a house rule rather than a one-off judgment:
- The keyword rule (an idiom goes under its first noun, else its first verb): both reviewers asked for read between the lines under
read-v, change of heart underheart-n, let's face it underface-v, turn of phrase underphrase-n, and similar. They were placed underline-n,change-n,let-v,turn-n. - Unrelated homographs are separate entries: the computing bit, the animal pound, and the side-of-the-head temple were refused as senses and queued as
bit-n-2,pound-n-2,temple-n-2. - Closed vocabularies: both called usually passive "not a grammar code" (
catch-v,grade-v); it is a value in the closed verb-pattern list. Both objected to offensive as a label oncrazy-adj; it is a closed register value. - The number-entry fence: ordinals (
sixth-num,ninth-num) keep their date sense and no adverb sense, as the house template says.
Two that turned on judgment: america-n, where both reviewers objected to the explanation of the Americas, now carries that use in the explanation; and east-n, where the East as a region needs continent names in its definition, which waits on your open question below.
What needs you
- Continent names in definitions (open since 2026-10-10). Working assumption: they go in the explanation, never the definition.
east-nandeurope-nfollow it. - The former, the latter, the following standing alone: noun or pronoun? (open since 2026-10-11). Working assumption: pronoun.
- Names after a title (Ms. Kim, Dr. Lee) in
teacher-n,title-n,dr-abbr,doctor-n. Working assumption: allowed there until you rule. - The charter line in
PROJECT.mdstill says "two non-Anthropic reviewers"; since your 2026-10-25 ruling reviewer A is Claude. The line is yours to edit. - Reviewer A's pronunciation flags stay off by your 2026-10-05 ruling, which rested on the old model's precision (0.26). The new model caught all four planted transcription errors in the trial; switching it back on is your call.
- Pronunciation disputes in
reviews/needs_curator.txt:several-det(American),medicine-nandmilitary-adj(British),outside-n(noun stress),part-time-adj, and the newensure-vandinsure-v, which made opposite British choices; one ruling would settle both.
The other curator lines only note merged branches this tool set cannot delete.