Founding session, day 2 (2026-09-17): the seed set is reviewed and published, and the dictionary is ready for its scheduled runs

This is the final report of the founding session of the TKG English Learner's Dictionary. It continues journal/2026-09-16.md, which describes the foundations, the word list, the tools, and the drafting of the seed set. Everything below is in the repository tkgally/eex-dict on main.

What happened today

The reviews. After the UTC day rolled over and the daily budget reset, two reviewer models from different labs (the roles reviewer-a and reviewer-b in config/models.md) read all 82 entries field by field: the 77 seed entries and the five determiners (a, all, an, another, any) that the dry run of the scheduled Routine had drafted. They returned 787 issues. I read every one, decided apply or reject with a one-line reason, and logged 798 decision lines in reviews/decisions.jsonl (some fields carried more than one issue): 367 applied, 431 rejected, none escalated to you. The accepted fixes were made by hand in the entries, every entry was re-validated and re-checked against the defining vocabulary, cross-reference back-links were added, and 78 entries now carry the status reviewed. Four small entries (can, child, each other, explain) were read by only one reviewer before the budget cap stopped the panel with about US$0.01 left; they stay draft and are the first thing the next run finishes.

What the reviewers were right about. Roughly one flag in two from the first reviewer was worth acting on. The real faults were of a few kinds: notes about other languages that claimed too much ("most languages", "English always needs it", "a hand, never hands"); a dozen subsense definitions that repeated their own headword as a prefix (look after yourself: ...); a few misplaced items (smell of under the cause sense of of; take someone to court hiding in the movement sense of take, now a phrase of its own; a radio's pick up merged with noticing a mistake); some overbroad rules for learners ("never", "only", "always"); a handful of wrong or unnatural examples; and two entries whose core-idea line defined an adjective as a feeling. Two of these patterns became rules in the style guide.

What they were wrong about. The second reviewer spent four flags in five on one complaint: a phrase with a single example, which the house style allows. Both reviewers repeatedly flagged house conventions they had been told about: no length marks in American transcriptions, no stress mark on one-syllable words, entry-level labels that cover every sense, grammar patterns limited to the closed list, explanations on verbs such as have and take. The precision by reviewer and by kind of issue is now measured and recorded (wiki/notes/reviewer-precision.md); three kinds fall below the line at which the Routine's lint pass switches them off. The reviewer prompt's list of conventions was extended so the next runs waste less.

The site. docs/ was built from the reviewed entries (78 entries on 77 pages; drafts are withheld) and passed the browser check: 54 checks on phone and desktop widths, including search by any inflected form, hover previews of linked words, the translator's view, dark mode, and no horizontal overflow. Nothing is visible on the web yet; that waits for item 1 below.

The Routine dry run (done yesterday, reported here because its pull request merges with this one): a Sonnet-class model followed routine-prompt.md literally through the unpaid half of a build cycle and produced five valid entries. It exposed five gaps in the prompt (the reading order, the schema files, the convention for recording the drafting model, what a run does when it may not build, and pacing advice), all fixed; the record is wiki/notes/routine-dry-run-v1.md.

Spend

Day 1: US$4.995 of the 5.00 cap (word lists, the pronunciation test, twenty reviews). Day 2: US$4.985 (the reviews of 78 entries by two models, and 17 by one). The ledger is config/budget-ledger.json; every call is recorded with its billed cost. A full two-reviewer read of a substantial entry costs about six cents.

What did not work

The budget cap cut the last four reviews. The order was alphabetical; a cheaper order (small entries first) would have finished. The Routine's selector sets a per-run budget below the cap, so a scheduled run will not hit this.
Reviewer noise: about 430 of 800 decisions were spent rejecting flags that repeat the house style. The extended prompt should help; if not, the lint pass switches those families off.
Nothing else failed today. The adjudication itself took most of the session: reading 787 flags against 82 entries is slow work that no tool shortens.

What needs you

Enable GitHub Pages for tkgally/eex-dict: Settings, Pages, Source: "GitHub Actions". The workflow .github/workflows/pages.yml then publishes docs/ on every merge to main.
Review the seed set. journal/2026-09-16-seed-review-guide.md says what to look at and how; drop your comments as files in inbox/. Anything there outranks the queue for the next run.
Schedule the Routine on Sonnet 5, one to six times a day, once you have looked at the seed set. The trigger prompt is one line: Read routine-prompt.md in the repository root and follow it exactly.
Decide the drafting model. The seed set was drafted in-session by the founding model; scheduled runs will draft with the Routine's own model unless config/models.md names a flagship OpenRouter model for the drafter role (about three to five cents per entry at current prices, on the same US$5 daily cap). The working assumption until you say otherwise: drafting stays in-session.

The baton for the next run is NEXT.md; the questions I answered with working assumptions are in wiki/open-questions.md.

All journal entries