Founding session, 2026-09-16: the dictionary now has its foundations, its word list, its tools, and a first seed set
This is the report for the first day of the founding session of the TKG English Learner's Dictionary. The second day (the reviews of the seed entries, the site, and the final baton) is reported in journal/2026-09-17.md, which also lists what needs you.
What exists now
The repository tkgally/eex-dict holds the charter (PROJECT.md, your decisions restated), the rules every session reads (CLAUDE.md), the knowledge base on the framework you supplied, the entry schema and every closed vocabulary (parts of speech, grammar codes, 29 verb patterns, and the original usage-label sets: 11 registers, 16 regions including "online", 47 domains, 5 currency labels, 6 attitude labels), the style guide, and the tools the pipeline needs, all in the Python standard library with 280 unit tests. Two pull requests have merged: the scaffold and the word lists with the tools.
The defining vocabulary and the queue
Three models from different labs (OpenAI, Google, DeepSeek), asked independently in 37 chunks by semantic field, proposed the words a learner's dictionary needs in order to define everything else. Words proposed by two or more form the core (1,671); I read the 927 single-vote words one by one and accepted 682; then I added function words, number words, the dictionary's own grammar terms (clause, participle, vowel), and a few plainly basic words the chunking had missed. The list holds 2,394 lemmas, and a rule lets transparent derivatives (politely, unkind, successful) count as listed. The same panel then rated candidates into frequency bands; the headword queue holds 4,278 rows (2,606 defining-vocabulary rows in band 1, 1,294 in bands 1 to 3 from the panel, and the rest words the seed entries refer to). No external word list or frequency list was consulted at any point.
The pronunciation test
Four models transcribed 500 words (200 defining-vocabulary words, 100 mid-band, 100 rare, 50 heteronyms, 50 loanwords) in American and British IPA. American transcriptions were scored against the CMU Pronouncing Dictionary, fetched to a temporary directory and discarded. The drafting model was right on 98.5 percent of defining-vocabulary words and 92.6 percent overall; rare words and loanwords are the weak spot (about 80 percent). The rule fixed in advance, that a transcription counts as verified when two of the votes (three panel models plus the CMU dictionary for American) agree with the drafter, reached exactly the required 95 percent precision and was adopted; every real drafter error in the sample (username with a syllable lost, electronic stressed on the last syllable) was caught. Reading the disagreements showed that most were notation conventions, and the comparison was amended for five of them before any verdict was adopted; both sets of numbers are recorded. The British hand check against Wiktionary is in progress at the time of writing.
The seed set
Seventy-seven entries were drafted to exercise everything: the heavy verbs run, take, make, get, go; be, have, do; say / tell, speak / talk, borrow / lend; nouns with idioms (hand has 32 phrases, time 36); grammar traps (information, scissors, police, people); the heteronym record; color with its British variant; shit under the taboo policy; adjectives with position restrictions; discourse adverbs; function words with explanations; modals; phrasal verbs; un- and -ness; ASAP and TV; o'clock. All validate; inflections are in from the rules and the exceptions table; the pronunciation panel has read them. The two-reviewer panel and my adjudication followed on day 2 (see journal/2026-09-17.md), since day 1's spend reached US$4.99 of the 5.00 cap. Three of the entries (run, hand, the) are the style guide's model entries. journal/2026-09-16-seed-review-guide.md tells you what to look at.
What did not work
Two things cost money without result: the first pass of the pronunciation test lost about a third of its replies to an output limit set too low (the models put hidden reasoning tokens against the same limit), and one model's banding replies had to be redone with a more robust format. The fixes (reasoning switched off for transcription and word-list calls, numbered replies) are in the tools. An agent building the inflection tool failed once by trying to emit its whole table at once; the second attempt worked in pieces.
What needs you
The four items are listed at the end of journal/2026-09-17.md: enabling GitHub Pages, reviewing the seed set, scheduling the Routine, and deciding on the drafting model.