The lemma is the unit
Stanza graduates from audition to integration, and the reason is accounting: Severo the agent was updating vocabulary itself and kept booking phrases as words (je m'appelle counted whole, then Apple counted again separately), so the seen-count that drives spaced repetition was fiction. Lemmatization fixes the unit, run, ran, running all book to run, conjugation coverage becomes a stat (correr known in five of six tenses), and the dependency data adds a bonus: multi-word entities like The United States of America get spoken as one fluid thing. One honest caveat survives: Stanza is a neural net, not a judge, so if Gemini writes a broken sentence it will happily draw confident arrows over the wreckage.
// trace: where this idea came from
- ↳ video diary @ 6:21 (correr, corrió, corriendo: un solo lema)
- ↳ video diary @ 12:23 (el caveat: Stanza no flaguea errores)
- ↳ Entry 228-2: X-ray for sentences (la exploración que aquí se vuelve integración)
- ↳ Entry 224-2: Score the word, not the sentence (la medición que estos lemas hacen honesta)
The x-ray audition becomes an integration, and the motive is bookkeeping, not beauty. Until now Severo the agent updated vocabulary himself, and an agent is a sloppy accountant: he’d book je m’appelle as one word, then Apple alone as another, so the same knowledge got counted twice or as phrases ▸ 6:59. Since the whole measurement thesis rests on counting how often a user meets each word, a broken unit breaks everything downstream ▸ 7:27.
Stanza, Stanford’s NLP library trained across their whole language list, fixes the unit: the lemma, the root under the surface forms, run, ran, running all booking to run ▸ 6:21. With it come the side dishes: conjugation coverage as a stat, you know correr in five of six tenses, these are missing ▸ 8:24; part-of-speech tags for free ▸ 8:37; and multi-word entities, so The United States of America is recognized as one thing and can be spoken as one fluid phrase instead of five stiff words ▸ 10:00.
correr, corrió, corriendo: una sola palabra →
The honest caveat gets stated before the feature ships: Stanza is a neural net, not a judge. It won’t flag an error, it will parse whatever arrives as best it can ▸ 12:23, so if Gemini 2.5 Flash-Lite emits a scrambled sentence, the arrows still connect, confident and wrong, and a learner who trusts connected arrows inherits the error squared ▸ 13:40. Meanwhile the coin economy gets its edge case resolved in character: what if a user goes broke betting against the sensei? Let the balance go negative, “si está perdiendo más de lo que gana es que no está aprendiendo, y póngase las pilas” ▸ 14:42, softened by Julia’s show-up bonus and a starting allowance ▸ 15:04. The ledger, at last, has a unit…