Score the word, not the sentence
The review system gets a redesign born from two honest admissions: Severo can't really know when you've internalized a word (an LLM grading whole sentences is subjective), and updating every word in a long text is so heavy it bugs out, sometimes rescoring every word ever seen and making the mastery graph spike and crash. The new plan isolates measurement: plain word-translation drills where speed is the signal, translate it fast and the algorithm marks it interiorized, echoing the three-times-in-two-seconds formula. What survives of the LLM: multi-word concepts (ne...pas, the three characters that mean train station) that naive splitting would destroy, plus anti-cheat details like per-letter timing and gentler penalties for typos.
// trace: where this idea came from
- ↳ video diary @ 19:02 (el bug que recalifica todas las palabras)
- ↳ video diary @ 22:24 (traducir rápido es haberla interiorizado)
- ↳ Entry 219-2: Count the words, not the feeling (la fórmula de las tres veces en dos segundos, ahora implementación)
The redesign starts with two admissions against his own architecture. First, mastery is unobservable from where Severo currently stands: an agent grading whole exchanges can note your response time, but calling a word internalized stays subjective ▸ 18:27. Second, the bookkeeping doesn’t scale: with longer texts the agent must rescore twenty or thirty words per turn ▸ 18:45, and it visibly bugs out, sometimes deciding to update every word he has ever seen ▸ 19:02, which is why the mastery graph climbs and then collapses in ways that can’t be true of a human memory ▸ 19:18.
The fix is isolation, the same instinct as counting words instead of feelings: retire the review button nobody uses and replace it with bare word-translation drills where the measured thing is speed. Translate it fast and you’ve interiorized it; hesitate and you haven’t ▸ 22:24. The edge cases get designed on camera: long words take longer to type, so maybe clock from first keystroke; typing fast then stalling mid-word is a tell, so per-letter gaps matter more than totals ▸ 23:05; and rushed fingers produce typos, so near-misses shouldn’t be punished like ignorance ▸ 23:57.
la velocidad de la traducción es la nota →
What the LLM keeps is the thing dumb splitting destroys: concepts. Ne…pas is one idea wearing two words ▸ 19:47; the three Chinese characters of train station shatter into fire, station, nonsense if scored separately ▸ 20:24, and Language Reactor’s concept-grouping suggests dictionaries might do it without AI at all ▸ 21:17. His own summary of the stakes: get this right and Severo turns much more serious as a learning instrument ▸ 24:17. Measurement first, magic second…