The backend got infected

Severo’s backend, Juan says, keeps “getting infected”. The job is to read a reply that mixes the user’s language and the target language and record which words were in the target language, so they can be counted and reviewed. But sometimes it files a word under the wrong language, or stores a stray word from the wrong language entirely, and once that happens the rot spreads ▸ 6:45. A poisoned entry in the vocabulary store poisons the review deck built on it, which poisons every minigame stacked above that, so everything from that point down breaks ▸ 7:16.

una palabra en el idioma equivocado envenena todo lo que viene después →

The diagnosis is load. Right now Severo has to teach the lesson and, in the same breath, decide which fragments of what it just wrote are French. That is a very high cognitive load, like asking someone to hold ten thousand things in mind at once, and LLMs hallucinate under exactly that kind of pressure ▸ 7:48. So the move is to take the job away from the model. A deterministic language-identification framework, Lingua, would look at what Severo wrote and rule “this part is English, this part is French”, and Severo would no longer carry that task at all ▸ 7:58. The general shape is the same one the whole app keeps rediscovering: the model is good at the fuzzy, human part and bad at the crisp, mechanical one, so you hand the crisp part to a program that cannot hallucinate. Make Severo’s life simpler ▸ 8:10, and the errors that compound stop starting…