Four decades of second-language-acquisition research converge on four mechanisms: comprehensible input, explicit attention to grammatical form, retrieval practice, and pushed output. Most products pick one and ignore the rest. Here is what we built instead, and why — with the evidence for each.
Stephen Krashen (1985) argued that languages are acquired, not memorized, and that the engine of acquisition is comprehensible input — language one step beyond what you already understand. He called that step i+1. The idea is unromantic and quantitative. To pick up a word from context, you have to understand the context. Read a sentence with one new word and you can usually infer it. Read a sentence with five new words and you can infer nothing; the unknowns alibi each other.
Paul Nation (2006) gave the threshold concrete numbers. To follow a text with help, you need to already know roughly 95% of the words on the page. To read unassisted, the bar climbs to about 98%. Below those thresholds, the text stops teaching and starts costing — every sentence becomes a dictionary errand.
Most learners who feel slow are not lazy. They are reading at 85% coverage and wondering why nothing sticks.
Below 95% coverage, the text stops teaching and starts costing.
The i+1 ceiling is widely cited and almost never enforced. Walk through the popular options and the math falls apart in the same place every time.
LingQ and the catalog-based reading apps serve native text with a click-to-translate layer on top. The catalog is enormous, but it is unleveled — every article was written for fluent adults. A click-to-translate dictionary turns native text into readable text, but readable is not comprehensible. You are decoding, not acquiring.
YouTube channels marketed as comprehensible input run into the same wall from the other side. A single creator writes for a single audience, and any given video either assumes too much vocabulary or covers ground you already knew. There is no way for the creator to know what is in your head, because they are not writing for you.
Textbooks solve the problem by ignoring it. Chapter four is chapter four for every learner who buys the book, whether you have studied for six weeks or six years. The result is content that is too easy for half its readers and too hard for the other half, and dead in either case.
The honest summary is that to find native content sized to your level, you would have to curate it yourself. Curation is most of the work. Most learners give up before they finish that work, and call it a willpower problem.
Native text with a click-to-translate layer isn't comprehensible input. It's just text with a dictionary attached.
The weekly article is generated against your vocabulary. We hold a list of the words you have plausibly encountered — the CEFR pool at your level, plus every word that has appeared on a card in your monthly deck — and we feed that list to the model as a whitelist. On top of the whitelist sits a small, budgeted frontier of new words: the i+1 layer, the words that should enter your reading this week and then turn up in next month's deck.
After the article is generated, we measure it. The body is tokenized, normalized, and checked against the allowed set. If unknown-word density exceeds 5%, we regenerate the article with a stricter prompt that names the offending words as forbidden. Two retries, then we ship and log a warning so we can tune the prompt later. The number on the page is real; we publish it on every sample.
The monthly Anki deck carries the same vocabulary forward into spaced repetition. The vocabulary quiz tests retrieval from that same set. The AI tutor draws from the same set. One vocabulary, shared across reading, retrieval, and output — the channels reinforce each other because they share the same words, not by coincidence but by construction. That is the input pillar. It is necessary, and it is not sufficient.
One vocabulary, shared across every channel. The reinforcement is not a side effect — it is the whole design.
Here is the uncomfortable finding the input-only crowd tends to skip. Give adults a great deal of comprehensible input and their vocabulary grows, but their grammar quietly stalls. In a controlled 2023 study, adults exposed to hundreds of sentences of a new language session after session improved on words and plateaued on grammar — and the learners who did make grammatical progress were the ones paying conscious, explicit attention to the patterns (Pili-Moss et al., 2023). The classroom meta-analyses point the same way: across decades of studies, explicit grammar instruction reliably outperforms hoping the rules will be absorbed from exposure (Norris & Ortega, 2000; Spada & Tomita, 2010; Goo et al., 2015).
This is not a license to drill conjugation tables. Decontextualized rule-cramming transfers poorly to real use. What works is focus on form: a brief explanation of the pattern, then practice where the grammar carries the meaning. We do both. For each structure, Frank Lingo gives you a short, plain-language note, then two kinds of practice — interpretation, where you have to parse the form to get the meaning right (this is VanPatten's structured input, and it transfers to production even without speaking), and production, where you produce the form yourself (Swain's pushed output, which input alone never fully builds).
And we sequence it. Grammar is not a flat list; structures depend on each other. You cannot meaningfully learn the subjunctive after a wish before you can form the present tense. So each language's grammar is a skill tree ordered by prerequisites, and you climb it from your level up — the next thing you are taught is always the next thing you are actually ready for. You can watch the tree fill in as you go.
Input grows your vocabulary and stalls your grammar. Grammar you have to be taught — briefly, then practiced.
Merrill Swain (1985) noticed something Krashen's framework did not predict. French-immersion students in Canada had years of comprehensible input and still hit a ceiling on grammatical accuracy. Input alone was not enough. The students needed to produce the language under pressure — to hunt for the word they did not have, notice the gap, and try again. Swain called this the Output Hypothesis. Output forces you to commit, and committing is what surfaces the holes.
This is the answer to the apparent contradiction on our homepage. Reading comes first, and the tutor is always on. Output without input is roleplay — you are performing a language you do not yet have. Output on top of input is acquisition. Our tutor runs from day one because it draws from the same vocabulary the reading and the deck are giving you. It is not a separate course; it is the production channel on the same vocabulary set.
Output without input is roleplay. Output on top of input is acquisition.
vs. AI conversation apps. Speak, Loora, Univerbal, Google Translate's conversation mode, ChatGPT voice — they all sell the same thing: talk to a bot in a language you do not yet speak. Talking to a bot in a language you do not know is not learning. It is roleplay with a partner who is contractually obligated to be impressed. Output is real practice only when you have a vocabulary to draw from. You cannot have a conversation in a language you have never read. Reading comes first.
vs. catalog-based products like LingQ. Reading native content is the goal, not the method. Until you have the vocabulary, native text is not comprehensible input — it is text with a dictionary attached, and the dictionary is doing all the work. We generate input sized to your level so that reading is actually driving acquisition. Once your vocabulary catches up to native text, LingQ and the rest of the native-content ecosystem are excellent. We are the on-ramp, not the destination.
vs. dead-content textbooks. Babbel, Pimsleur, traditional course series — the same chapter four for every learner since 1985. Static content cannot be i+1 for anyone in particular, because anyone in particular is exactly what it was not written for. We write you a weekly article. You get the next chapter that was written for you.
vs. input-purist apps. The strict comprehensible-input crowd insists you should never be taught grammar at all — just absorb it from exposure. The evidence says adults who do that plateau. We don't only expose you to grammar; we teach it, briefly and in order. Equally, we don't go the textbook route of decontextualized drills divorced from meaning. Focus on form sits in between, and it is where the research lands.
You can't have a conversation in a language you've never read. Reading comes first.
None of this is new. Krashen has been right about input since 1985. Swain has been right about output since the same year. The grammar-instruction findings have been piling up since the 1990s. Nation has been telling us the numbers for two decades. The reason language learning still feels slow is not a missing theory. It is that the theory has been hard to deliver — until it became cheap to write a custom article every week, teach the grammar in the right order, and run all four pillars from a single vocabulary and skill tree we already keep for you.
That is the whole pitch. The science is forty years old. The plumbing is finally here. Try a free month and see whether learning at your actual level feels different than learning at someone else's.