What a closed learning loop actually is
Practice that does not accumulate
A learner can have a pleasant English conversation every week for six months and end up roughly where they started. The conversations are not wasted — talking is better than not talking — but nothing carries from one to the next. Each session opens with the same warm-up, drifts wherever the topic goes, closes with a vague “that was good,” and leaves no trace. The article mistakes made in week two are still there in week twenty, because nothing ever wrote them down.
That is an open loop. Effort goes out, nothing comes back. It is the default shape of most language practice: a conversation club, a friendly tutor without notes, an app that starts every session from a blank slate. The learner is doing the hard part — producing language out loud under time pressure — and the system around them is throwing away the evidence.
A closed loop is the alternative, and it is a much less glamorous thing than it sounds. It is not a smarter conversation. It is bookkeeping.
The five stages, and what each one owes the next
Five stages, in order: the conversation itself; the analysis that reads it; the plan that analysis revises; the homework the plan assigns; and the conversation after that, which loads all of it before it starts. Each stage exists because the next one cannot function without its output.
The conversation produces a transcript. That is its second job; the first is the practice itself. The transcript matters because it is the only honest record of what a learner can actually produce, as opposed to what they can recognize on a multiple-choice screen.
Analysis turns that transcript into decisions. The plan turns those decisions into a sequence. Homework turns the sequence into work that happens between lessons. And the next conversation consumes all of it — which is the step that closes the loop and the step that is easiest to leave out, because a conversation looks fine without it.
Analysis is the stage everyone skips
Analysis is skipped because it is invisible. Nobody buys a language product for its post-session batch job. But it is where a transcript becomes something the next lesson can use, and if it produces prose — “the learner did well, some tense confusion” — then nothing downstream can act on it.
So the output is structured and validated against a schema rather than written as a summary. Concretely, a single session produces: a level estimate carrying a confidence rating and the specific evidence behind it, not just a letter; a list of errors, each with a type, an example in the learner’s own words, and a frequency; the vocabulary the learner reached for and did not have; updates to the review queue, marking each due item recalled or missed and setting its next due date; mastery updates for the curriculum objectives; the plan for the next lesson; the homework, each item with a due date; the report card the learner actually sees; and a short summary written specifically to be read into the next lesson.
That last item is worth dwelling on. The summary is not for the learner. It is a few sentences addressed to the next session, so that the tutor opens the conversation already knowing how the previous one went. It is the handoff note a human teacher would write to themselves, made explicit because a system cannot rely on remembering.
Mastery states, and why they should move slowly
Each curriculum objective — “narrate past events in the past simple,” “manage problems and complaints politely” — carries a state: being introduced, being practiced, or mastered. Analysis moves objectives between those states, and the temptation is to move them fast, because progress feels good.
The rule that prevents this is dull and important: an objective still in the practicing state is never advanced on the strength of a single good session. It takes at least two clean ones. A learner can produce a correct present perfect once by accident, by echoing the tutor, or because the sentence happened to be one they had rehearsed. Twice, in different contexts, is weaker evidence than a formal assessment and much stronger evidence than once.
This is also what stops the plan from drifting upward faster than the learner. A system that advances on every good turn ends up with a B2 label on an A2 speaker, and every lesson after that gets pitched slightly too high — which does not feel like a data problem to the learner, it feels like failure.
The plan carries the memory
The next lesson plan is where the loop’s state lives. It has to contain three things that a generic lesson does not.
First, the review items that are due today — specific past errors and specific vocabulary, scheduled to come back at the right moment rather than whenever the topic happens to allow it. Second, activities themed on what this learner actually said they wanted: someone learning English for work gets workplace role-play, not a lesson about hobbies with a work-shaped example bolted on. Third, a difficulty level consistent with the mastery states, so the lesson is not quietly asking for structures the learner has not been introduced to yet.
None of that is the tutor improvising well. It is the tutor being handed a plan that already encodes what the last session revealed.
The half of the loop that runs while you are away
Homework is the stretch of the loop that operates between conversations, and it is built from the review queue rather than from a shared pool of exercises. Flashcards come from vocabulary this learner reached for and missed. Cloze and transformation exercises are generated against this learner’s own recurring error patterns. Listening and reading material is themed on their stated interests, which is not decoration — interest is what makes an optional exercise get done.
Results flow back the same way everything else does. Completing an exercise updates the review schedule and the mastery state, and the summary of what got done goes into the next lesson’s context. The tutor knows you did the homework. More usefully, it knows which parts you got wrong.
When the loop cannot close in time
Analysis takes a few minutes, and sometimes a learner books two lessons close together. Rather than making the second lesson wait or pretending the analysis happened, the honest behaviour is to degrade visibly: reuse the current lesson plan, and generate the between-session summary from the last stretch of the transcript with a faster call. The loop stays closed, just with less depth on that one iteration.
Systems that have no such fallback tend to fail in the other direction — they produce a confident plan built on nothing, which is worse than a plan that is openly one cycle stale.
How you would check the loop is real
Two tests, both of which a learner can run without any access to the internals.
The first: by the third lesson, two people who started at different levels with different goals should be getting visibly different content. Not different topics inside the same script — different objectives, different difficulty, different scenarios. If the third lesson looks the same for a beginner aiming at a relocation and an intermediate speaker preparing for interviews, the personalization is cosmetic.
The second: within about three sessions, one of your own past mistakes should come back. Not announced as a review drill, but as a natural opportunity to use the structure you got wrong last time. If you never notice your old errors reappearing, nothing recorded them.
Fail either test and what you have is a chatbot with a database attached — pleasant, sometimes useful, but not a course. Passing both is not proof of good teaching, but no amount of good teaching survives failing them.
The unglamorous conclusion
Closing the loop is not a conversation-quality problem. A very good conversation inside an open loop still leaves nothing behind, and a merely adequate conversation inside a closed loop compounds. The engineering that matters here is the boring part: recording what happened, deciding what it means in a form the next step can read, and making sure the next conversation actually loads it.
The method page describes how the curriculum and the report card fit around that loop. And if the premise underneath all of this — that speaking is a separate skill that has to be practiced on its own — is not yet obvious, that is the argument in why speaking is the skill that does not transfer.