Why you are asked, not told.

What you actually want.

Not to understand the lesson while you are reading it. Understanding it while reading is easy, and it is not what you came for. You want to still have it in six months, in a situation that does not look like the lesson — in a conversation, in an exam, in front of the actual problem.

What you run into.

The obvious way to get there is to go over the material until it sticks: read the lesson, read it again, highlight it, come back and read it once more. It is what most people do, and it is what most people would tell you to do.

It fails, and it fails in a way that is designed to be invisible. Rereading reliably increases how fluent the material feels — how quickly it goes down, how familiar it looks, how confident you are that you have it. It does much less to how much of it you can produce later, from nothing, without the page in front of you.

The problem is that fluency is exactly the signal you use to judge whether you know something. So the method that works less well is also the one that most convinces you it is working. You do not discover the gap while studying. You discover it at the moment you needed the knowledge, which is the one moment it is too late to fix.

What we do instead.

Being asked to produce something from memory is not a way of checking whether you learned it. It is one of the things that causes you to learn it. Pulling a fact back out changes how retrievable that fact is next time, more than spending the same minutes putting it back in. This is one of the better-replicated results in the study of memory, and it is the single assumption the app is built on.

A second result narrows it further: producing an answer beats picking one out of a list. Recognising the right option can succeed on a trace far too weak to have generated it, which is useful information about familiarity and poor information about recall.

How that shows up in the app

• Answering is the default, not the reward. A session is a sequence of things to produce. The written lesson stays available the whole time, but reading it is not what advances your schedule — answering is.
• Recognition is capped, in code. Multiple-choice is offered to at most 30% of the slots in a session. It is a hard bound on the offer set rather than a preference in the mix: once spent, the type is dropped from the choices for the rest of the session (MCQ_CAP, sessions/grid.rs), and there is a test that fails the build if a generated session exceeds it.
• Most answers are graded on your device. Typed and constructed answers are checked locally, instantly, with no AI call — which is what makes producing an answer cheap enough to be the default. Only open-ended writing and code review need a model to read them.
• A wrong answer is not a dead end. The correction arrives immediately with the reason, and the point moves up the review schedule. Missing something you genuinely attempted is a normal and productive event here.

The exercise types themselves, and what each one is testing, are documented on exercise types. How a given session gets assembled is on exercise sessions.

What this does not buy.

Four things, and they matter as much as the rest of the page.

• It does not tell you what to practise. Retrieval says answering beats rereading. It says nothing about which point to ask about, or when. That is a different mechanism, and it has its own page.
• It does not replace being taught. You cannot retrieve something you never encoded. An app that only tested you would be useless on new material — which is why every course is written out as lessons before any exercise is generated from it.
• It does not mean difficulty is good in itself. The benefit comes from a successful effortful retrieval, not from struggle as such. An exercise that is simply too hard, or unclear, or asks for something the lesson never covered, is a defect — not a desirable difficulty.
• The 30% cap is a convention, not a finding. The direction is supported — recognition should not dominate a session. The specific number is a line we drew. No study says 30%, and we would rather tell you that than imply a precision we do not have.

Questions.

Sources.

What each one is doing here, so you can tell which claim rests on which work. Every one of these is a published paper you can look up.

• Roediger, H. L., & Karpicke, J. D. (2006). Test-enhanced learning: Taking memory tests improves long-term retention. Psychological Science, 17(3), 249–255.

  The core result this page rests on: being tested on material produces better long-term retention than restudying it for the same time.

• Karpicke, J. D., & Roediger, H. L. (2008). The critical importance of retrieval for learning. Science, 319(5865), 966–968.

  Retrieval is what produces the retention, not the extra exposure that comes with it — the finding that makes answering the app's default verb rather than a way of checking up on you.

• Slamecka, N. J., & Graf, P. (1978). The generation effect: Delineation of a phenomenon. Journal of Experimental Psychology: Human Learning and Memory, 4(6), 592–604.

  Producing an answer beats recognising one. This is why multiple choice is capped rather than merely discouraged.

• Koriat, A. (1997). Monitoring one's own knowledge during study. Journal of Experimental Psychology: General, 126(4), 349–370. · Dunlosky, J., & Metcalfe, J. (2009). Metacognition. Sage.

  Why the failure is invisible from the inside: learners judge what they know from how fluent the material feels, and rereading raises fluency without raising recall.

You will find effect sizes in these papers. We deliberately do not print them on this page: an average across studies does not predict what any one person will get, and a number quoted without its conditions reads as more precise than it is. In the papers they come with those conditions.