Each person gets back what they got wrong

You set the pool. Inside it, each person's own missed questions come back on a schedule of their own. Those are two separate layers, and the second never overrides the first.

The first layer is yours. A team lead assigns topics and skills; a tutor sets a class plan for the term. That decides what the group meets at all. Nothing in the schedule narrows that pool without telling you, or narrows it to what somebody already answers well.

One thing drives the second layer, and only one: how that person answered. A question they got wrong comes back soon; a question they got right can come back later and less often. The first time a missed question returns it is usually the same question, word for word. That is deliberate. Re-testing the identical item and re-testing a reworded version with the same answer come out about level in the literature, while a different question about the same idea is the weakest of the three and not reliably better than nothing at all. We do what the evidence supports rather than what looks cleverer.

Your record belongs to you, not to the place you answered. Answering on the web, in a direct message from the app, or from inside a coding agent moves the same schedule, because it is the same person answering.

No model chooses the question and no model grades the answer. The schedule is arithmetic written by people: a fixed set of intervals, a rule for moving up them, a rule for moving down. You could work any single decision out on paper, and if you ask us why somebody got a particular question on a particular day, we can tell you.

Nobody's answers are ever sent to a language model, on any path, ever. No training, no retrieval, no third-party inference. It is not a setting and it is not a tier. The Privacy Policy carries the same commitment, and Security covers where your data lives and who can reach it.

A reviewer can check a rule, which is the thing they are actually being asked to do. Every decision here can be reconstructed after the fact, because there is nothing in it but rules somebody wrote down.

The answer differs between the workplace product and the school product, so here they are separately. Do not apply the answer for one product to the other.

At work

An admin configures the rep and never sees anybody's answers, accuracy or skill signal, on any plan. There is no setting for it, because there is no such view to switch on. Only the person who earned them sees their own record, and the channel never learns who answered what. Privacy and the daily rep lifecycle say the same thing.

At a school

A school is the controller of its students' data under a signed agreement, and it can see more, because teaching requires it. What a tutor sees is still bounded: only what they switched on for that class, and only observed counts. How many questions were asked, how many were missed, which parts of the plan the class found hard.

Never a score, a level, a rank or a risk flag about a student. We show any class-wide count covering fewer than three people as “fewer than 3” rather than as a number, so a small class cannot become a report about one identifiable child. Every tracking switch starts off. The terms are in the Data Processing Agreement.

Yes, in one specific way: what comes back, and when. Nothing else about the product changes for a person, and the list of things it deliberately does not do is as much of the design as the list of things it does.

• The schedule never ranks people against each other. Nobody's next question is computed from anybody else's answers.
• It does not decide anything about anybody. No output of the schedule gates access, changes standing, or gets attached to a person as a judgement.
• How long you took is never read as a signal about you. Not as confidence, not as effort, not as engagement, not as mood. The schedule does not look at it.
• A question nobody answered is never counted as a miss. Ignoring a rep on a busy day costs you nothing.

There is also a ceiling. Only part of anybody's day can be questions coming back, and questions they got wrong take that part before questions they got right, so a bad week cannot turn the product into a remediation queue. There is always something new in the day.

Published research on retrieval practice and spaced practice, read with its limits stated. The reading is on Research, with citations, including the parts that did not go our way.

The intervals in any spaced schedule, ours included, are carried over from laboratory work on facts and vocabulary, and there is no randomised trial of retrieval practice on programming. We say so on the research pages rather than presenting the design as settled science. You should trust a page that states its own limits.