An AI that hands you the answer makes the session feel good and leaves less behind. That effect has been measured, and it shaped everything the Tutor is allowed to do.The AI Tutor works with you while an item is open. The AI Validator, which grades what you have already written, is a separate thing with its own page.
In a school trial, students given an unguarded chatbot did better while they had it and worse than the control group once it was taken away. A guarded version that withheld answers removed that harm. The lesson we took is that an AI helper needs a shape, and that the shape is mostly about who decides when the answer appears.
That study is about high-school mathematics with a chat window, not about this product. It tells us which failure to design against. It does not tell us we avoided it.
The smallest useful help first
Tutoring gains in the classic studies tracked what the learner constructed, not what the tutor explained. So the Tutor asks you to do something before it tells you anything, and it picks that something from what you asked for.Stuck on the answer starts with a recall prompt or one focused hint, then waits for your retry. A real why question gets a direct grounded explanation, because turning a genuine question into a riddle is not scaffolding. A wrong idea it can identify gets a contrast between what you produced and what the rule does, plus a chance to repair it. Story language above your level gets a short explanation in your configured language and no exercise attached.Examples, another hint and a comprehension check stay available as things you choose. The staged prompt-hint-answer structure comes from the dialogue-tutor literature; the application owns the order so the model cannot skip to the end.
You control the reveal
The answer policy you set decides whether the Tutor may show a worked answer, and on the guarded setting it asks first. Hints exist because learners click past them, not because withholding is virtuous: the help-seeking research documents both help avoidance and hint-clicking, and neither produces learning.When the AI service is unavailable or your Tutor budget is spent, a free static explanation is offered under the same reveal policy. A limit changes who wrote the help, never whether the answer is held back.
Why help marks the item assisted
An item you got help with cannot also be the check that you did not need help. So it is recorded as supported practice: it counts as study time and exposure, and it does not move your rule states.
This is not a deduction. Nothing is taken away from you. It is a statement about what that particular answer can prove, and it comes with the replacement described below.
The check that comes later
After supported work, the product stores a source-distinct opportunity on the same skill: another content lineage, a changed source context, a comparable difficulty. It becomes available after at least 24 hours and stays open for seven days.A scorable check comes from audited deterministic material and can update your profile through the ordinary independent path. Unscored practice is a retrieval or self-explanation prompt: worth doing, and it never touches mastery or practice accuracy. The product says which one you are looking at before you start it.
The 24-hour delay and the seven-day window are our product heuristics. The second-language spacing meta-analysis supports spacing and longer gaps for delayed tests. It does not establish these two numbers for this product, and we will change them if the usage data says to.
You can find a due item on the Learning landing page and on your Overview. Remind me schedules exactly one in-app notification at the due time, under the notification preference you already have. It is not a new channel and it does not repeat.Opening a stored opportunity costs no AI credit. If you ask the Tutor for help inside one, that is a separate budget and it is shown and confirmed before anything is reserved.
What the evidence actually supports
Human tutoring and intelligent tutoring systems both post real average effects in the reviewed literature, which is why building a tutor is a reasonable thing to do. A structured AI tutor in one university physics course also reported strong results with authoritative solutions and controlled sequencing, which is promising for this shape.None of that is evidence about Infinite Story. The measure we will watch is the delayed independent opportunity: whether learners who used the Tutor complete a later unassisted, source-distinct check correctly. In-session success is the easy number and the wrong one.
That measurement has not run yet. Until it does, this page describes a design and its reasons, and claims no learning outcome for the product.
What is sent, and what is kept
Before you confirm a Tutor budget the product names the provider it will call, the bounded material it will send and how long that provider keeps it, with a link to the full privacy explanation . The Tutor receives your request, the exact material you were served and only the profile signals its route allows. It does not receive your balance, your account or your history.Tutor message text is kept for 30 days and then removed, and the panel shows that date. Your submitted follow-up responses are kept for 90 days. Both are included in your Grammar profile export while they exist, and a profile reset deletes them. Logs and analytics never contain the message text.Every Tutor response carries a report action for an explanation that is wrong, reveals too early, arrives in the wrong language or has nothing to do with the item.
Tutor help inside a Rich check-up
A Rich check-up is an assessment, so any Tutor use inside it makes the whole run diagnostic-only: it can no longer confirm or change your grammar learning level. You keep the full report, and the price of the Rich run does not change.
That consequence is shown and acknowledged before the Tutor budget is confirmed, not discovered afterwards. If you want the level outcome, finish the run first and ask afterwards.
Getting the most out of it
Try the item before opening the Tutor. The first hint is written for someone who has already been wrong once.
Ask the actual question. "Why is it dative here" gets an explanation; "help" gets a hint you may not need.
Do the later check when it turns up. It is the only part of the exchange that can move your profile.
Report a response that is wrong or in the wrong language. That is the only signal we get.
Inside a Rich check-up, decide about the level outcome before you ask for help, not after.
Under the hood
Each Tutor turn is a fresh bounded call. The model fills a stage of a server-owned route under structured validation; it does not choose the route, the stage order or whether an answer may be shown.Grounding defaults to your course material plus bounded general language knowledge, and can be restricted to course-only. Explanations default to your native language with examples in the target language.Any help that could affect the open answer marks that attempt assisted before the response is rendered, so the marking cannot depend on whether you read it.A failed AI operation consumes no response budget, and a provider retry inside one priced operation is never charged twice.The stored follow-up carries the skill, the response mode where one applies and a difficulty band, and it is drawn from a different content lineage than the item that created it.