The method

Study with evidence, not with a feeling of progress.

For most students the exam is the first trustworthy measurement of what they know, and it arrives too late to repair what was missing. This page describes the method NodeStudy uses to move that measurement earlier: what counts as evidence of learning, why locating the step where the solution went off course matters more than the score, and how a mistake becomes a scheduled review instead of disappearing.

1. Fluency is not mastery

Rereading notes, following a worked solution and watching a lecture all produce fluency: the material becomes familiar and the sense that you understood it climbs quickly. That feeling is a poor predictor of performance, because an exam does not ask for recognition, it asks for production. They are different operations, and training the first does not train the second.

The practical test: if you close the material and cannot rebuild the argument from scratch on a blank page, you recognise the content but do not yet own it. The useful part is that this distinction is measurable, and it can be measured early.

2. Active recall is what produces evidence

Active recall means producing the answer with the material out of sight. It is the same operation the exam asks for, which makes it the only cheap way to find out today what you would otherwise find out on exam day. In NodeStudy this is the tutor's default: it explains from your material and then asks for your attempt, instead of handing over the solution and moving on.

3. Where you went wrong matters more than how often

Knowing that you missed 6 of 10 questions says very little. A wrong final answer can come from an arithmetic slip, from a rule applied halfway, or from a concept you understood backwards: and each of those calls for a different repair.

The first invalid step in an attempt is more informative than any of them, for two reasons. It identifies the misconception that has to be taught again, and it is the only point where a correction still changes the rest of the solution: everything after it is already contaminated.

It is the difference between “you got the derivative wrong” and “you applied the chain rule to the outer function and never multiplied by the derivative of the inner one”. The second sentence is teachable; the first only reports the score.

4. The error comes back before the exam, not during it

A mistake that is identified once and never met again comes back on the exam. So the diagnosis does not end with the explanation: the error goes into the error notebook with the step that failed and the misconception behind it, and the notebook feeds the spaced review queue.

The interval between reviews is computed by FSRS, a spaced-repetition algorithm that fits the next encounter to how hard that item was for you, rather than applying one fixed interval to everyone. An item you struggled to recall comes back sooner; one that came instantly comes back much later. And an entry only counts as resolved after you reproduce the correction unaided more than once, not the first time you read the explanation.

5. The plan respects the exam date

With evidence per topic and a review queue, planning stops being a generic timetable. The study plan is built from what your own evidence leaves open and the time left before the assessment, and it can be re-paced when a week does not go as planned: sessions you already completed are never lost in the re-plan.

The full loop

1. Your material goes inPDFs, slides, problem sets and past papers. Answers cite the passage they came from, with file and page.
2. The explanation is groundedThe tutor answers from the retrieved material, and says when the material does not cover the question instead of filling the gap.
3. You attempt itYour own solution, without looking. It is the only step in the loop that produces evidence.
4. The first invalid step is locatedNot the score of the attempt: the step where the solution left the road, and the misconception that produced it.
5. The error enters the notebookWith the step that failed, so it can be met again later rather than only corrected in the moment.
6. The review returns at the right intervalThe spaced queue brings the item back before you forget it, at the interval your own difficulty computed.
7. The plan adjusts to the examWhatever still has no evidence moves up the queue while there is still time.

Join the waitlistSee plans →

Why the verification is done in code

A fluent answer and a fabricated one look exactly alike at the moment you read them. That is why in NodeStudy the model proposes and code verifies: arithmetic, derivations, stoichiometry, units, complexity, truth tables and numeric results have deterministic verifiers available. Inspect which checks support an answer. The presence of these tools does not mean every claim has been verified.

The part that matters most is the refusal. A legal-norm verifier that cannot find the article says so instead of completing it by analogy; an essay rubric refuses to invent a score. A synthetic 160/200 would be treated as a measurement and planned around: which is precisely what this method exists to prevent.

Which verifiers serve which exams is listed under exams and olympiads.

How the tutor answers: four help formats and a router

The same question can ask for different things: a full explanation, a nudge at the step where you got stuck, or just confirmation that your answer is right. So the tutor does not have one way of answering. It has four help formats, which are four answers to the question “how much of this should I get ready-made?”:

• Direct. A clear, complete answer, straight to the point, with the reasoning and the sources. For when you ask for the answer.
• Step by step. The solution built in small steps, with explanations and checks where applicable. For when the calculation or derivation is long.
• Socratic. Questions that lead you to the answer instead of the answer. For when you want to leave knowing how to do it, not only what it is.
• Practical. Examples and exercises first, theory after. For people who learn by doing.

The default is Adaptive: a router reads your request and what you have already shown you know, and picks the format for that one answer. It explains when you ask for an explanation, corrects the first wrong step when you show your attempt, gives the answer with its justification when you ask for the answer, and proposes a new task when you ask for practice. If you prefer, you pin a format for the conversation in its settings, and the footer of every answer says which format was used.

Auto: the right model for each question

NodeStudy is not an AI model; it is the structure around one, and that structure chooses the model. In Auto mode every question is read before it is answered: how hard it is, what field it is in (a real-analysis proof and a SAT essay do not ask for the same thing), and whether there is an attempt of yours to correct. With that, the router picks, among the models in your plan, the one that answers that question best, and spends more reasoning on the hard ones and less on the simple ones, so you do not wait the same time for a definition and for a proof.

Three rules always hold. It never costs more than the models in your plan: Auto chooses within what you already have. You know who answered: the footer of every answer names the model that produced it, and says when the model you asked for was unavailable and another answered instead. And verification does not depend on the model: calculations, derivations and citations go through the same code verifiers whichever model wrote the answer. If you prefer one model, pick it in the selector and it stays fixed.

What this method is not

• A summary is a study aid. The tutor can generate summaries and documents, but reading an explanation does not show that you can solve the problem yourself. Use them alongside practice and review.
• Not a shortcut to the answer. Asking for the finished solution works for handing in the problem set and leaves no evidence behind.
• Not a measure of effort. Minutes connected, number of prompts and PDFs opened are not learning, and the product does not treat them as if they were.

Related research behind the method

Roediger & Karpicke (2006), Test-Enhanced Learning: experiments on retrieval practice and delayed retention.

FSRS · Free Spaced Repetition Scheduler: open implementation of the review scheduler used in the product.

Settles & Meeder (2016), A Trainable Spaced Repetition Model for Language Learning: related reading on memory modelling; this is not the FSRS algorithm.

These works inform mechanisms. NodeStudy’s conversational loop is a product adaptation, not an intervention validated by these studies.

Frequently asked