Why an adversarial tutor

Being told the answer feels like learning. It isn't.

The problem

Recognition is mistaken for knowledge

Re-reading a chapter produces the warm feeling that you understand it. Then the exam asks you to produce the argument from nothing, and it turns out you could recognise it but never reconstruct it. The gap only shows up when it is expensive.

A general chatbot teaches the wrong course

Ask a general assistant about your subject and you get the internet's average version of it — not your lecturer's notation, not your syllabus's scope, and occasionally not even something true. It also answers instantly and completely, which is precisely the wrong behaviour for someone trying to learn.

Nobody schedules the forgetting

What you understood on Tuesday is mostly gone by the following month unless something brings it back at the right moment. Left to yourself, you revise what feels familiar — which is the material you already know.

Order matters, and it gets lost

Hard subjects have prerequisites. Study tools that shuffle flashcards randomly will happily quiz you on something three concepts deeper than where you actually are, and the failure teaches you nothing except that the subject is hard.

How TrypoLearn answers it

  • Grounded in your own material. Every answer is retrieved from the documents that were actually loaded for your subject — hybrid vector plus full-text search over your syllabus, with citations back to the source.
  • Adversarial by design. The tutor makes you produce the answer instead of handing it over, and it tells you plainly when you are wrong. Comfort is not the goal; being able to do it unaided is.
  • A map, not a pile. Material is decomposed into knowledge components with prerequisite edges — including edges that cross document boundaries — so practice follows the structure of the subject rather than the order files happened to be uploaded in.
  • A human approves the map. The model proposes knowledge components and prerequisites; a teacher reviews, edits and approves them. Automatic extraction sets the pace, human judgement sets the truth.
  • Mastery measured over time. Review scheduling uses FSRS, so an item you got right today comes back exactly when you are about to lose it — and "mastered" means it survived that, not that you nodded along once.
  • Evidence from ordinary conversation. You do not have to sit a quiz for the system to learn about you: mastery signals are also inferred from what you demonstrate in normal chat with the tutor.

    • What this approach costs

    Grounding the tutor in your own material means someone has to load and curate that material, and approving knowledge components is real work for a teacher. A tutor that argues back is also less pleasant than one that agrees with you.

    Those are deliberate trade-offs. The alternative — instant, agreeable, ungrounded answers — is already available everywhere, and it is not what makes anyone competent.