A real ML model, not a chat log
Most "AI tutor" demos just chat with the student and hope something sticks. Mentora instead runs every answer through Bayesian Knowledge Tracing (Corbett & Anderson, 1994) — a probabilistic latent-variable model that turns a stream of right/wrong answers into a calibrated probability the student has actually learned a concept. It's the same class of model used in production tutoring systems like Carnegie Learning and ALEKS.
Four numbers describe a learner
Each concept gets its own tracer with four parameters. None of this is prompt engineering — it's a small, closed-form statistical model.
Prior probability the student already knows the concept before any instruction.
Probability of transitioning from "not known" to "known" after one learning opportunity.
Probability of answering incorrectly despite knowing the concept — a careless mistake.
Probability of answering correctly despite not knowing the concept — a lucky guess.
The Bayes update, in full
On every observed answer, Mentora computes a posterior mastery estimate P(Lt), then — only if the answer was correct — applies the learning transition on top of it.
Why gate the learning transition on correctness? Classic BKT applies pLearn after every observation, reasoning that any attempt is a learning opportunity. In practice, with a non-trivial pLearn, that unconditional bump can swamp the evidence step — a wrong answer could still net-increase mastery, which looks broken live: mastery should visibly drop when you get something wrong. Gating pLearn to fire only on correct answers keeps wrong answers honest.
Live simulator
Drag the parameters, then feed the tracer correct or incorrect answers and watch the posterior mastery estimate move in real time — the exact same engine (bkt.js) that runs during a real Mentora lesson.