A living picture of what each learner knows.
Behind every student, Maestro maintains a learner model: per-concept mastery beliefs, updated by every piece of evidence, calibrated by confidence, weighted by the learner's own baselines — and only ever consolidated when independent witnesses agree.
Not a gradebook. A model.
A gradebook stores outcomes. The learner model maintains an interpretation — what the system currently believes about each knowledge component, and how much that belief deserves to be trusted.
Mastery beliefs
For every knowledge component, an evolving probability that the learner has mastered it — never a checkbox. It climbs and dips with evidence, in the spirit of Bayesian knowledge tracing.
Calibration
How well the learner's confidence tracks their correctness. Well-calibrated confidence is a strong sign of real understanding; confident wrongness signals a misconception to confront.
Evidence records
Every Evidence Check writes its full record — the diagnostic band plus the four signals behind it — so any belief can be traced back to the exact evidence that built it.
Personal baselines
The learner's own pace, process, and engagement patterns. Behavioural signals are read against their baseline — not against other learners — and only ever adjust the weight evidence carries.
Beliefs consolidate only when witnesses agree.
A single result is one witness, never proof. Four mechanisms of longitudinal triangulation decide how far any piece of evidence moves a belief.
Prediction versus result
The learner's engagement in a node — practice performance, time, help requests — predicts an expected outcome. A result far above what their own history predicts is held as unconfirmed, pending more evidence.
Gradual belief
A clean check nudges a belief up; it never jumps it to “mastered.” Coherent evidence — practice and check agreeing, confidence calibrated, reasoning quality matching answer quality — moves it strongly. Incoherent evidence moves it weakly.
Survival of re-encounter
Knowledge components resurface in later nodes and as spaced re-checks — a fresh variant over the same tested substance. Real mastery survives; a belief that collapses is revised, and the collapse discredits the earlier result.
The milestone anchor
The controlled-conditions milestone is the strongest witness, and the whole model is calibrated against it. Beliefs that disagree with the milestone are the loudest incoherence the engine knows.
What the model will never do.
A model of a person carries responsibility. These rules are architecture, not policy.
A state, never a label
A low belief is a temporary state to support — not a judgment to pin on someone. The model exists to route help, not to rank people.
Signals support, never accuse
Behavioural signals are never shown as flags and never trigger penalties. The same pattern that might suggest outsourcing also describes an anxious learner — both get the same supportive check-in.
Honest uncertainty is valued
“I'm not sure” is a first-class, penalty-free answer. It routes to support and improves calibration — the model treats honesty as signal, not weakness.
The learner sees the picture
Nothing is modelled in secret. The learner always sees where they stand, why the system believes it, and what the best next step is — transparency is a tenet, not a feature.
See a learner model come to life.
Watch beliefs form, hold, and consolidate as a real learner moves through a course — and how routing decisions trace back to evidence.