Learner model
Studium keeps, for each concept, two separate estimates: how well the learner has shown they know it (mastery), and how likely they are to still recall it now (retention). This page gives the equations exactly as the code applies them, and the rules that keep both estimates honest. You do not need it to use Studium; read it to judge what a mastery figure means.
Mastery: a Beta estimate
Each concept starts with a uniform prior, α = 1 and β = 1. α collects evidence of knowing, β evidence of not knowing. Mastery is the posterior mean:
p = α / (α + β)
A graded answer at difficulty w (between 0 and 1, from the concept's Bloom level, see Scene types) with h hints updates it:
correct: α ← α + (2 − w) / (1 + 2h)
wrong: β ← β + (1 + w)
Hints discount the credit
Each hint shrinks the credit of a correct answer. At w = 0.5, an unhinted correct answer adds 1.5, one hint 0.5, two hints 0.3. Leaning on hints cannot certify a concept as cheaply as answering alone.
Correction for guessing
On a choice item, a learner can be right by luck. With k right options out of n, the guess rate is g = k / n. A wrong answer then also moves the credit a lucky guess would have earned from α to β:
τ = min(α − 1, (2 − w) · g / (1 − g))
wrong: α ← α − τ, β ← β + (1 + w) + τ
For a learner who knows a fraction of the items and guesses the rest, the expected estimate tends to that fraction rather than to the fraction plus lucky guesses. A coin-flipper on true/false items no longer reads 0.5. α never drops below its prior of 1, and correct answers are not changed.
Cap on old evidence
When α + β exceeds 20, both are scaled down by the same factor. The mean is kept, and a new answer always carries at least about 5 % of the total weight, so a learner who stops knowing a concept is not hidden behind a long early streak. Attempt counts are not scaled.
The update is in beta-updater.ts.
When a concept counts as mastered
| Use | Condition |
|---|---|
| Unlock a dependent concept in the graph | each direct prerequisite at p ≥ 0.8 (0.6 in a lenient universe), with at least 3 graded attempts |
| Certify mastery (a mastery claim is written) | p ≥ 0.8, at least 3 graded attempts, not measured as overconfident, not measured as over-trusting |
| Count as "due" in the study plan | p ≥ 0.6 and the review is due |
The three-attempt floor means two lucky answers can let the learner move on but never certify the concept.
Retention: FSRS
Each graded answer, and each self-rating, is a review of the concept in a scheduler derived from FSRS-4.5 (Free Spaced Repetition Scheduler) with its published default weights. The memory of a concept has a stability S, in days, and a difficulty between 1 and 10. Retrievability, the probability of recall t days after the last review, is:
R(t) = (1 + (19/81) · t / S) ^ (−0.5)
A concept is due when R falls to the target retention of 0.9; with these constants that happens after S days. A successful review increases S, more so when it comes late (the spacing effect); a lapse resets it to a lower value, never a higher one.
| Event | FSRS grade |
|---|---|
| Graded answer, wrong | again |
| Graded answer, right with hints | hard |
| Graded answer, right without hints | good |
| Self-rating | the learner's choice: again, hard, good or easy |
The scheduler is in fsrs.ts. Memory is kept per lesson and on the learner's cross-lesson profile, so a concept reviewed in one lesson is known in another.
Effective mastery
What the learner can use now is the posterior discounted by recall:
m_eff = p · R(t) when the concept has been reviewed
m_eff = p when it has never been reviewed
Skipping a concept as already mastered reads effective mastery, so a concept mastered weeks ago and since faded is not skipped. The function is effectiveMastery in learner-state.
Self-rating
After a free-text answer that no judge assessed, the learner compares it with the reference and rates their recall: again, hard, good or easy. The rating is FSRS input only, for every concept of the rated scene. It schedules the next review and never changes α or β. A request waiting for a rating is saved with the learner's state, so it survives a server restart.
Calibration
Confidence
A meta_checkpoint, or a confidence given with a quiz answer, is compared with measured mastery:
gap = c / 4 − p c = declared confidence, 0 to 4
A positive gap is overconfidence, a negative gap underconfidence. A checkpoint sets the gap to its latest value; confidences given with answers form a running average (weight 0.4 per answer). Within ±0.4 the learner is calibrated. Only measured overconfidence withholds a mastery certificate; an underconfident learner who has proven the concept is not penalised.
Trust in AI output
A spot_the_flaw grounded in a real executed result measures how far the learner trusts a flawed AI reading of it:
over_trust = missed flaws / planted flaws
over_distrust = sound claims wrongly flagged / sound claims on offer
gap = over_trust − over_distrust (running average, weight 0.5)
Measured over-trust beyond 0.4 withholds a mastery certificate. A signal-detection summary (discrimination A′ and bias c) is also kept for reading; it never gates.
Unmeasured is not zero
The model never turns an absence of evidence into a value.
- An answer nobody graded (free text without a judge) moves neither α nor β, does not count as an attempt, does not start or clear an error streak. It is counted apart, as an ungraded attempt.
- A concept never reviewed has no retention evidence: its effective mastery is its posterior, not zero.
- A calibration never measured is neither calibrated nor miscalibrated, so no gate fires on it.
- Over-distrust is unmeasured when a scene offers no sound claims; the gap then reflects over-trust only.
- A concept only self-rated is listed in the study plan as "Due for review", never as "To relearn".
- The class report states accuracy only where a key or a judge graded.
See also Honesty and provenance and the glossary.