Learning principles

This page turns the evidence from learning science into Estudar's design rules. Each rule states the evidence, the rule and its current status in the app. The rules are invariants: if a feature conflicts with a rule, the feature changes.

Strength of evidence

• HIGH — replicated experimental results; high utility according to Dunlosky et al. (2013).
• MODERATE — good experimental support, under narrower conditions.
• EMERGING — few studies; implement, instrument and check against our own data.

Observational reports serve as context: they motivate the problem, but do not justify decisions on their own.

Implementation status

Rule	Evidence	Status
Core model	HIGH	Implemented
R1 Attempt before revealing	HIGH + EMERGING	On paper
R2 AI only after an attempt	EMERGING	Met
R3 Spacing with a deadline	HIGH	At session level
R4 Interleaving	MODERATE	At plan level
R5 Self-explanation	MODERATE	Optional prompt
R6 Confidence and calibration	HIGH	Implemented
R7 With help vs. unaided	EMERGING	Implemented
R8 Low-utility techniques	HIGH (negative)	Implemented
R9 Time per subject	Heuristic	Implemented

"On paper", "at session level" and "at plan level" mean the rule is applied to what the app knows today — sessions and logs — and not to individual items, which will only exist with the question bank (roadmap).

Core model

Learning is measured by unaided retrieval, after some time has passed — never by performance while studying. Performance during practice is a misleading signal:

• rereading felt better and retained less (Roediger & Karpicke, 2006);
• blocked practice did better during training and much worse a week later (Rohrer & Taylor, 2007);
• with an AI tutor, performance with help went up and unaided performance went down (Bastani et al.).

In Estudar: mastery comes only from unaided closed-book tests; hours, practice and practice with help are shown separately.

R1. Attempt before revealing

Evidence. The testing effect (Roediger & Karpicke, 2006; Karpicke & Blunt, 2011). The crutch effect (Bastani et al.): a tutor that gave answers harmed unaided performance; one that made students reason first removed much of the harm.

Rule. Nothing that contains or implies the answer is shown before an attempt has been logged. An empty attempt counts, but as a failed one.

In Estudar (paper mode). The block log asks for attempts, confidence and help used before marking; those fields are then locked; only after that does the number correct appear, capped at the number attempted. A blank exercise counts as attempted and wrong. Self-assessment is flagged as such and compared against closed-book tests (R6).

R2. AI is a post-attempt tool

Rule. Before the attempt, only hints that do not reveal the answer. Afterwards, marking, explaining and generating variants. In closed-book tests and mock exams, none.

In Estudar. The AI only generates the weekly plan and reads the image of your degree record; it never sees exercises or answers. Practice done with AI is logged as "with help" and never counts towards mastery. A phased tutor is future work and will have to respect this table.

R3. Spacing, with an exam deadline

Evidence. Ebbinghaus (1885); Cepeda et al. (2006, meta-analysis); Cepeda et al. (2008): the optimal gap between reviews is around 10–20% of the time until the test.

Rule. Every subject has an exam date; spacing guarantees review in the final window; the day before the exam is review only.

In Estudar. Exam dates per subject (or taken from the next assessment in the degree record); the day before is marked "retrieval and review only, no new material"; priority in the last 7 days; the plan requires every medium or heavy subject on 2+ days, with reviews 1–3 days later. Per-item scheduling (FSRS) arrives with the question bank.

R4. Interleaving

Evidence. Rohrer & Taylor (2007): mixed practice ~63% vs. blocked ~20% on a test a week later — the gain comes from having to identify which method applies.

Rule. First exposure can be blocked; review and practice interleave topics and problem types.

In Estudar. The plan asks for mixed exercises and cross-reviews on the same day; closed-book tests ask for mixed topics; first exposure is logged separately.

R5. Self-explanation

Rule. After an attempt, sometimes ask the student to explain why a step works — in moderation.

In Estudar. On every third practice log, an optional field asks for a one-sentence self-explanation.

R6. Confidence and calibration

Evidence. The illusion of competence: people who reread predict better results than people who test themselves, and are wrong (Roediger & Karpicke, 2006). Bjork (1994): fluent study feels effective and is not.

Rule. Before revealing, confidence (1–4). Calibration per subject. Subjects showing overconfidence rank above those that are merely weak.

In Estudar. Exactly that — see Block log.

R7. With help vs. unaided

Rule. Mastery is computed only from unaided retrieval. Performance with help is shown separately and never inflates it. Measure the gap between the two: a growing gap is the crutch effect in our own data.

In Estudar. Mastery comes only from unaided closed-book tests; practice with help and unaided side by side; a relies on help warning; tests taken with help become practice.

R8. Low-utility techniques are not a study mode

Evidence. Dunlosky et al. (2013): rereading, highlighting and summarising have low utility.

In Estudar. Reading new material is logged as first exposure and never counts towards mastery; the app does not offer AI-generated summaries; the prompt forbids those techniques as the main activity and the plan check warns when they appear.

R9. Time allocation

Status. A planning heuristic, not learning science. It decides how much time; R1–R8 decide how.

priority(d) = ECTS(d) × deficit(d)
share(d)    = max(10%, priority / Σ priorities), renormalised
deficit(d)  = 1 − unaided mastery(d)     — no test yet: 1

ECTS are the European credits that measure each subject's workload. Overconfidence increases the deficit; the nearest exam jumps the queue; exams already sat drop out. Details and parameters in AI weekly plan.

How these rules are enforced

Every rule with a measurable consequence has an automated test, and every test is checked by deliberately breaking the rule (mutation) to confirm the test catches it. See Tests.

Measure before deciding

After two weeks of use, compare your closed-book tests with your practice with help before tuning any parameter. The semester report shows whether your tests predicted your grades well.