Exercise types.

How grading works.

By cognitive level.

Following Bloom revised. The app picks the level based on what the learning point actually requires.

Recall

Get the fact back from memory. Graded on your device.

• Flashcard

  A question on one side, the answer on the other. Self-rated; the FSRS scheduler reads that rating to set the next review.

• Fill the blank

  AI fallback

  A sentence with one or more blanks. Each blank is checked locally, ignoring whitespace and case. If some blanks match and others do not, the judge reads the rest for wording the answer set did not anticipate.

• Multiple choice

  Pick one or more correct answers from a list. At least three options, so the result is never a coin flip. Capped at 30% of any plan to avoid recognition-without-recall drift.

• True or false

  Statements designed around the misconceptions the subject actually creates, not generic claims.

Understand

Show you grasp the structure. Deterministic.

• Matching

  Pair items from two lists. Cross-checks every pairing. At least five pairs, so one slip cannot decide the verdict.

• Categorize

  Sort items into groups. Partial credit for partial sorts. At least five items, so one misfiled item cannot decide the verdict.

• Ordering

  Arrange a sequence in the right order. Penalty per swap.

• Word order

  Rearrange scrambled fragments into the one correct statement. All or nothing: unlike ordering, a sentence that is nearly right is not partly right.

Apply

Produce an answer, not just recognize one. Graded on your device, with the judge as a fallback on the two free-text types.

• Short answer

  AI fallback

  A word or short phrase, matched locally against the answer set after normalization. When it does not match exactly, the judge reads your wording rather than marking a synonym wrong.

• Numeric answer

  Calculate a result. Tolerance and units configurable per exercise; the answer is read in either decimal convention.

• Equation solving

  AI fallback

  Solve a math or physics problem. Numeric and normalized-form answers are settled on your device; a form that is equivalent without being identical goes to the judge, so 2x + 4 and 2(x+2) both score full.

• Code trace

  Read code, predict the output. String match with whitespace normalization.

Analyze and create

Produce something or diagnose what is wrong. AI judge.

• Written response

  AI judge

  Explain a concept, make an argument. The judge applies a rubric specific to the exercise, returns a score and per-criterion feedback.

• Code review

  AI judge

  Find what is wrong in a piece of code and propose a fix. Judged on both diagnosis and remediation.

What happens when you get one wrong.

Four things, in this order, and the order is deliberate — a correction that arrives after an invitation to reflect interrupts the reflection instead of informing it.

1. 1. Your answer is marked against the right one. On the six types with options or arrangements, the correct choice is marked in place, and yours alongside it, so you can see the contrast rather than reconstruct from memory what you picked. Every mark carries a symbol and a text label, never colour alone, and your own choice is never painted as a failure.
2. 2. The correct answer is stated. Always, immediately, never behind a tap. For the types whose answer appears nowhere else on the card, it is stated in its own panel. This fires whenever a graded answer falls below the same threshold the memory model uses to call something correct — so the correction and the scheduling agree by construction.
3. 3. The why, and what to check next. On types graded by a language model this is guaranteed and specific. On types graded by exact comparison it depends on whether the exercise was written with an explanation — see the honest limits, because today it often is not. When it is missing the card says so and points you back to the section that covers the concept, rather than leaving the verdict standing on its own.
4. 4. Then, and only then, an invitation to explain it yourself. Optional, never auto-expanded, and it tells you upfront what it buys: the next exercises on that concept are tuned to what you say.

One exception, and it is the important one. Exams, placement tests and recalibrations reveal nothing per question. You find out at the wrap-up. Withholding feedback until the end of a diagnostic is what keeps it a measurement rather than a lesson — and it is what makes the wrap-up worth reading.

There is no retry on a missed exercise in a practice session. Once the answer is on screen, a second attempt would be a copy, and it would bank a full-strength memory signal on an answer you were just given.

From your answer to a number.

A graded answer produces two different numbers, and confusing them is a mistake we made and fixed. One is whether you were right. The other is how much evidence of durable memory being right on that particular type provides.

• Being right is being right. It does not depend on the type. A correct true/false answer is a correct answer, and it counts as one.
• How much it proves does depend on the type. Some types hand a share of the credit to someone who knows nothing — a true/false is a coin flip, a four-option multiple choice is one in four. On those types, and on matching and ordering, the signal the scheduler receives is discounted by roughly what guessing alone would earn. Types where nothing is handed to you — anything you have to write or compute — are not discounted, because there is nothing to discount. Two correct true/false answers buy less spacing than one written explanation, which is the honest reading of them.
• Some shapes cannot prove anything, so they are not written. A two-option multiple choice is a coin flip whose best possible result still says nothing, and a matching set of three pairs turns one ordinary slip into an outright failure. The authoring rules refuse both: multiple choice carries at least three options, and matching and categorize at least five items.
• A lapse means you forgot, not that you were imperfect. Only an answer the memory model reads as a genuine retrieval failure counts as one. An answer that was recalled but shaky is neither a success nor a lapse — it is what it is, and it is scheduled accordingly.
• The four flashcard buttons mean what they say. Again, Hard, Good and Easy are the four grades the scheduler itself uses, so the rating you pick is the grade it applies. In particular, "Good — solid recall" is recorded as a recall. Honest self-rating should never cost you more than generous self-rating does.

And a wrong answer has to be your mistake, not the grader's. Numbers are read in either convention, so 2,5, 2.5 and 2 500 are all understood as the numbers you meant — the app is bilingual, and the decimal separator your own keyboard produces was never a reason to be marked wrong. Predicted program output is compared ignoring spacing and line endings, so [1,2,3] and [1, 2, 3] both pass; capitalisation and punctuation still count, because those are genuinely different output.

Disagreeing with a grade.

Both deterministic and AI-judged grades can be contested. For deterministic exercises (typos, alternate phrasings the answer set missed), the contest can update the exercise's answer set. For AI-judged exercises, you write a counter-argument; the judge re-reads your original answer plus your argument and revises the score with new feedback.