Daily practice for Högskoleprovet
A new session every morning, built from what you got wrong yesterday.
Sofia plans the day's practice for you, marks it exactly — with code, never with a language model — and brings back what you missed just as you are about to forget it. On screen, or printed on paper the way the real test is taken.
Nothing to pay now.
If you are new to it
Högskoleprovet is Sweden's national university aptitude test. It is held twice a year, taken by around 70,000 people each sitting, and it opens a second route into higher education: for ordinary first-cycle programmes, at least a third of the places are allocated on the test result rather than on school grades, provided you meet the entry requirements. A result is valid for eight years, you may sit it as often as you like, and only your best result counts.
-
The marking never guesses
No language model judges your answer. It is compared by code against a stored, checked key: exact comparison, unit conversion and a computer-algebra engine. And when the code cannot reach a verdict, it says so instead of guessing.
-
On paper, like the real thing
The test is written by hand, in five 55-minute sections, with no screen and no calculator. So the whole day's session prints to A4 with working space and an answer sheet — and is marked afterwards by exactly the same code.
-
A view for parents
An overview for guardians of how the week went: sessions completed, days missed, minutes trained, strength per section and the weakest skills. It shows outcomes and discipline — not every question and every answer.
The hard part is rarely the studying. The hard part is knowing what to study today — and going back to what you have already got wrong, instead of practising more of what you can already do.
Around the clock
What a day looks like
Six steps, every day, without anyone having to remember anything.
-
The night before
The planner builds tomorrow's session from your own skill profile: the skills you are weakest at, that you are about to forget, that carry the most marks in the real test and that you get wrong most often — divided by how long each one takes. No randomness: on the same inputs, the same day always produces the same plan.
-
Half past six
A letter with a personal link straight into the day's session. It shows how yesterday went, how many days in a row you have trained, which section is weakest right now and how many days are left. No language model writes a word of it.
-
The session
About 75 minutes, split into blocks: revision, targeted practice on one skill, mixed questions, a timed set, and yesterday's errors. Screen or paper, your choice.
-
Marking
Multiple-choice, numeric and expression answers are checked the moment you submit them; a freely written solution you mark yourself against the worked solution. A wrong numeric answer often gets a diagnosis rather than just a cross — a sign error or a power-of-ten slip is named.
-
The mistake books itself
What you got wrong comes back after 2 days, then 3, 7 and 14 — the base schedule, stretched out for skills you are already strong at. Get it wrong again and it drops a step. Your estimate for that skill moves at the same time.
-
That night, and on Monday
Every night the system reads back over three weeks of your own answers and works out the pattern: which skills you struggle with and how — a misconception, carelessness, a guess, or working too slowly. Monday brings the weekly report.
Marking
The marking is code, not a model
Multiple-choice answers are compared exactly. Numeric answers are read with the Swedish decimal comma, fractions, roots, π and scientific notation, converted between units, and accepted within the tolerance the question specifies. Algebraic expressions go to a computer-algebra engine, so 2x+1 and 1+2x both pass.
This matters because it is the one thing in the system that is never allowed to be wrong. Language models are used elsewhere — to read old papers, to write explanations, to check questions and to draft new ones — but no model looks at your answer. The key it is compared against has itself been checked before the question is used; how that works is further down.
And when it cannot reach a verdict in time, it says so. No verdict is recorded, your skill profile is untouched, and you get the turn back. A right answer that happened to take too long to check should not become a wrong answer in your statistics.
The skill model
The system tracks what you know — and what you are losing
Each skill carries an estimate that moves rather like a chess rating: by the gap between how you performed and how you were expected to, in smaller and smaller steps as confidence in the estimate grows. A lucky guess counts as half a mark's worth of evidence. A slow correct answer as four fifths.
Between sessions the estimates decay along a forgetting curve — towards a floor, not to zero. The planner sees the decayed number, not the one from the day you last practised. That is why revision arrives when it does.
Revision is capped at 20 items a day, oldest first. The rest wait in the queue instead of swallowing the day's session.
Paper
The test is written on paper. Practice can be too.
The whole day's session prints to A4: the questions with their figures, working space sized for the type of question, and an answer sheet. The key on a separate page, if you want it.
Afterwards you type the answers in, and they are marked by exactly the same code as on screen. The revision queue and the skill model update. The one thing paper cannot carry over is how long each answer took, which the screen measures by itself. Practising on paper costs you next to nothing in follow-up — which is the whole point, because that is how the test is actually taken.
The questions
No question is served until its answer key can be traced
Questions come from past papers in the official archives. The system fetches from nowhere else — the source list is an allow-list, re-checked at every redirect.
Each page is read by a model that can see the image. But the answer key is never taken from one model's reading: two independent models must return an identical value for a cell to be accepted. Where they disagree the answer is withheld rather than guessed. The models' own confidence scores are recorded — and deliberately ignored.
Before a question is served, its key must have a traceable basis: either a panel of four models from different families has solved it blind, never shown the answer, and landed on the stored key — or the key comes from the official published answer sheet and the copy has passed a mechanical integrity check. A question the system generates itself is normally only released if every model on the panel agrees, not a majority.
Some whole papers are sealed. They are never used for practice, never appear in revision, never feed the vocabulary drill and are never shown to the model that writes new questions. They exist for one purpose: so that on test day you sit a paper you have genuinely never seen.
The tests
What you are actually training for
Sofia is built for Högskoleprovet today. Support for Matematik- och fysikprovet is being built now.
Available
Högskoleprovet
Five 55-minute sections of 40 questions each. One of the five is a trial section that does not count — but it is indistinguishable from the rest, so all five are written in earnest. 160 questions are scored: 80 verbal and 80 quantitative.
The eight subtests are XYZ (mathematical problem solving), KVA (quantitative comparisons), NOG (quantitative reasoning), DTK (charts, tables and maps), ORD (vocabulary), LÄS (Swedish reading comprehension), MEK (sentence completion) and ELF (English reading comprehension). The test is in Swedish.
- Next sitting
- 18 October 2026
registration closed - Then
- 10 April 2027
registration 7–14 January - Scaled score
- 0.00–2.00 in steps of 0.05
- Result valid for
- eight years
- Attempts
- unlimited — the best valid result counts
- Registration fee
- SEK 550
The mean in autumn 2025 was 0.90. About 8 % of those who sat reached 1.50 or above, and just under 2 % reached 1.75. Source: UHR's own norming and distribution tables.
In progress
Matematik- och fysikprovet
A selection test run jointly by Chalmers, the University of Gothenburg, KTH and Stockholm University. It confers no eligibility — it gives one more group to compete in. At KTH, up to 11 % of the places on Materials Design are filled this way, and up to a third on Engineering Physics and Engineering Mathematics.
Held once a year, in spring, in Gothenburg, Stockholm and Umeå. Three hours of mathematics for 45 points, two hours of physics for 30, roughly half of it mechanics. No formula sheet and no calculator. A pass is 30 of 75 points. The result is valid for the current admission round only.
Past papers are freely available back to 2007 — but with bare answer keys only. There are no official worked solutions at all. That gap is what the work on Matematik- och fysikprovet is aimed at.
- Skill taxonomy
- in place
- Marking
- shared with Högskoleprovet
- Questions and scoring
- being built
Why it exists
Built for one student
Sofia was built by one engineer for a single student — his daughter, who was preparing for Högskoleprovet. That is why it looks the way it does. Nobody builds a marker that says “I did not manage to check that in time; that is my fault, not yours” because it looks good in a brochure. It exists because a right answer was once marked red, at a kitchen table, and that was unacceptable.
The same goes for the sealed papers, for the rule that two models must read the same answer-key cell identically, and for the button that lets the student say the key looks wrong and take the question out of her own rotation immediately. Every one of those rules came from something that went wrong first.
The system has run in production every day since August 2026, and has over 700 automated tests that run against a real database before anything changes. It is also, honestly, young. It opens to more families a few at a time — not to manufacture a queue, but so that each one has time to go wrong in a way that can still be put right.
Waiting list
Register your interest
Sofia runs today for one student. It is opening to a small number of families at a time. Leave your address and we will be in touch before the next intake — with what it costs, what is included, and what is still missing.
Nothing to pay, no commitment. We will not send you anything else.
Questions
What does it cost?
That is not decided. Sofia is not sold today, and the waiting list costs nothing. When there is a price, everyone on the list hears about it before anything takes effect.
When does it open?
We will not set a date we cannot keep. The system runs every day for one student, and some of what is needed to run it for several families simply is not built yet — separate exam dates per student, a real parent account, registration. That work is under way, and it is measured in months, not weeks.
Will it raise my score?
We do not know, and we are not going to claim it. Sofia has run for one student for a few weeks — there is no cohort, no measurement, and therefore nothing to promise. What we can describe is the method and the content, and that is what this page does. If we ever publish a number, the method will be published with it.
Is this another AI app?
No, and the line is drawn in a specific place. Language models read in past papers, write explanations, check questions and summarise your error patterns. But nothing that decides whether an answer is right is a model — that is code, deliberately. A model that is wrong about an answer key does real damage: it teaches the wrong thing and destroys trust in everything else.
What does a parent see?
Whether today's session is done, the streak, minutes trained, missed days, strength per section, the last four weeks of sessions, the weakest skills, timed-test results and the weekly reports. Not question texts, and not individual answers. Practising should not feel surveilled.
Do you need a computer?
A browser is enough, plus a printer if you want to work on paper. There is no app to install. The morning letter carries a personal link that works once and then stops working.
What happens to the details I leave here?
Your name and e-mail address are stored so we can contact you about Sofia, and for nothing else. They are not passed on. You can remove yourself from the list at any time without contacting anyone. How your data is handled.