Digital SAT · first cohort
The tutor that never
forgets a mistake.
A tutor sees your kid two hours a week and remembers what a person can. Ultralearn reads every answer — with how sure your kid was before each reveal, and how long it took — names the exact error behind every miss, and forgets none of it. Aimed at one target score and one test date.
Try the real thing — this is a turn
2x + 6 = 20. What is the value of x ?
Pick an answer — wrong ones are the interesting part.
The plan
A target, a date, and a plan that just works.
Setup is three questions: the target score, the test date, and when your kid can work. Week one is a full-length diagnostic in an exact replica of the College Board’s Bluebook app — then the model of your kid exists, and daily sessions begin: typically 20–50 minutes, recalibrated weekly from your kid’s actual pace, not a brochure’s. And if the date and the target stop being compatible, it says so early — it will not quietly let a plan fail.
Why so few minutes? Because none of them are wasted.
Practice is spaced — a skill comes back right before your kid would forget it, so it consolidates instead of evaporating by test day. Skills are interleaved — mixed the way the real test mixes them, because the SAT never labels a question with its chapter, and neither do we. And the testing is the teaching — pulling an answer from memory is what makes it stay, so there’s no rereading, no highlighting, no busywork that feels productive and does nothing. This is the learning science the research has been shouting for decades. Most prep ignores it, because done right it feels harder. It is. It’s also why it works.
The mechanism
One model of your kid. Everything follows from it.
You’ve seen prep that “worked” right up until test day — that’s practice measuring the wrong thing. Here’s what runs instead:
Test
It starts with a full-length diagnostic in an exact replica of the College Board’s Bluebook app — the two-pane passages, the timer, the calculator — and every daily session keeps testing with real SAT-style questions, confidence rated before each reveal. Testing does double duty here: every answer pulled from memory sticks harder, and every answer teaches the model something new about your kid.
Model
Every answer updates a model of your kid — what’s solid, what’s shaky, and the exact named error behind every miss:
…written onto a map of all 30 skills — tap a skill to use it:
Arrows are teaching order — the engine will not drill circles into a student who can’t yet complete the square.
Linear functions — 14 named misconceptions tracked
slope-intercept-swapslope-formula-invertedparallel-slope-not-equal+ 11 moreReading the slope where the intercept lives is a specific, recurring move — not general “bad at graphs.”
These names are verbatim from the engine’s registry — the same entries the model writes to when your kid picks a wrong answer built to catch exactly that move.
Lesson
Each lesson is built from what the model found — aimed at the one fix that buys the most points for the study time. Taught, then proven: nothing counts as learned until your kid does it cold, days later, on a version they’ve never seen. Sure and wrong gets interrupted and broken first; mastery that slips is revoked and retaught from a new angle.
Generic is what software does when it doesn’t know the student. This knows more every day, and forgets none of it. That’s the whole machine.
For parents
Built to be trusted, not to be sticky.
You judge prep from one chart: the target, the date, and progress we can defend — if we can’t stand behind a number, you don’t see it. Past that, the honesty is structural:
The deeper honesty is in the incentives. Progress is denominated in work done, not minutes-on-app — right answers earn nothing by themselves, because rewarding correctness teaches answer-hunting. Catch-up after a missed day is deliberately capped, and past what a brain consolidates overnight, extra time is redirected to review instead of new material. We couldn’t sell you junk hours if we wanted to.
Why this is a category of one
Tutoring was never the ceiling. Bandwidth was.
In 1984, Benjamin Bloom measured what parents have always suspected: an average student, tutored one-on-one under mastery learning, ends up ahead of 98% of a conventional classroom. His famous catch: one tutor per student is “too costly for most societies to bear.” Everything since — prep classes, practice apps — imitates the tutor cheaply. Ultralearn is the mechanism itself, with a memory no human has. It doesn’t beat a great tutor’s insight; it beats their bandwidth, their memory, and their bar — the three places tutoring actually leaks.
| Prep class | Practice app | Private tutor | Ultralearn | |
|---|---|---|---|---|
| Sees every answer | No | As right / wrong | Two hours a week | Every one — with confidence and timing |
| Names the exact error | No | No | The ones they catch | Every wrong answer → one of 600+ named misconceptions |
| The bar for “learned” | The course ends | The streak survives | The tutor’s judgment | Does it cold, days later, on a version they’ve never seen — and it’s revocable |
| Remembers week 1 in week 9 | No | No | If they kept notes | Perfectly |
| Mirrors the real testing app | Sometimes | Rarely | No | Down to the calculator |
| Aimed at a target score and date | No | No | Informally | It’s the literal input: score + date → daily dose |
| Typical cost | $$ | $ | $1,500–6,000 a season | First cohort — below |
Fair questions
Ask us the hard ones.
Every answer below is a piece of architecture that already exists — not a policy, not a promise.
Who writes the questions — and how do you know they’re good?
Items aren’t conjured on the fly. The bank is authored against the full public corpus of released digital-SAT questions — all 2,929 — so skill coverage and frequency mirror the real exam. Every wrong answer choice is engineered as the fingerprint of one named misconception, and the compiler refuses to ship an item that breaks that contract. On top of that, the founder grades items in a review console, side by side against real College Board questions of the same skill and difficulty.
Is this just a chatbot in a costume?
No. The core is bookkeeping, not conversation: a model of your kid plus a selector computing what comes next. There is a tutor to ask “why?” — but it unlocks only after an answer is submitted, deliberately: between the question and the answer, your kid is on their own, because an answer that isn’t theirs teaches the model nothing.
How will I know it’s working?
One chart: the target score, the test date, and honest progress toward it — if we can’t defend a number, you don’t see it. And the engine grades itself: every night it replays its predictions of your kid’s answers against what actually happened. When it’s wrong about your kid, that’s a defect — measured, logged, fixed.
What if it’s wrong about my kid?
It’s built to find out. Mastery is revocable — a failed cold re-check reopens the skill. A lesson that didn’t land is retaught from a different angle, not repeated louder. And nothing about how it teaches changes on a hunch: every tuning change is proposed from replayed evidence and approved by a human.
First cohort
Fifty students. Not fifty thousand.
Ultralearn is onboarding its first cohort personally — about fifty students, each trajectory watched by the founder. An access code unlocks the full engine: the diagnostic, the daily loop, the tutor. If someone shared this page with you, ask them for their code.
Already practicing on another device? Sign in and your progress follows you.