How it works

The fourteen exercises

They come through three doors: memory, reasoning and speed. Every problem is built fresh each time from a starting number, so there’s no bank of questions to memorise.

Memory

Relational n-back. The usual n-back asks whether what you’re looking at now matches what you saw a few steps back. That’s easy to game: you learn to group things, and your grouping improves rather than your memory. This one asks whether the relationship between the last two things matches the relationship from a few steps back. The things themselves are never the answer, and one that repeats is a trap.

Premise n-back. Same idea, applied to sentences. Does this one have the same shape as the one a few steps back?

Interrupted memory. Hold a sequence while something keeps interrupting. The interruptions are the point: they stop you rehearsing, so what’s measured is holding rather than repeating under your breath.

Tonal span. Hear a run of tones, then tap them back in order on a ladder of eight pitches. They’re too close together to name, so what you hold is the shape of the tune: where it rose, where it dropped, and how far.

Reasoning

Seven families, each closing a different shortcut: order chains, same-or-opposite, compass directions in 3D, syllogisms, logical connectives, mixed problems that combine two kinds at once, and alternating reasoning, where the kind of puzzle changes from one problem to the next, so you work out which rules apply before you apply them.

Order chains covers three scales in one exercise — bigger than, earlier than, and took longer than. They used to be two separate exercises, until we measured them against each other and found they were the same inference wearing different words. Keeping all three scales inside one family means more variety in what you meet, which is better training than any of them alone.

Hidden shapes sits here too. A shape is buried in a busy field and you pick which of four patches it is. Finding it means holding a target in mind and testing it against what’s in front of you, rather than recognising it on sight.

Speed

Sweep pairs. Two quick slides of sound, one after the other. Did each one go up or down? They’re fast on purpose: there’s no second listen, so the skill is answering on the impression you got.

Split-second glance. Something flashes up and is wiped almost immediately. You identify what was in the centre and say where a second thing sat out at the edge. Both, from one flash.

Two things are going on. One is how fast you can take something in at all. The other is how much of the edge of your vision you can use while your attention is busy in the middle. As you get better the flash gets shorter, so the task keeps taking about as long and stays about as hard.

The real version of this one is the Android app. The flash is timed to your screen’s own refresh — down to a single frame — and a browser can’t promise that, so the web version stops at a gentler floor of about a tenth of a second. Good as a preview; the training happens on the phone.

Every problem gets checked before you see it

A generator that marks its own homework will eventually ship a broken problem, one with two right answers or none, and nobody will notice. The thing that should have caught it is the thing that made the mistake.

So each family has a second program that solves problems independently. It was written before the generator existed and shares no code with it. The generator proposes, the checker judges, and a problem only ships if the checker agrees exactly one of the offered answers works. If a wrong answer turns out to be defensible, the problem is thrown out and rebuilt.

One useful side effect: every wrong answer is provably wrong, not just different from the one we had in mind.

And then checked again for accidental giveaways

Checking stops a problem being broken. It doesn’t stop it being guessable.

Anything that changes how a problem looks can leak the answer. Count the premises, count the negatives, notice where the long word sits, and you can beat chance without thinking about the problem at all. You don’t even have to notice you’re doing it.

So the build measures how much each surface feature gives away, in bits — a bit being one yes-or-no question’s worth of information. If any single feature gives away more than 0.05 bits, roughly 3% of what you’d need to know the answer, the build fails.

It’s caught five real ones. The worst was worth 0.93 bits: one family’s wording printed “it is not the case that…” on exactly the problems where that was the answer.

The five leaks, in detail →

The difficulty moves on twenty-three dials

Difficulty here isn’t one number. Twenty-three separate things can be turned up, each loading something different, and only one or two move in any session. Here are the ten that do most of the work.

Two of those deserve a note. Relations at once and number of premises are separate dials on purpose, because ten premises in a straight line are easier than four you have to combine. Most trainers scale difficulty by adding items, which grows the reading rather than the thinking. Both are adapted here, but only one is the reasoning load.

Only one or two dials move in a session, and which ones is decided by a number you never see. So you can’t work out what’s being pushed and put your effort there.

Decisions come from a run of twenty trials, never one. A single trial tells you almost nothing. The staircase aims to keep you at roughly 80% correct: above 85% it makes things harder, at 75% or below it backs off, in between it holds.

One bad session can’t undo months. No dial drops more than a fixed amount below your best, so illness or a noisy train costs you a bounded amount. Come back after a week away and everything eases off 10% on the way in.

Interruptions don’t count against you

A phone call mid-problem produces a wrong answer that says nothing about you. The app notices from three directions — the app going to the background, audio being taken away, the browser tab being hidden — and records what happened.

That trial stays on your record, marked, and feeds into nothing. It won’t lower your difficulty and it won’t earn you a step up.

During a test the rule is stricter: any interruption voids the whole sitting, nothing is stored, and you’re offered a retake. Training is practice either way. A test taken across a phone call isn’t comparable to one taken whole.

The five tests

Here’s the half that answers the actual question.

They’re built to be everything the training isn’t: never practised, never adaptive, never feeding the training. The difficulty is identical every time. Only the specific problems change, drawn from the same pool, so a score from today and a score from next month mean the same thing.

Matrix reasoning — a grid with one piece missing and five candidates.

Figural series — five shapes in a row, each attribute following a rule down the sequence, and five candidates for the sixth. The rule language is small and closed on purpose, so “there is exactly one right answer” is something we can prove rather than hope.

Relational integration — a part-filled grid where each symbol appears once per row and column, and one square has only one possible answer if you follow the constraints far enough.

Symmetry span — remember a run of locations while judging whether patterns are mirror images in between. The judging is the point: it stops you rehearsing.

The fifth one is different

3D mental rotation. A shape built from cubes, and four candidates, one of which is the same shape turned around. We deliberately don’t train this.

If it climbs as much as the other four, you got better at taking tests rather than at reasoning. If the other four move and this one doesn’t, that’s the result worth believing. It’s shown on its own, below a line, and never averaged into anything.

Why we picked an ability we’re least likely to improve →

When you take them

Your first sitting is your baseline, and you take it before you train. It’s open the moment the app is installed, and everything later is compared with it.

First contact with a new test format flatters you for reasons that have nothing to do with ability — learning what the buttons mean, how the timer feels. So every test opens with two warm-up items at its easiest level. They’re marked as warm-up, you’re told whether you got them right, and they never count.

After that the clock is your training rather than the calendar. A check-in comes due once you’ve done 200 minutes of real training since the last sitting and at least a week has passed — both, not either. Sixty minutes of training in a month and sixty hours are not the same month, and a check-in exists to measure what the training did. The week is there because practice effects need a gap however hard you train: 200 minutes in four days still waits out the week. And after eight weeks you’re invited back regardless, because if something faded that’s worth knowing too.

Only training counts toward the 200 minutes. Practice runs, ranked rounds and sessions that were badly interrupted don’t.

A reminder when one’s due, and no nagging if you skip it. Nothing is ever locked behind a test.

Results come back as accuracy per test with a trend line. There’s no combined score and no IQ conversion, and there won’t be. A short test taken at home can’t support either.

Does it actually work? → · What to expect →

Your data stays yours

Everything lives on your device, so the ways out matter.

Export as JSON whenever you want, in a documented format that doesn’t change without notice. Encrypted backup through your phone’s own file picker, so saving to Drive or Dropbox needs no account and no sign-in. On Android it also backs up on its own while you train, keeps the last five, and tells you loudly if it ever stops working.

What the optional online features would send →