ScaleUp
Nobody can hear themselves play.
Pick a skill off the tree, play it, and find out what you actually did — pitch, timing, dynamics, and how you were sitting. The things you stop practising fade, and come back as tomorrow’s quest.
Free to try. No model decides your grade — the numbers come from the recording.
Quartz. Poke it.
The problem
An hour of practice can make you worse.
You can’t hear it while you do it
Your ear is busy playing. You can’t tell a flat third from a nervous one, or rushing from unevenness, and you certainly can’t see your own right shoulder. So the mistake gets repeated until it’s fluent — and fluent feels like progress.
And you won’t notice it going
The things you lose are the things you stopped touching, so nothing you’re currently playing will surface them. Scales you had cold in March are gone by June and there’s no moment where you find out. Until there is.
Why lessons cost what they cost
A teacher is a second pair of ears with a long memory.
They hear the flat note while it’s still in the air, and they remember your left hand collapsing the same way two months ago. That’s the job. The scales and the repertoire you could get from a book.
It’s also why an hour costs what it does. Someone has to be in the room, paying attention, every week, for years — and it doesn’t get cheaper the more people want it, because it isn’t one thing being copied. It’s one person listening to one person.
The listening half sounds like a signal-processing problem, and signal processing is a solved field. It isn’t that simple.
Why this is hard
Every obvious way to grade a performance is wrong in a way you can’t see.
Not hard to build — hard to build so the number means something. Each of these is a version that runs, returns a score, and would quietly teach you the wrong lesson.
The nearest note is not close enough
Round what you played to the nearest semitone and almost everything comes back correct. A note a quarter-tone flat rounds to the right name and passes. That is the entire problem a string player has, and it is invisible at that resolution.
What we do insteadIntonation is measured in cents against the written pitch, so being nearly right is a number rather than a pass. Fifty cents is a quarter-tone; the failure threshold sits at half of that.
backend/app/evaluation/violin.pyRight note, wrong moment
Line your notes up against the score one for one and time disappears. Play the passage perfectly but a beat late and nothing objects. Compare strictly by the clock instead and playing it slowly to get it right becomes an error.
What we do insteadTakes are aligned elastically against the score rather than compared position by position, so a slow tempo and a wrong rhythm stop being the same mistake.
backend/app/evaluation/dtw.pyLoud is a property of your room
Absolute level tells you about the microphone, how far away you sat, and the size of the room. Score it directly and moving your laptop counts as a change in your playing.
What we do insteadLevels are centred on your own take and dynamics are scored as rank agreement — did the crescendo actually happen — which no amount of gain can flatter.
backend/app/evaluation/dynamics.pyA number that measures nothing looks exactly like one that does
Silence, a hand out of frame, a hip the camera cannot see. The tempting thing is to score it anyway, because a grade with a gap in it feels broken — and nobody can tell the difference by looking.
What we do insteadAn unmeasured dimension is reported as absent and the rest are renormalised around it. A drummer is not marked down for pitch. Nothing is scored zero because it was not observed.
backend/app/evaluation/registry.py
The through-line is the last one. A plausible number is far cheaper to produce than a true one, and on a screen they look identical. So the rule here is that a dimension nothing measured is reported as missing, never as zero — and the score you get is the one the recording can actually support.
What you get
A grade you can argue with.
Every take comes back with the dimensions it could measure, the ones it couldn’t, and an examiner’s read on what to fix first. The numbers are measured off the recording; a model may improve the wording and never touches them.
Shipping today: banjo, drums, guitar, piano, trumpet, violin.
- 6
- instruments with a published, versioned curriculum
backend/app/curricula/ - 53
- skills across those 6 prerequisite graphs
backend/app/curricula/*.json - 4
- dimensions scored in a take — pitch, rhythm, dynamics, technique
backend/app/evaluation/registry.py - 16
- posture rules, each declaring the landmarks it needs to be honest
frontend/lib/posture.ts
And then it fades
The tree goes dull where you stop going.
Every skill carries a review schedule. Play it well and the gap before it comes back gets longer; leave it and the gap closes. None of that is stored as a number — it’s worked out from when you last played it and how it went, so changing what “solid” means changes your whole history rather than whatever a nightly job last wrote down.
What it feels like: the thing you were best at in March turns up as tomorrow’s quest, and you find out you’d lost it before anyone else does.
Quartz has opinions about your semiquavers. It doesn’t get a vote on your grade — nothing it does changes a number.