Paper-first: why our tutor watches a worksheet instead of a screen.

The decision, in one paragraph.

Most kid-facing math products ask the child to look at a tablet. Koda doesn't. We built our tutor to watch a real paper worksheet on a real desk — pencil in hand, notebook open, no screen between the child and the work. We did this for a specific reason: the cognitive-science literature has been pointing in this direction for over a decade, and elementary math in particular is the wrong place to absorb the costs of getting it wrong. This note lays out what the research says, what we think it means for a 9-year-old learning fractions, and the honest counter-arguments we've considered.

What changes between writing and tapping.

Writing by hand and tapping a touchscreen are not equivalent. They use different muscles, produce different motor traces, generate different feedback signals to the brain, and — this is the part that matters — appear to leave different memory residues. The literature comes from two separate places: adult note-taking studies, and developmental neuroscience on pre-readers. They point in the same direction.

The note-taking literature: Mueller & Oppenheimer (2014).

In a series of three experiments at Princeton and UCLA, Pam Mueller and Daniel Oppenheimer asked undergraduates to take notes during a TED-talk-style lecture, either on a laptop or in longhand. Both groups did equally well on factual-recall questions. On conceptual questions — the ones that required actually understanding the lecture rather than reciting it — the longhand-takers did better, and kept doing better after a 1-week delay.

The mechanism the authors proposed has held up well: laptop typists transcribed verbatim, while longhand note-takers had to summarize because they couldn't keep up. The summarizing forced encoding. The transcribing didn't. This finding is sometimes overstated in the popular press; the conceptual-question advantage is real but moderate. What's under-stated is that the same effect appears even when laptop note-takers were explicitly instructed not to transcribe — they did it anyway. The form factor pulls the behavior.

Mueller, P. A., & Oppenheimer, D. M. (2014). The pen is mightier than the keyboard: Advantages of longhand over laptop note taking. Psychological Science, 25(6), 1159–1168.

The developmental literature: James & Engelhardt (2012).

A different kind of evidence comes from Karin James's lab at Indiana University. James and Engelhardt studied pre-literate children — kids who hadn't yet learned to read — and split them into three groups that practiced letters in different ways: one group printed letters by hand on paper, one traced letters, one typed them on a keyboard. The kids in the printing group later showed brain activation patterns, measured by fMRI, that resembled the patterns adult readers show when looking at letters. The tracing and typing groups did not.

The interpretation: the act of producing a letter shape — the variable, error-prone, motor work of making it appear from nothing — engaged a network of brain regions tied to letter recognition that tracing or typing didn't engage. Recognition followed production. The kids whose hands had drawn the shapes recognized those shapes differently afterwards.

James, K. H., & Engelhardt, L. (2012). The effects of handwriting experience on functional brain development in pre-literate children. Trends in Neuroscience and Education, 1(1), 32–42.

Why this matters for math specifically.

Most discussions of writing-versus-typing focus on prose. We think the argument is even stronger for elementary math, for three reasons.

1. Math notation is two-dimensional.

Adding two three-digit numbers requires aligning the columns: ones above ones, tens above tens, hundreds above hundreds. The standard subtraction algorithm makes the child cross out a digit and write a smaller number above it. Long division places the quotient above the dividend, with successive products written below in carefully indented columns. None of this fits comfortably on a touchscreen keyboard. A 9-year-old using a tablet for arithmetic is fighting the interface; a 9-year-old with a pencil and a sheet of grid paper is using a tool that was designed for the job.

2. Errors are visible.

On paper, when a child writes "23 − 17 = 14," the work is right there. The teacher (or a tutor, or a camera looking down) can see where the slip happened. Did they regroup correctly? Did they subtract the ones column or skip it? The work shows the shape of their thinking. On a tablet, after the kid hits "submit," the answer is either right or wrong, and the trace is gone. Paper retains the data structure of the work.

3. Drawing is part of the curriculum.

Modern elementary math curricula (including the Common Core standards) ask kids to draw fraction circles, number lines, area models, and bar diagrams as a routine part of solving problems. These representations aren't decorative — they're how the child learns to see quantity, before symbolic notation takes over. A pencil draws a fraction circle in 3 seconds. A finger on a tablet doesn't, and the workarounds (preset shapes, drag-and-drop manipulatives) train the kid to use the app rather than the math.

What "watching paper" actually buys you.

A note on status. The description below is the design Koda is built around — the overhead camera reading handwritten work in real time. That loop is the next big milestone on the build plan and isn't wired into the production session view today (today's build uses typed input while the handwriting OCR and capture path are in flight). For the shipped-today vs. roadmap split, see the first 15 minutes.

When that loop is live: Koda's overhead camera looks down at the worksheet at about 15 frames a second during an active session. The vision model reads handwritten work as it appears, not after submission. So the tutor sees the regrouping mark, the scratched-out digit, the stalled pencil mid-line. It can ask a question while the work is still in progress — not after the answer is wrong. The pencil stays your child's; the tutor's job is to ask the small question that hands the next move back.

This is a different shape of feedback than a tablet app provides, and it's the shape the literature implies kids learn from. The child writes (encoding); the tutor reflects what the child wrote (drawing attention without supplying the answer); the child writes again (re-encoding). Three cycles a minute. The paper holds the trace.

The honest counter-arguments.

We've thought about these and want to name them out loud.

Tablets aren't bad for everything. For drilling math facts, for adaptive flashcards, for visualizing 3D solids, for gamified pattern matching — a tablet often beats paper. The argument here is specifically about the working math phase: the moment a child is in the middle of doing a problem on the way to an answer. That's the phase that wants a pencil.

The Mueller & Oppenheimer effect has been challenged. Replication efforts have produced mixed results, and several follow-up studies found smaller effects than the original. The developmental literature (James and others) has held up better, but the field is still in motion. We're not claiming a clean settled science. We're claiming that the direction of the evidence, in the domain we care about (kids working on arithmetic in a non-laboratory setting), points the same way enough times that betting against it would require an argument we haven't been able to construct. The strongest replication evidence is in the elementary years; we extend the same hand-on-paper default to Pre-K and middle grades, where the literature is thinner but the cognitive logic carries.

Some kids can't use a pencil comfortably. Children with dysgraphia, fine-motor delays, or specific physical conditions write slowly or painfully. We take that seriously. Koda's roadmap includes a spoken-answer path — a dysgraphia accommodation where the child says the answer instead of writing it — and the parent portal will be able to configure a slower pace and larger physical worksheet. The spoken-answer pipeline is on the workbench and not yet live; for now, the typed-input fallback is the alternative. We want hand-on-paper to be the default; we don't want it to be the only option. That's the responsibility that comes with picking a default.

What would change our minds.

We try to write down what evidence would make us reconsider, so that future-us has to argue with past-us rather than quietly drift.

• A well-powered, pre-registered randomized trial of paper-first versus tablet-first math instruction across the elementary or middle grades, run over at least one school year, with conceptual-understanding outcomes that include transfer questions. If a study like that finds equivalent or better outcomes for the tablet condition, we owe the field a serious rethink.
• A new generation of styluses + paper-feel display surfaces that reproduce the kinematics of writing on paper closely enough that the developmental-fMRI literature stops distinguishing them. We watch the ReMarkable / Boox / Onyx / iPad-with-paper-screen-protector space; today nothing is quite there for an 8-year-old's hand.
• Field evidence from Koda's own families that the paper-first design is itself excluding a group we wanted to serve — say, families where a worksheet is hard to source, or where a child has a strong tablet preference for sensory reasons. If that pattern emerges in real data, we adjust.

Closing.

We picked paper-first because the literature points there, the math curriculum demands a 2-D notation tablets don't comfortably support, and the visible trace of the work is the surface our tutor needs to actually help. The decision constrains the product — Koda needs an overhead camera, can't run on a phone, costs more to ship — and we accept those costs. The alternative is a tutor that lets the screen take over the part of math that should stay in the child's hand. We aren't building that one.

If you'd like to know when Koda ships, the waitlist is here. If you'd like to read the related architecture note, we wrote about the local-only build choice.