Why we picked paper over a stylus.

The two architectures.

There are basically two ways to put an AI tutor next to an elementary-age kid (we're mostly thinking about K–5, with the canonical example a 9-year-old doing arithmetic). The first is the tablet path: an iPad with an Apple Pencil, or a ReMarkable, or any of the stylus-on-glass products that have shipped in the last decade. The kid writes on the screen, the strokes arrive as a digital signal in real time, and the tutor lives in the same app. The second is the paper path: the kid writes on an actual sheet of paper with an actual pencil, a camera looks down at the desk, and the tutor reads enough of the work to respond the way a teacher would.

Both can capture handwriting. Both can integrate with a tutor that watches what the child does. The choice between them isn't about whetherthe model can see the work — it's about what surface the child is working on, what the family already owns, and what the design optimizes for when those two things are in tension. The choice quietly determines a lot of downstream design, so it's worth being honest about why we picked one.

Why a stylus seems like the obvious answer.

The tablet path has real advantages and we want to name them before arguing against it. There's no setup — the kid taps an icon and starts writing, the screen lights itself, the page never tilts or slides off the desk, and erasing is a button. The data is also better: stroke-by-stroke pen events (pressure, velocity, writing order) arrive natively from the operating system, where a camera-based system has to infer the same signal from a video stream. And the install base is real — a meaningful fraction of families with elementary kids already own an iPad. Most edtech ships on a tablet for exactly these reasons. On the engineering merits, it is the obvious answer.

Why we picked paper anyway.

Five reasons, in roughly the order they came up in the design conversation.

Screen time, and the second screen.

Earlier AAP guidance for school-aged kids (2016, reaffirmed 2022, updated again in January 2026) has consistently asked families to set limits, protect sleep and exercise, and watch what screen time displaces. The exact phrasing has moved — current AAP framing leans on family media plans, quality and context, and whether media crowds out the rest of childhood — but the directional advice hasn't. Most parents we've talked to already feel like there is too much screen time in the house. Koda still has a screen in the room — the avatar, the XP bar, and the ambient feedback all live on a display — but we deliberately don't make the screen the work surface. The child writes on paper, and Koda handles the feedback. We didn't want the answer to “how does my kid practice math” to be “by adding another screen for them to write on.”

Pencil grip and the conservative default.

The grades-2-to-5 window is still part of fine-motor development, and handwriting is treated as a real school-participation skill in this age band — pediatric OT resources include handwriting interventions for ages 5–21. We found less direct evidence on stylus-on-glass versus pencil-on-paper for grip and handwriting mechanics than we wanted, so we chose the conservative school-default surface: pencil on paper. That's a product default we'd happily revise if someone runs the comparison study; it's not medical advice, and it won't be right for every child. We name it as a default, not as a recommendation about what your kid's hand should be doing.

Cost, durability, and what scales across siblings.

We want to be honest here, because the lazy version of this argument is wrong. Koda is not free hardware. A typical Koda setup is an IPEVO overhead camera (roughly $200), a TOTEM front-facing camera (roughly $120), and a Mac that the family may already own. The iPad path, depending on configuration, is somewhere between $400 and $1,000 for the tablet plus an Apple Pencil. If you are buying everything from scratch, the up-front numbers are in the same neighborhood — the cost argument is not “Koda is cheaper than a tablet,” it's narrower than that.

Where the picture changes is on the two axes we actually care about. First, per-session consumables: a printed worksheet costs a few cents, where a tablet's wear-and-tear cost is a slow drip toward an eventual replacement or repair. Second, scaling across kids: the camera rig on the desk is shared hardware. One kid uses it Monday, a sibling uses it Tuesday, the same setup handles both. An iPad in a household with multiple kids is usually one kid's iPad, and adding a second child often means adding a second device. So the honest cost story is “the up-front numbers are comparable; the per-session cost is lower; and the rig scales across siblings in a way a personal tablet doesn't.”

The “real homework” intuition.

There's a parent-trust dimension we couldn't put a number on but kept hearing in interviews. Math homework that looks like math homework — pencil, paper, the same kind of page the kid uses at school — gets accepted as work by parents and teachers in a way that an app session doesn't always. We aren't claiming a clean piece of research on this; we're reporting what parents said. A kid sitting at the kitchen table doing problems on a worksheet reads as “doing homework.” A kid on an iPad reads, to a non-trivial number of grown-ups, as “on the iPad.” That impression matters for whether the session actually happens in the evenings the parent is tired.

Local-only processing, and a paper artifact.

Koda is not “privacy by absence” — there are cameras, a microphone, profile data, and a local event log on the device. The honest version of the claim is that the processing stays on the device (no cloud round-trip for the model or the camera frames), and the durable artifact of the session is a piece of paper, not a cloud-synced file. If Koda's cameras are off, the worksheet still exists; the kid can still do their homework, the parent can still look at it. With a tablet-based system, the worksheet isthe tablet — if the device is dead or lost, the work is gone with it. The paper path leaves the durable artifact in the child's notebook.

What we give up.

The honest cost ledger, because the post doesn't work without it.

Stroke-by-stroke modeling is harder.A stylus hands you pen events; a camera hands you frames. To know what order the digits were written in, or to flag a long hesitation before a regroup, we have to infer from a video stream rather than read a clean signal off the OS. That's a more expensive engineering path, and the inferences are noisier than the equivalent stylus data. We accept that.

Worksheet setup matters.A camera looking at paper has to actually see the paper. The page has to be reasonably in frame, reasonably lit, and reasonably flat. Koda's worksheet rectification handles ordinary desk geometry — the overhead camera doesn't need to be perfectly square to the page — but it doesn't handle a worksheet that's outside the camera's field of view, or a desk lamp so dim the OCR can't read pencil. Page detection also assumes meaningful contrast between the paper and the desk: a white worksheet on a pale desk surface can confuse the corner finder until you tuck a dark placemat under it. A tablet has no equivalent failure mode. The setup tax is real.

There is a learning curve.The first few sessions involve the family learning where the camera looks, what counts as “the worksheet is positioned correctly,” and how the room should be lit. We try to make that curve short, but we don't pretend it's zero. A tablet app is closer to zero on this axis.

Some kids genuinely prefer the screen.Children with certain sensory profiles, or with fine-motor difficulties that make pencil grip painful, may do better on a stylus. The paper-first default isn't the right default for every child. We name it explicitly because the alternative is to pretend our defaults work for everyone, which they don't.

Closing.

None of this is dogma. If your kid is already fluent with an Apple Pencil and a tablet-based math tutor is meeting your family where you are, that's a reasonable path — Koda just isn't the right product for that household, and we're not going to pretend it is. What we're building is the version of the tutor that lives next to the paper your kid was going to write on anyway, because the screen-time, cost, and “math homework should look like math homework” arguments all stacked in the same direction for us. The decision is a values choice as much as it is an engineering one, and we've tried to be honest about the engineering parts we're paying extra for.

If you'd like the longer cognitive-science argument for hand-on-paper, that's in paper-first pedagogy. If you'd like to see how the cameras are framed at a system level, we wrote that one up too.