A Transaction Science Open Standard
Access is an open standard for people the usual channels fail. Express builds an utterance and speaks it. Read takes a document in and eases it back out. Both build on one kernel, access-core, which defines how a person reaches the screen: eight access methods, from touch and switches to eye gaze on an ordinary webcam. Apache-2.0.
Two surfaces
One person may need to say something and another to take in something — often the same person, on different days. Both are access to language. Both are here, live, free.
AAC — say it
Build an utterance and speak it. A symbol-and-text board you drive however you can reach it; the words come out in a voice. For when speech itself is the channel that fails.
Lumina — take it in
A document comes in; it is eased back out. Read-aloud, a line highlight that follows your gaze, and one-word-at-a-time pacing — so a page meets you where your reading is.
On languages, plainly: Express offers 45. English has a core vocabulary drawn from studies of English speakers. The other 44 use glosses of that English list — translations of its words, not a core vocabulary studied with speakers of each language — and the standard labels them as glosses. In September 2026, 538 of those glosses were corrected.
One kernel
The hard parts of access — how you select, how gaze becomes a point on the screen, how a voice is configured, how a result is accounted for — live in a single surface-agnostic kernel, and each surface re-exports it. A fix to a part lands in every surface that uses that part. It depends on four crates (serde, serde_json, thiserror, web-time) and embeds no model runtime. Gaze and voice are clock-free: the caller passes the time in. Dwell is the one part that keeps time, reading a monotonic clock (through web-time, so the same code runs natively and in a browser).
How a person reaches the screen: the eight access methods and a motor profile, switch scanning over a grid, dwell selection, and display preferences such as contrast, type size and reduced motion.
Screen and gaze geometry, calibration data, the iris math for a 478-point face mesh, a Kalman smoothing filter, and an estimator that turns face landmarks into a screen point.
Voice settings, speech requests handed to whatever speech engine the surface has, and voice banking. Clock-free: the caller passes in the time a recording was made.
An honest record of how a result was produced — measured, model-based, or estimated — and the energy it took, in microjoules. Transparency, never a bill.
Surfaces are meant to ride the kernel rather than fork it. A new surface — a sign-board, a switch-only keyboard, a captioner — is a new face on the same access methods, gaze, voice, and receipts.
Eight ways in
Access is not one motion. The kernel defines eight access methods, each with its own recommended target size, and a person can move between them as the day or the year goes.
Finger or stylus on the screen. The direct path — large targets, forgiving hit areas.
Mouse or trackpad. The most precise path, so targets can be smallest here.
One switch, two, or more. Choices are highlighted in turn and a press selects — the whole screen reachable from a single button.
A joystick or directional pad moves between targets, one step at a time.
Look at what you want. A webcam reads where your eyes land, and dwell selects. No mouse, no touch, no special rig.
Head movement, read by a camera or a sensor, moves the pointer; dwell selects.
For people with some speech: the words and sounds they can say, used as input.
Brain–computer interface, defined for research and emerging devices, with the largest targets of all.
Dwell is a selection technique, not a method: rest on a target and it activates, no click needed. Eye gaze and head tracking select this way. Hold time is yours to set. From your own use, the kernel can suggest a longer or shorter hold, which you accept or decline; automatic adjustment is there too, and off unless you turn it on. Switch scan speed works the same way.
Express offers six in its setup — touch, pointer, switch scanning with one or two switches, joystick, eye gaze, and head tracking (which, for now, runs the same webcam eye-gaze pipeline) — and runs the kernel's own scanning and dwell. Read takes its gaze math from the kernel; its scanning and dwell are still its own. Partial voice and BCI are defined for any surface to adopt; neither reference surface offers them yet.
Eye gaze, on a webcam
Eye gaze has long meant a dedicated camera bar and a setup fee. Here it runs on the webcam that ships with the laptop. Both Express and Read take their iris math from the kernel, so the eyes are read the same way in either surface.
The camera you already have. Express opens it only when eye gaze (or head tracking) is your access method; the video never leaves the device.
MediaPipe finds a 478-point face mesh, including the iris, in each frame — in the browser, on the GPU.
The iris position between the eye corners and lids becomes a look direction. This is the kernel's math, and both surfaces use it.
The kernel's estimator maps that direction onto the window and smooths it against jitter. Given four or more calibration points it interpolates between them instead; Express does not collect any yet.
In Express the point lands on a cell of the board, a ring fills while you rest there, and dwell selects it.
The video is read frame by frame in the browser. Nothing is uploaded; the camera stops when you stop.
Express boots and works without a network. Eye gaze is the exception: when you switch it on, the eye-tracking library (MediaPipe Tasks Vision, pinned at 1.0.1) comes from jsDelivr and the face model from Google's model storage. The app's offline cache does not hold them, so plan on a connection for gaze.
The iris math lives once, in the kernel, so a fix to it reaches both surfaces. Express also runs the kernel's estimator; Read still maps and smooths gaze its own way, with its own calibration step.
An open standard
Access is one of the Transaction Science open standards. The point is not an app to sell but a kernel anyone can carry — so access to language is a shared foundation, not a feature someone owns.
The kernel and both reference surfaces are permissively licensed. Read it, build on it, ship your own surface — no fee, no asterisk.
Express and Read are not the product behind a paywall; they are the standard, demonstrated, live and free. They show one way to build on it.
Every result can carry a receipt — measured, modelled, or estimated, in microjoules. It is there for transparency and is never billed.
The kernel bundles no model runtime: callers supply the face landmarks, and for gaze and voice the time as well. Dwell alone keeps time, from a monotonic clock that works natively and in the browser. That keeps it small, portable, and easy to embed.