Modulate
Each canvas object is driven by a circular shift of one shared pseudo-random binary sequence. One base code, one calibration, every target.
Wega Labs · Our thesis
Attention, not gesture.
Blinks and nods are voluntary movements. They fatigue, they are visible, and they cap out at a handful of distinguishable commands. There is a paradigm with thirty years of literature behind it that has no such ceiling, and it needs a stimulus grid we already happen to own.
Paradigmc-VEP · 100–260 bits/min
The paradigm / 01
Each target carries its own signature.
In code-modulated visual evoked potentials, every selectable target on a screen is modulated by its own pseudo-random binary sequence. When a person attends to one, their visual cortex produces a response time-locked to that target’s code and to no other.
Cross-correlate the recorded signal against each code, and the target the person was looking at falls out of the arithmetic. There is no imagined movement to train, and no gesture to perform.
The user does not move. Attention alone is the input.
Information rate / 02
The highest-bandwidth non-invasive paradigm published.
Motor imagery
~15 bits/min
Imagined movement, extensive per-user training
Blink / gesture selection
~20–40 bits/min
Visible voluntary movement, fatigues quickly
c-VEPour direction
100–260 bits/min
Occipital electrodes, precise stimulus timing
Published c-VEP results report 0.77–0.84 accuracy across subjects and 0.98 for the best performer. These are literature results, not ours.
Pipeline / 03
From a modulated glow to a committed instruction.
Attend
The operator looks at the object they were already thinking about. No blink, no nod, no dwell timer, nothing visible to anyone watching.
Record
Occipital electrodes at the visual cortex. The canonical minimal montage in the literature is three: Oz, Fz, and a forehead ground.
Correlate
Cross-correlate the recorded signal against each object's code. The attended target produces a response time-locked to its code and to no other.
Rank
The winning object becomes the referent. A model proposes candidate actions for that object, because it knows what the object is.
Commit
One more selection carries the verb. Two selections, roughly four bits, and a specific instruction has been issued to a specific agent.
Two selections, roughly four bits: one to say which object, one to say what to do with it. The model supplies the rest, because it already knows what the object is.
Cost of entry / 04
The barrier collapsed and almost nobody noticed.
₹2,600≈ $30
One publication-grade single-channel biopotential amplifier, manufactured in Bengaluru.
The canonical minimal setup in the c-VEP literature is three electrodes — Oz, Fz, and a forehead ground — reporting around 100 bits per minute on half a second of data. That is one single-channel biopotential amplifier.
Not the fifteen hundred dollars a consumer headset commands. Serious work in this area is now accessible to far more people than realise it, and we intend to demonstrate that publicly.
Plate IV / 05
The target is chosen before any word exists.
A recorded scanpath over a canvas: a fast survey, a return, then a long commit fixation on the object the operator actually meant. Overt attention is already doing the addressing — c-VEP is a way of reading a decision the body has already made.
FIX07/07
ADDRESSterminal_01
DWELL210 ms
PHASEverify
ELAPSED2150 ms
Why addressing is the expensive part / 06
Most of a command is pointing.
This is the cost c-VEP removes. Point at any block and watch what the sentence on the left has to become in order to mean the same thing.
Demonstration A
The address problem
Point at any block on the canvas. Watch what the sentence on the left has to become in order to mean the same thing.
Point at a block to compose the message.
GAZE→ —
Typed—
vsSpoken—
Every word beyond “dark mode” is the cost of not being able to look.
Open problems / 07
Four things standing between this and a product.
None of these are hypothetical, and two of them have nothing to do with neuroscience. That is the part of this work we expect to be hardest.
Stimulus timing is the real adversary
Software timestamps lie. A 60 Hz display quantises to 16.7 ms and frame delivery jitters unpredictably, and c-VEP decoding depends entirely on knowing when a glow actually reached the eye. This is why Stage 02 is a photodiode and no participant.
Legible stimuli versus strong signal
Response amplitude scales with luminance contrast, so the strongest signal comes from the ugliest interface. There is a genuine tension between a decodable canvas and one anyone wants to look at, and we do not yet know where it resolves.
Target count against code separation
Published forty- and hundred-target spellers come from tuned laboratory rigs. Adding targets means codes bleed into one another. Our benchmark starts at four targets; the larger systems remain literature references.
Calibration nobody will sit through
Shared-code circular shifts are what make a single calibration generalise across targets. Getting that calibration toward zero is the difference between a research demo and something a person uses on a Tuesday.
Claim limit / 08
We have not run a c-VEP session. Stage 03 is where that begins.
Every figure on this page — 100–260 bits per minute, accuracies of 0.77 to 0.98, the three-electrode montage — comes from published literature and is cited as such. None of it is a measurement of our system, because our system does not yet exist.
What exists today is Stage 01: selection with no sensor at all, and a counter that reports what it costs. Stage 02 is a photodiode. Neither involves a participant.
The substrate
The grid is a product that already ships.
October is the spatial canvas where agents, screens, and terminals are addressable objects. It was built to supervise agents, and it turned out to be the precondition for controlling them without moving.