Measurement session
One recording that can settle an argument
Seven blocks, about nine minutes, every frame labelled. It varies the three things that have been confounding every previous recording — how far away you sit, whether your head moves, and which screen you are looking at — and writes down which was which. It also records a channel nothing else here has: the screen emits a faint flicker, and the session measures whether that flicker comes back off your cornea.
Why this session exists
- You can read the monitor above this one and the system says you are fine. That is the case to settle. A screen stacked above sits a couple of degrees outside this one across the bezel, which is below the error of every geometric method that exists — ours and everybody else's.
- Two previous recordings disagreed by a factor of seven about how much the eyes contribute, and the reason turned out to be behavioural: you followed the pointer with your head one time and your eyes the other. Neither file recorded that. This one makes it a column.
- The quiet reading block at the end is the most important one. It is the only thing that bounds the false-alarm rate. Recall measured without it flatters every proctoring system ever built, because a detector that fires constantly catches everything.
The new channel: the screen as a beacon
Every other method here measures DIRECTION — where the eyes point, where the head points, where that ray lands. This one does not. The display is modulated a few per cent, a few times a second; the cornea is a convex mirror and reflects whatever it is aimed at. If our pattern comes back, you are looking at our screen. A different monitor shows different content and carries no such pattern, however close to ours it sits.
The trap a naive version falls into: screen light lands on your whole face, so a modulated display makes your forehead flicker whether you are looking at it or at the ceiling. So the measurement is the eye's modulation relative to the face's — the face is a control that arrives free with every frame. Whether any of this survives an 8-bit sensor, automatic exposure and a dark iris is exactly what this recording is for; it is not yet known, and nothing on this page pretends otherwise.
Emitting at 4% depth, 4 Hz.
Does this room support it at all?
The first real session answered no: the FACE channel — the control, which is lit by the screen whether or not you are looking at it — saw nothing at 4% depth. A bright ceiling lamp was in shot, the face was lit mostly by the room rather than by a 14-inch display, and the camera was running background blur. So measure the depth this room needs instead of guessing at one. Twenty seconds, and it ends either with a number or with an honest "not here".
The two keys
Held, not pressed — and keys rather than buttons for a reason worth stating: a button has to be seen to be clicked, so labelling "I am not looking at the screen" by clicking something on that screen contaminates the very label it sets. A key can be found by touch.
Right now: Looking at this screen
Run it
Not started.
No block running
Press the three buttons above in order.
- Faces
- –
- Yaw
- –
- Pitch
- –
- Eye crop
- –
- Gaze
- –
- Beacon eye:face
- –
- Frames recorded
- 0
The files
The JSON is the one that matters — it carries every channel, every label, and what the screen was emitting at each moment. The video alone shows a face, not what the model made of it.
Reading material for the quiet block
The last block asks you to read normally without looking away. Here is something to read, so that block is a real reading task rather than a minute of staring at a progress bar.
A proctoring system makes two kinds of mistake and they are not symmetrical. Missing a candidate who cheated costs the institution something; accusing a candidate who did not costs that person something, and the record of every public proctoring incident of the last decade is a record of the second kind. ExamSoft flagged roughly a third of one bar-exam cohort. One deployed system detected faces for 78% of darker skin tones against 92% of lighter ones, with six times as many missing-face flags. A United States court has held a room scan unconstitutional.
That asymmetry is why the default in this prototype pauses an attempt and calls a person rather than failing anybody by machine, and why the gaze model refuses to judge at all when its own measured error is too large to accuse somebody with. A system that cannot recognise its own failure is the dangerous kind; one that says "I cannot tell" is merely a limited one, and a limited system that knows it is limited can be deployed honestly.
The hardest case in this whole product is the one you are testing today. A second monitor stacked directly above the first is, from the camera's point of view, a couple of degrees away across the bezel. Every method that works by measuring the direction of a gaze — ours, and every published method, and every commercial product — has an error larger than that gap. Reading the other screen is detectable when the head moves with it. Peeking across the bezel is not. The only way past that ceiling is to stop asking which direction the eyes point and start asking a different question entirely, which is what the flicker on this page is for.