Window & tab
Page Visibility, window focus, full-screen exit, second display, clipboard, idle time, and a heartbeat that notices its own absence. No permissions, works on every platform, prevents nothing.
Detection: reliable · Prevention: none
Prototype · runs entirely in your browser
Browser proctoring, with the ceiling drawn in
Three things a plain web page can watch — the tab, the window, and the candidate's face — wired to a policy engine that warns for five seconds and then decides. Switch the signals on and off, sit the exam, and try to beat it. It is not hard to beat, and that is the most useful thing this prototype has to say. The second most useful is what it does with a violation by default: it stops the clock and calls a person, rather than failing anyone by machine.
Camera
MediaPipe Face Landmarker on the device: face present, face count, head yaw and pitch calibrated against the candidate's own neutral pose, iris direction, eye openness. No frame leaves the page and no emotion is inferred.
Attention monitoring · not eye tracking
Gaze
The model's ten iris landmarks, a nine-point calibration, and a fit scored by predicting each target it was not trained on. Good enough for "on the exam or off it"; nowhere near a hardware eye tracker — and it reports which of those it managed, in numbers, for your face and camera.
Measures its own error · refuses to judge when that error is too large
The screen as a beacon
Every method here — ours and everybody's — measures which DIRECTION the eyes point, and a monitor stacked directly above this one sits a couple of degrees away across the bezel, under the error of all of them. So this asks a different question: the display is modulated a few per cent, the cornea mirrors whatever it faces, and if our pattern comes back the candidate is looking at our screen. A different monitor shows different content and carries no such pattern, however close it sits.
Measurement session · unproven, and built to find out
A head we already know the angle of
One photograph becomes a rotatable 3D head, rendered back into the same detector an exam uses. The angle it reports is compared with the angle that was drawn — which is exact, because it was typed in. It has already earned its keep: it settled the yaw sign convention nobody had ever checked, and proved that moving only the eyes leaves head pose alone.
Head rig · ground truth, not a person's best effort
Policy
Grace, then a five-second warning with an audible alarm that is heard even from another tab, then penalty points, a pause for a human, or termination. Seven presets from "never acts" to "one strike", all editable and shareable as a link. The strict ones exist here to show what they do: inside a platform, every one of them pauses instead of ending the attempt.
Default: pause and call a proctor, never auto-fail
Where this sits against the things it will be compared with
Two different markets get confused here, and the honest answer is different for each. The proctoring vendors are ahead on integrations, scale and live human proctors. The webcam eye-tracking vendors are ahead on calibration craft. Nobody is ahead on saying what their own error is.
| Product | Gaze calibration | Validation | Reports its own error |
|---|---|---|---|
| Proctor Lab | 9 points, 4 of them with a head-pose target | 4 fresh points after the fit closes | Per candidate, per camera, and refuses to judge above 33% |
| RealEye | 39 points by default (5 → 137 px, 39 → 106 px, 78 → 92 px) | 3 points, session refused above 150 px | Publishes aggregate accuracy; virtual chinrest and live pose/distance monitoring |
| Eyedid (ex-SeeSo) | 1–5 points | — | Exposes an OutofScreen boolean; commercial licence |
| Tobii (hardware) | 5–9 points, with a positioning box and a distance bar | Yes | Hardware-class accuracy, and not a webcam |
| Proctorio, Honorlock, Talview, Proctortrack | None published | — | Identity and face presence; gaze is reported as a flag, not as a measurement |
Read the last row carefully before repeating it: it says their published material, as read in September 2026, describes no gaze calibration step, which is not the same as proving none exists. The row that matters commercially is the last column. Every one of these products will produce a number; this one is built to tell you when its own number is not good enough to accuse anybody with, and that is the only claim here worth defending in front of a regulator.
Read this before demoing it to anyone
- Detection is not prevention. A plain browser produces evidence. Blocking other applications needs a browser extension or a locked-down browser like Safe Exam Browser.
- A second device is invisible. A phone next to the laptop is outside every browser API, and outside most of the paid products too.
- "Eye tracking" is three different problems. Iris direction relative to the head is free. Which region of the screen you are looking at needs a nine-point calibration and is realistic. A precise gaze point, like a hardware tracker, is not achievable from a webcam. The gaze lab runs the middle one on your own face and tells you its measured error. Held up to a still photograph the error comes out at more than half the screen — the automated suite measures 50–56% as the calibration stands — and the model refuses to place a gaze point or call anything off-screen at all. That refusal is the claim worth checking, not the figure: the figure moves whenever the calibration changes, and an earlier version of this page quoted one from a design that no longer exists.
- Mobile cannot be locked down. App switching is detectable; screenshots, split view and a second app are not. Full-screen lock is not a control on iPhone.
- Auto-failing is the dangerous part. ExamSoft flagged roughly a third of candidates on one bar exam; 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 US court has ruled a room scan unconstitutional. Flag for a human, do not fail by machine.
- Getting into a platform means a connector per platform. 1EdTech's LTI Proctoring Services on LTI 1.3 is the open standard and this tool implements it, but as of September 2026 no LMS is certified for it. Open edX accepts it only through 2U's edx-exams service, which gates the exam and never pauses it; Blackboard Learn Ultra documents it; Canvas, Moodle and D2L do not implement it at all. So Moodle needs a quiz access plugin, Canvas an exam delivered in our own tool, and Mapache test the embed. The embedding demo shows that route — one script tag and a framed monitor — for a host that owns its own page.
- The EU treats this as high-risk. Exam proctoring is an Annex III high-risk use under the AI Act, and inferring emotion in education has been prohibited outright since February 2025.