Concussion · VORx1
After a concussion, turning your head can make the room slide. The exercise that retrains that reflex only works above a certain head speed — and at home, without an instrument, almost everyone drifts below it and nobody finds out. Gimbal is a webcam dose meter for that prescribed gaze-stabilization exercise. It does not choose your therapy, set your parameters, or diagnose anything. It measures whether the exercise your clinician prescribed was actually delivered — at the velocity that makes it a dose rather than a movement.
No accountNo keyNo upload pathNothing leaves the tab
npm test, on a clean clone, with no flags and no keys.
package.json.
Gaze-stabilization exercise is prescribed with parameters, exactly like a drug: a head-frequency band, a peak-velocity floor and ceiling, a block length, a number of blocks. Two of those numbers are angular velocities, and they are the part a stopwatch cannot see.
The clinic measures none of them at home. The only instrument is the question “how did the home exercises go?”, answered from memory, in minutes-claimed. Every other tool in this space watches, asks, or chats. None of them measure the quantity the prescription is actually written in.
Gimbal does not argue with the numbers on your handout. It holds you to them.
A dose meter that credits everything is a stopwatch. So the interesting surface of this product is the set of things it declines to count — and there are two kinds, which is the whole point.
Reason one · about the movement
Rep not counted — too slow (below 150 °/s; measured 101 °/s).
Proves the velocity is measured and the band is enforced. Both numbers are shown: the threshold that came off your card, and what was actually measured. Never a bare verdict.
Reason two · about the instrument
Rep not counted — tracking unreliable. Try more light.
Head-pose fidelity at 2 Hz on a cheap webcam is the biggest technical risk in this build, and it is answered by refusing to emit rather than by smoothing. Cover part of your face, or dim the room, and the instrument declines to produce a number it cannot stand behind.
Red says you did something wrong. Two of the five refusal reasons are the instrument's condition, not the patient's — and telling someone who slowed down because they got dizzy that they made an error is clinically backwards. A refused rep is drawn as an absence: an unfilled, hatched cell, with the reason in words, and the dose numeral simply does not advance.
And then
No celebration, no streak, no badge, no confetti. In-zone is the resting state; out-of-zone is the only thing that produces change. A reward loop that encourages more of a symptom-limited therapy is clinically wrong, so the quiet is the design thesis rather than a missing feature.
network requests during a full session. Open DevTools → Network. Clear it. Run a block. Not “no personal data”, not “anonymised telemetry” — no requests at all, after the page and its model have loaded.
This is enforced by the browser rather than by a promise. The
Content-Security-Policy below ships as a response header
and as a meta tag on both pages, and connect-src 'self' means
a future commit that added an upload would be blocked, not merely
caught in review.
Content-Security-Policy:
default-src 'self';
script-src 'self' 'wasm-unsafe-eval';
connect-src 'self';
font-src 'self';
object-src 'none';
base-uri 'none'
Permissions-Policy:
camera=(self), microphone=(), geolocation=(), interest-cohort=()face_landmarker model bundle
(3,758,596 bytes) are vendored same-origin and
committed. Nothing is fetched from a CDN, which is what makes the
DevTools proof honest rather than a stunt.
'wasm-unsafe-eval' is the only relaxation in the policy. It is
required by MediaPipe's WASM instantiation and named as such in
METHODS.md, never quietly widened to 'unsafe-eval'.
Permissions-Policy header denies the microphone outright. The
AudioContext is output-only — Gimbal makes sound, it never listens.
Finish a session — or press Esc, which ends it instantly with no confirmation and no penalty, because stopping on symptom provocation is the clinically correct behaviour — and you get the page a physical therapist actually receives.
Verbatim, at body size, in the same words the printed report uses. This block is
byte-compared against src/report/limitations.ts by a check in
npm test, so the page and the artifact cannot drift apart.
No bench-agreement figure appears anywhere in this project, because none has been recorded. The intended validation is a phone gyroscope logged rigidly at the temple while the webcam runs Gimbal, the two traces aligned by cross-correlation and compared per cycle. The tolerance for that comparison is already committed, and it was chosen so that deleting the central-difference bias correction makes it fail.
Two instrument thresholds — qFloor and
deadbandFraction — are likewise provisional.
PROVISIONAL_FROM_SPIKE in src/dsp/limits.ts names
exactly which ones, and the report says so on its face rather than presenting
a guess as a calibrated threshold.
Until that recording exists, the honest statement is that it does not exist.
npm test@mediapipe/tasks-visionnpm auditnpm ci
npm test # unit tests + the source checks
npm run check:build # builds the bundle, then greps it
No flags. No environment variables, no keys, no account, no
seeded state, no --dry-run. The judged capability executes for real
on the default path or it does not execute at all, and a check in
npm test fails the build if one of those tokens ever appears in
README.md or DEMO.md.
sin(2π·2·t) at 30 fps with a bin width of exactly
30/256 = 0.1172 Hz — and it returns
no peak, not NaN, for an all-zero input.
/app?demo fills the eight fields from a set of numbers
published for evaluation, shows a persistent banner labelling them as an
example, and stamps EXAMPLE beside every value. It does
not tick the clinician gate for you, and the values print
on the report labelled as an example rather than as clinician-sourced.
/app with no query string is the real entry path and is still
completely empty.
localStorage and is cleared with one button. There is no
cross-device history and no upload path of any kind — see section 03 for
how to verify that yourself in fifteen seconds.
Thirty seconds is enough
Open it, turn your head gently, and watch the ledger acquire a hole with your name on the reason. Then speed up into the band, and watch nothing at all happen.
Not a diagnostic device. It supplements your clinician; it does not replace them.