One honest number per fist.

What we measure

The camera finds 21 hand landmarks per frame (Google MediaPipe, running in your browser on the GPU). From them we take one headline reading: the distance from the index fingertip to its knuckle, divided by the palm length (wrist to middle-finger knuckle). Dividing by your own palm makes the number independent of how close you sit. Lower means a tighter fist.

Each capture is the median of a steady window: the pose has to hold for 0.7 s (about 20 camera frames) before it counts, so one blurry frame can’t move the result.

Open hand

0.766

Fist

0.183

Fist closure

1 − fist ÷ open

Validation run 4, 20 October 2026, one adult right hand: the open and fist readings above are that run’s medians. The weekly report draws one gauge per day from the fist closure.

Form guards

A 2D camera can be fooled, so the coach refuses a capture until the form is right, and says what to fix:

• Palm to the camera. If the knuckle width drops below 0.45 × the palm length, the hand is side-on and 2D distances lie. Cue: "Turn your palm to face me."
• Thumb outside. If the thumb tip is closer than 0.6 × the palm to the little-finger knuckle, the thumb is tucked inside and the fingers can’t close. Cue: "Thumb across the outside." Tucked thumbs tripled the noise in run 3; the guard brought it back down in run 4.

The noise band (MDC95)

Make the same fist ten times and the reading still wobbles. In run 4 the ten re-formed fists had a standard deviation of 0.023. The minimal detectable change at 95 % confidence is 1.96 × √2 × SD = 0.063, which is 10.7 % of the open-to-fist range. A day inside your usual reading ± 0.063 is normal day-to-day noise; only a change beyond it is called real.

• Your usual is the median of your earlier sessions. Flexion needs at least 3 before it calls any change.
• A flare is two days in a row looser than your usual by more than the band. The coach then asks whether to switch to the gentle routine, and nothing changes until you say yes.
• Progress only rises after 3 complete, pain-free sessions and a week at the same level, by at most 10 %, and never past the limit on your therapist’s sheet.

Limits, said plainly

• The guard thresholds and the noise band come from one person’s right hand. They need more people before they mean anything general.
• The MDC95 uses the spread of repeat trials as the measurement error, a shortcut; a test-retest study (ICC-based) should replace it.
• These are ratios of a 2D image, not joint angles, which is why the site never shows degrees.
• Speed: 30 fps from a laptop webcam in Chrome; 3–5 ms per frame for the landmarks alone (Day-1 gate G1), 7–12 ms with gesture recognition (the current engine).

What leaves your device

The video never leaves your browser tab. Only the numbers of a finished session (readings, reps, the grip result, form fixes) are saved, under an opaque user id with no name.