It's reading the camera, not just you
An Omoggle-style score comes from landmark positions detected in a video frame — eye corners, jaw edge, cheekbones, nose bridge. Move the camera three inches closer, tilt your chin down a few degrees, or change the room's lighting, and those landmark positions shift in the frame even though your face hasn't changed at all. The model can only read what the lens hands it.
The usual suspects, in order
Across repeated scans, four variables account for most of the swing: distance (changes apparent width-to-height ratios), angle (a few degrees up or down distorts jaw and midface reads), lighting (overhead-only light casts shadows a landmark model can mistake for structure), and expression (a slight smile or squint introduces asymmetry). Jaw- and symmetry-related measurements tend to be the most sensitive of the group.
What actually confirms it
The way to find out whether a swing is real or just camera noise is to hold every variable fixed except one, then change only that one — the practice guide covers what to change first and why. If the reading moves consistently with the variable you changed, that variable was the cause. If it still swings with everything held fixed, that's a different question. Guessing between scans on a live service doesn't let you isolate anything — you need a private place to test one variable at a time and see the same six measurements each time.
Mogalytics is that test bench
Mogalytics reads six camera-visible geometry ratios — eye line, jaw ratio, frame symmetry, midface, cheek ratio, and bridge — into a weighted composite over a 30-frame trimmed-mean window, so a single blink or head turn doesn't skew a saved reading. Change one variable, save, change the next, save again, and read the pattern across sessions instead of any one scan. Camera frames never leave your browser.