What this is testing. Not the games. Whether browser input latency
and resolution are good enough that muscle memory built here transfers to a real geared
or remote head. Watch the smoothness figure, not the score.
How smoothness is measured. The metric is
log dimensionless jerk, the standard smoothness measure from the motor-control
literature: LDLJ = −ln( (T³ / v_peak²) · ∫|d²v/dt²|² dt ). Dividing by
duration cubed and peak speed squared makes it independent of how long and how fast a
move was, which is what lets one grade mean the same thing in Lander and in Centipede.
No per-drill constants exist.
Why your input device matters. Jerk is the third derivative of
position, so any quantization in the input is amplified enormously. Movement is sampled
on a fixed 100 Hz grid and low-passed at 8 Hz before differentiation — voluntary movement
lives below about 5 Hz and physiological tremor at 8–12 Hz, so nothing above that is
signal. Residual quantization bias is removed with a correction curve computed at startup
from your device's own step size. Below roughly 120 counts per second the measurement is
genuinely impossible, and the app says so rather than inventing a grade. If the
Measured figure is low, raise the CPI or resolution setting in your wheels' own
driver.
The two constants. Grades map LDLJ onto 0–100 between
−11.0 (F) and −5.0 (A). These are the only tunable numbers in
the metric, they are global rather than per-drill, and they are provisional — fitting them
against real geared-head performance is exactly what this prototype exists to do. Every
round stores its raw LDLJ, so they can be refitted later without invalidating history.