Mouse motion violates the 2/3 power law (speed independent of curvature — synthetic path)
bh.power_law_violation ·
What it catches
Lacquaniti 2/3 power law (1983). v0.74.25 SECOND-SOURCE corroboration + FP-safe tightening (task biomech-corroborate, docs/behavioral-data.md): the floor was single-source (10-user Balabit) and the doc's 'p1=0.116' was a PER-SEGMENT number, but the detector fits one β over the WHOLE pointer stream — different quantity. Re-measured with the SHIPPED extractor: on long/clean captures (Balabit whole-session, SapiMouse 128-pt blocks) FP@0.1 is 0%, but on very short pause-contaminated windows Balabit's worst individual reached ~2.9% — the 0.1 floor sat AT the short-capture human p1, not above it. Added an INDEPENDENT second corpus — SapiMouse (Antal et al., 120 subjects, a different rig/era) — and measured β with the same biomech.power_law code: human p1≈0.384, median≈0.49, 0% of samples and 0/120 subjects below 0.1 (or 0.05). Both sources agree real aimed motion obeys the power law far above the floor; threshold tightened 0.1→0.05 to add margin against the only elevated-FP regime (short Balabit windows) at ZERO recall cost — a constant-velocity/Bezier synthetic path is β≈0 (corpus bot -0.004 still fires; the lab humanizer at β≈0.45 still correctly does not), and no corpus session sits in (0.05,0.1) so labels are unchanged. Corroborating-only (behavioral never convicts); gpu/Tier-3-real-device motion still the open frontier.
Signals it reads
behavioral.power_law_exponent
How it fires
below_threshold (threshold 0.05)