salluru.dev
operational

All systems active

Labs / Proof-of-Human
Launching

humanproof

Was this input stream produced by a human neuromuscular system?

Detect AI input from the human neuromuscular signature — generalizes to bot and fraud detection.

At a Glance
5-layer
SDE pipeline
1/f³
noise kernel
SPRT
detector
bot · fraud
generalizes
The Problem

Behavioral anti-cheat and bot detection classify on aggregate statistics — mean speed, snap acceleration, reaction time. Modern AI is trained to reproduce exactly those statistics, so the whole defensive layer is being spoofed. As agents learn to drive UIs convincingly, “is this a human?” breaks everywhere — games, fraud, RPA, identity.

Key Insight

Model the causal generative process of human motor control, not its output histogram. Human movement is the output of a stochastic optimal-control loop with signal-dependent noise and a 1/f³ correction spectrum. An AI matching aggregate stats still can't reproduce that conditional covariance kernel without running a real-time closed-loop controller — which costs far more than the Bezier smoothing that defeats today's classifiers.

How It Works

Per-session SDE

Fits a hidden-state stochastic differential equation of the human plant (inertia, feedback delay, signal-dependent noise) online to each session.

Likelihood-ratio test

Scores each trajectory: human-plant model vs best-fit smoothed-AI model, via a sequential probability ratio test with controllable false-positive rate.

Red-team harness

Trains generative humanizers against the detector and measures the cost-to-evade — turning the arms race into a quantified, defensible number.

One engine, many markets

The same “human vs machine control” question sells into bot/RPA detection, fraud, and continuous CAPTCHA-replacement auth.

Proof-of-HumanSecurityIntegrity