Day 4: 576 Experiments, 4 Engines — The Data Behind CPS-0001

#5
by ContinuityLab-Org - opened

We didn't write a whitepaper. We ran 576 experiments.

Most projects in the AI verification space start with a theory. We did it backwards.

We built 4 evidence engines, ran 576 experiments, and published everything — passes, failures, and limitations.

EE-001 — Presence Entropy Score (N=281)

Four dimensions of biological noise: micro-timing variance, noise residual, frequency entropy, perturbation. Cohen's d = 2.1.

Limitation: Single-session recordings. No cross-age or cross-cultural measurement.

EE-002 — Cross-Modal Causal Coupling (N=316)

Tests whether camera and gyroscope streams are causally consistent. Result: 100% temporal alignment when the subject is real.

AI would need to coordinate two independent sensor models in real time — 25x the cost of generating a single video stream.

EE-003 — Challenge-Response (N=200)

Randomized gyroscope challenges with unpredictable timing. 200 trials. 60% pass rate.

We publish the 40% failure rate because honest thresholds matter more than inflated numbers.

VS-001 — Verification Session Pipeline (N=60)

EE-001 passive detection + EE-003 active challenge in a single pipeline. 60 sessions. 93% pass rate.

Demonstrates the Forgery Cost framework: each verification layer multiplies the attacker's cost.

Why publish negative results?
54 subjects. Single-session only. Fluorescent lighting artifacts. No adversarial attack testing yet.

If you only publish what makes you look good, you're not doing research.

Open protocol (Apache 2.0): github.com/myshapeprotocol/myshape-protocol
Open data: HuggingFace ContinuityLab-Org/cps-0001-benchmark
Research hub: thecontinuitylab.org

Sign up or log in to comment