576 Experiments. 4 Engines. One Question.
Three days ago we laid out the problem. Yesterday we showed the math behind our approach. Today we show the evidence — all of it.
═══════════════════════════════════════
THE FOUR ENGINES
EE-001 — Presence Entropy Score (PES)
4-dimensional biological noise analysis. No user action needed — it passively measures whether sensor data carries the statistical signature of a living entity.
N = 281 | Cohen's d = 2.1 | AUC = 0.94
"Cohen's d = 2.1" means the distribution of human sensor data is separated from AI-generated data by more than two standard deviations. That's not a subtle difference. That's a detectable signal.
Crucially: inter-person discrimination is d = 0.04. PES was designed to measure presence, not identity. The null result on identity confirms the design intent.
EE-002 — Cross-Modal Causal Coupling
Temporal binding between independent sensor streams (camera + IMU). The question: do both sensors detect the same physical event at the same moment?
100% alignment across 316 trials.
This is important because an AI generating a video doesn't necessarily produce IMU data consistent with the camera motion. The two streams are physically coupled in reality. The coupling is detectable.
EE-003 — Challenge-Response
Randomized gyroscope challenges with jittered timing. The user must move their device in a specific direction within a time window.
60% pass rate, N = 200.
Only 60%? Yes. And that's expected. Challenge-response is noisy by nature — human reaction times vary, motion varies. But the architecture matters more than the score: a real-time challenge costs the verifier almost nothing, while it forces the attacker to solve a physics problem on demand.
VS-001 — Verification Session
Dual-engine pipeline combining passive PES with active challenge escalation. Protocol-level: hash-chained receipts, Ed25519 signatures, V₁-V₇ verification contract.
93% pass rate, 60 sessions.
The pipeline is the point. Each engine individually has gaps. Combined, they do what individuals cannot — because the attacker faces multiple independent tests simultaneously.
═══════════════════════════════════════
WHAT THIS MEANS
576 experimental runs total. Four independent engines. All on consumer hardware. All data open. All failure cases documented alongside results.
We are not claiming to have solved forgery. We are claiming to have built an architecture where the evidence is public, the protocol is engine-independent, and the direction of the data is consistent.
The next step: external verification. Can someone else — using different hardware, different subjects, different conditions — reproduce these results?
If you have a phone and some skepticism, the data is open.
github.com/myshapeprotocol/myshape-protocol
thecontinuitylab.org
Top comments (0)