PES Benchmark v0.2

Part 1 — What We Measured

The Presence Entropy Score (PES) is a 0-to-1 metric that quantifies the depth of biological entropy in a motion sample. It operates across four independent dimensions:

Micro-timing variance. The frame-to-frame jitter in human motion. Biological motor control introduces irreducible timing noise — muscles do not fire at perfectly regular intervals. Synthetic motion, generated by algorithms, tends toward periodic regularity.

Noise residual. The difference between observed motion and a smoothed trajectory. Human motion contains stochastic perturbation at every joint. AI-generated trajectories are typically smoother than any human can produce.

Frequency entropy. The spectral spread of motion across frequency bands. Human motion distributes energy broadly. Synthetic motion concentrates energy in narrow frequency bands — a consequence of interpolation and smoothing in generation pipelines.

Biological perturbation. Low-amplitude, high-frequency oscillations characteristic of living tissue. These micro-tremors are involuntary and present in all human motion, including deliberate stillness. AI generators do not model them.

This benchmark evaluates whether these four dimensions, individually and in aggregate, separate human motion from synthetically generated motion with statistical significance.

Part 2 — Dataset

The v0.2 benchmark dataset consists of 281 motion samples across two classes.

Human motion (81 samples). Collected from 54 distinct human subjects performing unstructured 30-second motion sequences captured via standard webcam at 30 fps using MediaPipe Pose (33-point landmark model). Subjects were instructed to move naturally — no choreography, no calibration, no controlled environment. Lighting, background, camera angle, and clothing were not standardized. This ecological variability is intentional: the benchmark must reflect real-world conditions.

Synthetic motion (200 samples). Generated using four distinct AI strategies (50 samples each):

1. Random walk. Each joint independently follows a Gaussian random walk with per-joint variance calibrated to match human joint range of motion. Produces plausible but statistically flat motion.

2. Spline interpolation. Keyframes sampled from human pose distribution, connected via Catmull-Rom splines. Produces smooth, continuous trajectories that pass through human-like poses.

3. GAN-generated. A conditional GAN trained on human pose sequences. Produces motion that is visually convincing but lacks the micro-perturbation signature of biological tissue.

4. Near-static. Minimal joint displacement around a resting pose with Gaussian noise injection at increasing power levels. Tests whether the PES can distinguish between genuine stillness and synthetic stillness.

Part 3 — Results

Aggregate separation. The weighted PES (four dimensions combined) produces a clean separation between human and synthetic motion:

In plain language: a score of 2.1 on Cohen's d means the human distribution and the synthetic distribution are separated by more than two standard deviations. An AUC of 0.94 means that if you pick a random human sample and a random synthetic sample, the human scores higher 94% of the time.

Per-dimension separation. Each of the four PES dimensions independently separates human from AI motion:

Micro-timing variance: Human mean 0.72, Synthetic mean 0.18. Gap: 0.54. Human motion is measurably more irregular in frame-to-frame timing.

Noise residual: Human mean 0.68, Synthetic mean 0.22. Gap: 0.46. Human trajectories contain irreducible stochastic noise; synthetic trajectories are visibly smoother.

Frequency entropy: Human mean 0.74, Synthetic mean 0.25. Gap: 0.49. Human motion distributes energy broadly across the spectrum; AI motion concentrates it.

Biological perturbation: Human mean 0.61, Synthetic mean 0.12. Gap: 0.49. The micro-tremor signature is the hardest for AI to simulate — current generators do not model involuntary physiological oscillation.

Classification threshold. Based on the empirical distributions, we set the human-vs-AI classification threshold at PES = 0.40 (Decision DL-001). All 54 human subjects scored ≥ 0.41. All four synthetic strategies scored ≤ 0.38. The 0.03 gap is the current entropy margin — the irreducible separation between biological and synthetic motion under current generation models.

Multi-source AI results. Four generation strategies produce PES scores consistently below the human floor:

Random walk: PES = 0.31 · Spline interpolation: PES = 0.38 · GAN white noise: PES = 0.22 · Near-static: PES = 0.08

The highest-scoring synthetic strategy (spline interpolation, 0.38) still falls below the lowest-scoring human subject (0.41). No overlap in the current sample.

Part 4 — Threats to Validity

We identify the following limitations of the current benchmark. These are not excuses — they are the starting conditions for v0.3.

1. Sample size.54 human subjects is sufficient to establish a large effect size (Cohen's d = 2.1) with statistical significance, but it is not sufficient to characterize the full distribution of human motion entropy across age, mobility, neurological conditions, and cultural movement patterns. The benchmark may overestimate discriminability for populations not represented in the current sample.

2. Ecological validity. All samples were recorded in a single session per subject. We have not measured intra-subject variability across days, moods, fatigue levels, or hardware. A subject who scores 0.72 on Tuesday may score differently on Friday. Longitudinal stability is an open question (see Research Roadmap: 2026 Q4).

3. Adversarial robustness. The four AI generation strategies are representative of current off-the-shelf techniques. They are not adversarially optimized to defeat the PES. A motivated attacker with knowledge of the PES architecture could potentially design generation strategies that narrow the entropy gap. We have not tested against this threat model.

4. Replay attacks. The current benchmark evaluates de-novo motion generation. It does not evaluate whether a captured recording of genuine human motion, replayed through a virtual camera, would pass the PES. Replay resistance is a separate investigation (see RN #003, in pipeline).

5. Sensor variability. All samples used standard 30 fps webcam input. Lower frame rates, higher compression, or different landmark models (e.g., BlazePose vs. MediaPipe Pose) may shift the PES distributions. Hardware calibration is not yet part of the protocol.

Part 5 — What This Does Not Prove

It is important to state what this benchmark does not establish:

It does not prove continuity.Distinguishing human motion from synthetic motion at a single moment is presence detection, not continuity verification. The PES tells you “this is a human right now.” It does not tell you “this is the same human who was here ten minutes ago.” Closing this gap is the central research question (see OQ-001).

It does not identify individuals. The PES is not an identity verification system. It measures entropy — a population-level property of biological motion — not identity. Two different humans with similar motor characteristics may produce indistinguishable PES scores. This is by design: the protocol separates presence verification from identity attribution.

It does not provide a security guarantee. The current entropy margin (0.03) is empirical, not theoretical. As AI generation improves, the margin may narrow. A theoretical lower bound on human entropy — derived from the physics of biological tissue — would provide a security guarantee. We do not yet have one.

Part 6 — Next Steps

The v0.2 benchmark establishes that the PES captures a real and measurable signal. The next phase moves from “does this work?” to “under what conditions does it fail?”

v0.3 (2026 Q3).Expand to 300+ human samples with explicit coverage of age, mobility, and hardware diversity. Publish per-subgroup analysis. If Cohen's d remains above 1.5 across all subgroups, the signal is robust. If any subgroup shows d < 1.0, the benchmark has identified a boundary condition — and boundary conditions are more valuable than confirmation.

Replay benchmark (2026 Q4). Capture genuine human motion, replay it through a virtual camera, and measure whether the PES detects the replay. This is a harder problem: the motion itself is genuine, but it is not continuously present. It requires temporal analysis beyond single-frame entropy scoring.

Longitudinal stability (2026 Q4). Measure the same subjects across multiple sessions, days apart. Does intra-subject PES variance exceed inter-subject variance? If yes, continuity verification requires session-to-session signal stability that the current benchmark does not assess.