Motion as Computation
How Comp-Core turns raw sensor data into justified N'Ko statements through an 8-layer architecture.
The premise is simple: motion carries information. When you move, you're not just displacing mass through space—you're expressing state, intention, and meaning. The challenge is extraction: how do you turn continuous, noisy sensor data into discrete, justified statements?
The problem with traditional approaches
Most motion analysis systems treat the pipeline as a classification problem. Take sensor data, extract features, run a classifier, output a label. "Walking." "Running." "Sitting." This works for coarse categories but loses everything subtle.
The real question isn't "what activity is this?" but rather: "what is the trajectory doing right now, and how confident are we?" We need a representation that captures dynamics, not just snapshots.
Key insight
Motion is not a sequence of poses. It's a trajectory through a semantic space, with structure that can be inscribed as justified claims.
The 8-layer architecture
Comp-Core processes motion through eight layers, each with a specific responsibility. Data flows upward from raw sensors to high-level synthesis, with feedback loops that allow higher layers to influence lower ones.
The 5-stage motion pipeline
Layer 2 is where raw sensor data becomes coherent motion. The pipeline has five stages:
- 1.Time Normalize
Align timestamps across devices. Handle clock drift.
- 2.Resample
Uniform sample rate (60Hz). Cubic interpolation.
- 3.Coordinate Unify
Transform to canonical skeleton. Mocopi → MediaPipe → Watch space.
- 4.Fuse
Extended Kalman filter. Confidence-weighted sensor fusion.
- 5.Window
Create MotionWindow (deterministic, replayable, hashable).
The output is a MotionWindow: a self-contained, deterministic representation of motion over a time interval. It's replayable (same input → same output) and hashable (content-addressable for caching).
From motion to inscription
Layer 3 is where the magic happens. The Semantic layer takes MotionWindows and produces inscriptions: typed claims about what the motion is doing, written in N'Ko script.
Why N'Ko? It's a unified script for Manding languages, designed to express meaning that doesn't map cleanly to Latin or Arabic. Using it as the surface language for motion inscription honors that legacy—the machine doesn't speak in borrowed tongues.
Example inscription
ߛ ⟦100.0–200.0⟧ : z(σ) ↓ ; home ; c=0.85"Stabilization from t=100 to t=200, dispersion decreased, at place 'home', confidence 0.85"
The 10 sigils
Each sigil is a detector for a specific trajectory pattern:
Stabilization
Dispersion
Transition
Return
Dwell
Oscillation
Recovery
Novelty
Place-Shift
Echo
Performance: 2.16ms end-to-end
The entire pipeline—from raw sensor packet to justified inscription—runs in 2.16ms average. That's 18× under our 40ms target for real-time applications. Cross-modal coherence is 99.94%, meaning inscriptions are consistent across different sensor sources observing the same motion.
What this enables
With motion as justified computation, we can build systems that:
- →Remember movement patterns across sessions (5D trajectory coordinates)
- →Generate music-aligned motion (cc-motiongen diffusion)
- →Anticipate what comes next (commitment, uncertainty, transition pressure)
- →Verify claims about motion with provenance chains
The body becomes a source of justified computation. Motion becomes code.