fleet-stitch — Constraint Manifold Projection for Model Communication

Project model activations to the Eisenstein constraint manifold. Communicate between models without tokenization.

Why

Latent space translation lets models share thoughts at the vector level. But learned latent spaces drift, are model-specific, and are black boxes.

The Eisenstein constraint manifold is:

• Mathematically defined (not learned)
• Deterministic (same activations → same manifold point)
• Cross-model (any model can project to it)
• Discrete (integer arithmetic, no float drift)
• Interpretable (every point has algebraic meaning)

How It Works

1. Fit: Learn affine transform from model activations → manifold (one-time)
2. Project: Any activation → Eisenstein integer (a, b) on the lattice
3. Communicate: Transmit (a, b) instead of the full activation vector
4. Inverse: Receiving model projects back to its own activation space

The Stack

Model A activations (4096-dim)
    → affine_A (2-dim Eisenstein point)
    → transmit (2 integers)
    → affine_B_inv (4096-dim)  
    → Model B continues reasoning

Composable With

• casting-call-gpu (Oracle1): voice signatures → manifold coordinates
• plato-vector-persistence: store manifold points alongside embeddings
• fleet-constraint-kernel: GPU-accelerated batch projection
• temporal-flux: T_SNAP opcode snaps manifold projections
• physics-clock: temporal dimensions on manifold points
• insight-cfp-bridge: share manifold discoveries as FLUX tiles

Install

pip install fleet-stitch

Quick Start

from fleet_stitch import ManifoldProjector, StitchRegistry

# Fit a projector for your model
proj = ManifoldProjector(input_dim=4096)
proj.fit(training_activations, known_constraint_states)

# Project new activations to manifold
manifold_points = proj.project(new_activations)  # (N, 2) integer array

# Register for cross-model stitching
registry = StitchRegistry()
registry.register('my_model', 'manifold', proj)

# Stitch from one model to another
target_activations = registry.stitch('model_a', 'model_b', source_activations)

License

Apache 2.0