Watching Habitat Load

1. This is NOT an Embedding Similarity System

Most "semantic" systems do this:

Habitat does something fundamentally different. The 768D embedding is just the starting point — the "white light" that gets decomposed.

2. The Extraction Pipeline: Compositional Semantics

From the logs, I watched the EnhancedNativeExtractor work:

Step 1: S-Token Extraction

Step 2: 12D ProcessAssert Projection (Predicates/Aspectuals)

This is decomposing WHAT the predicate asserts — aspectual structure from Bach/Vendler classification.

Parallel: 5D ProcessActor Projection (Actants/Modalities)

This is decomposing WHO is acting — modality structure from semantic roles (AGENT, THEME, EXPERIENCER, INSTRUMENT). Role assignment uses Dowty proto-role properties (1991):

• Proto-agent: volition, sentience, causation, movement, exists_independently
• Proto-patient: undergoes_change, incremental_theme, causally_affected, stationary

Proto-role scores are modulated by Bach/Vendler aspectual class to determine final role assignment.

Step 3: Rainbow's Ghost Preservation

The system projects DOWN to lower dimensions but keeps the original. The "white light" (768D) is preserved alongside the decomposition. This is reversible observation, not lossy compression.

3. Bach/Vendler Classification — Aspectual Structure

The system classifies predicates into aspectual classes:

• STATE — "knows", "believes" (no change)
• ACTIVITY — "runs", "swims" (unbounded process)
• ACCOMPLISHMENT — "builds a house" (bounded, has endpoint)
• ACHIEVEMENT — "arrives", "dies" (instantaneous)

This is linguistic ontology, not machine learning classification.

Plus Levin verb enrichment:

This connects to Beth Levin's work — verbs that undergo the same syntactic alternations share semantic structure.

4. 17D Compositional Vectors from Relations

The system extracts ProcessActor ⊗ ProcessAssert pairs — WHO does WHAT:

• ProcessActors (5D): agency, stability, influence, boundary, resonance — from semantic roles
• ProcessAsserts (12D): 5D aspectual core + 7D polyworld detection.

This is the tensor product: 5D Actor ⊗ 12D Predicate → 17D compositional vector.

5. Document Manifold: Σ and Eigenstructure

The 268 compositional vectors get aggregated into a covariance matrix Σ (17×17).

The eigenvalues reveal the energy distribution — where meaning concentrates. The first eigenvalue (2.858) dominates. The document has strong directional structure.

6. Fresnel Zones — Prismatic Observation at Greatest Fidelity

Fresnel zones are natural observation positions revealed by eigenvalue geometry — prismatic sectioning of the tensor at points of greatest observational fidelity.

From the architecture: "Eigenvalues = prismatic refraction angles. Cumulative eigenvalue energy = Fresnel zone boundaries. Natural observer positions emerge from geometry."

Each zone is a prism face where the tensor trio operates:

• Fall-line (g_ij gradient) — "Where am I pulled?"
• Descent (geodesic path) — "How do I get there?"
• The lag/experience — "What is it like to not-yet-know but move-toward?" (curiosity itself)

Zones form when the manifold's condition number drops low enough to "see across." Each zone represents a distinct perspective where the geometry allows clear observation. The geometry tells you: "I can see now."

Anisotropy = 91.391 — This document has HIGHLY directional structure. Tight focal beam, strong fall-line.

7. The Metric Tensor: g = Σ⁻¹

The metric tensor defines distance in semantic space. But it's observer-dependent:

• Different users have different Σ_user
• Same document looks different through different metrics
• Mahalanobis distance = geodesic distance through that user's metric

8. Observer-Dependent Semantics: Σ(observed | observer)

This is the key insight. From the foam model:

There is no "objective" semantic position. What Olivia sees when she observes a document is different from what Katrin sees — because their Σ matrices are different.

9. Redis Architecture: Event Sourcing

• Extraction requests go into Redis streams
• Workers consume and process
• Coupling traces are recorded (user↔document meetings)
• This is temporal provenance — every observation has history

10. What This Means

Habitat is not searching for similar content.

Habitat is:

1. Extracting compositional structure (ProcessActor ⊗ ProcessAssert)
2. Building geometric objects (Σ matrices, metric tensors)
3. Computing observer-relative observations (Σ(observed | observer))
4. Preserving plurality (different users see differently)
5. Detecting constitutional dimensions (high curvature = sovereignty-preserving)
6. Enabling coordination without convergence (bridgeable dimensions exist)

The LLM (if used at all) just formats geometric truth into language. It doesn't compute meaning — Habitat does.