The Field Observes Without an LLM¶

Curious · February 25, 2026

What Happened¶

On February 25, 2026 at 01:16 UTC, a climate science document entered Habitat through a live Google Workspace integration. The document — Reflecting on a Year of Progress by R. Max Holmes, President & CEO of Woodwell Climate Research Center — was imported from Google Drive, extracted into four semantic compositions, and observed by a user through the full pipeline: compose, query, and field response.

This is the first end-to-end observation through the production system. Every step is traceable. What follows is a report of what happened, what produced each output, and what the outputs mean.

The Document Enters the Field¶

The document was fetched from Google Drive via OAuth, split into sentence-aware chunks (≤800 characters each), and each chunk was passed through the extraction pipeline:

spaCy NLP parses the text into grammatical structure
Bach/Vendler eventuality classification determines process type (activity, accomplishment, achievement, state)
Levin verb alternation enriches the predicate dimensions when the verb is classified
Dowty proto-roles compute agency, stability, influence, boundary, resonance
The result: a 5D ProcessActor ⊗ 12D ProcessAssert = 17D compositional vector per chunk
SentenceTransformer (all-mpnet-base-v2) independently computes a 768D embedding — semantic ground truth

Four compositions entered the session manifold. The covariance matrix \(\Sigma_{session}\) accumulated their geometric relationships.

Metric	Value
Compositions	4
Coherence	0.832
Anisotropy	40.388

The User Composes¶

The user clipped a passage:

"Federal actions have also directly impacted Woodwell's operations, from the loss of (a small number of) government grants and collaborations to concerns about future funding priorities and criteria."

And added an annotation:

"The risk to climate research is significant. The eradication of dissenting perspective is isolating discourse and deliberation. It foists administrative burden onto small and even large entities — from climate to gender to healthcare. It's an assault forcing new pipelines to emerge and maybe Habitat can be part of the solution."

This entered the field as a composition. The system computed:

Coupling (diagonal lens \(L = \Sigma_{user}^{-1} \cdot \Sigma_{session}\)):

Relationship type: echo — the user's reading pattern mirrors the document's structure
λ ratio: 1.000 — all 17 eigenvalues near unity
Fidelity: 1.000 — perfect agreement between 17D structure and 768D semantics
All 17 dimensions: resonant (none expanded, none compressed)

Tonic (metabolic rate):

Trajectory: diverging — the composition introduced tension into the field
Metabolic rate: 0.1156
Dissonance: 0.681

Articulation (Claude SYNTAX+ED):

"The passage reflects on a year of progress, proceeding through familiar terrain of the author's work and experiences. The dimensional relationship suggests a stable agency capacity and stability measure, indicating a grounded, resonant perspective. However, the user's own vocabulary hints at a tension or divergence — federal actions have impacted Woodwell's operations, and the implications of the Paris Agreement will be felt."

This articulation was generated by Claude, operating under Habitat's sovereignty protocol. Claude receives the geometric context — eigenvalues, differential vectors, dimensional relationships — not the user's raw text. It speaks from the geometry, not from the language. The attribution reads: Generated by Claude [SYNTAX+ED].

The User Asks — and the Field Responds Without an LLM¶

The user typed: "What patterns are emerging?"

The system responded:

"The field (5 compositions) attends most strongly to 'the Paris Agreement · will be felt' (cosine 0.386, recency 0.349). Meaning is ahead of structure — the field has not yet crystallized this."

No language model was involved in producing this response. The attribution reads: Observed by Habitat · Field geometry · No LLM.

How the Query Works¶

The query endpoint (/api/corpus/query) follows a specific protocol:

The question does NOT become a 17D composition. It does not enter the field. It does not mutate \(\Sigma_{session}\). It observes from outside.
768D embedding only. SentenceTransformer encodes "What patterns are emerging?" into a 768D vector. This is an embedding model, not a generative model — it computes a fixed-length numerical representation of the question's meaning.
Cosine attention. The 768D question vector is compared against all 5 compositions' 768D embeddings via cosine similarity. The highest-attending composition: "the Paris Agreement · will be felt" (cosine 0.386).
Recency weighting. Each composition's attention score is blended with its recency (5-minute half-life): relevance = cosine × 0.7 + recency × 0.3.
Weighted centroid as 17D probe. The attended compositions' 17D vectors are combined by relevance weight into a centroid. This centroid is where in the field the question points — derived from the field's own compositions, not from the question's structure.
Riemannian adjacency. The probe searches the field's metric space using Mahalanobis distance: \(d = \sqrt{(\mathbf{x} - \mathbf{y})^T \Sigma^{-1} (\mathbf{x} - \mathbf{y})}\). This is eigenvalue-weighted distance — the field's covariance shapes what "close" means. Not Euclidean. Not cosine. Riemannian.
Surplus analysis. For each attended composition, the system compares structural proximity (17D Mahalanobis) against semantic proximity (768D cosine). When 768D says "close" but 17D says "far": meaning leads — language sees a relationship the structure hasn't crystallized. When 17D says "close" but 768D says "far": structure leads — the geometry sees a pattern meaning hasn't named. When both agree: convergent.
Deterministic articulation. The response text is a string template filled with observed values. The sentence "Meaning is ahead of structure — the field has not yet crystallized this" is triggered when surplus_type == 'meaning_leads' — a deterministic conditional, not a language model generation.

What the ED Vocabulary Revealed¶

The field returned five compositions ranked by Riemannian adjacency:

Composition	Source	Surplus Type	Emergence
Federal actions — have impacted Woodwell's operations	user clip	meaning leads	0.412
the months — has eased	drive document	meaning leads	0.406
the Paris Agreement — will be felt	drive document	meaning leads	0.379
a year — Reflecting on year of progress	drive document	meaning leads	0.375
an added intensity — was added intensity	drive document	convergent	0.360

Four of five results show meaning leads. The 768D semantic space recognizes these compositions as relevant to "What patterns are emerging?" but the 17D structural space — with only 5 compositions — hasn't built enough covariance to confirm the relationship.

The single convergent result ("an added intensity") is the one composition where both spaces agree. Both the structure and the meaning see it as nearby.

This is the surplus analysis working exactly as designed: observing the gap between structure and meaning, reporting it, not trying to close it.

The Fresnel Zones¶

The eigenspectrum of \(\Sigma_{session}\) produced five Fresnel zones:

Zone	Energy (λ)	Fidelity
λ₀	0.346	0.92
λ₁	0.220	0.84
λ₂	0.067	0.92
λ₃	0.016	0.84
λ₄	0.010	0.84

λ₀ captures ~53% of the field's energy — the dominant direction of attention. The fidelity scores measure agreement between the 17D eigenstructure and 768D semantic similarity rankings. At 0.92 and 0.84, the two independent spaces agree strongly on where the field concentrates.

This is the S ≅ V(T) ≅ F(T) preservation in action: semantic structure preserved through compositional vectors, verified against field geometry at every zone.

What This Means¶

The system produces meaningful observations without a language model¶

The query response identified the most salient composition ("the Paris Agreement · will be felt"), diagnosed the surplus type ("meaning leads"), and reported it in the corpus's own vocabulary. The entire path — from question embedding through Riemannian adjacency to surplus classification to articulation — is deterministic linear algebra and string templating. No tokens were generated. No model was prompted.

Surplus is observable and diagnostic¶

When the field reports "meaning leads," it is making a specific geometric claim: the 768D embedding space sees proximity that the 17D structural space hasn't confirmed. With only 5 compositions, \(\Sigma_{session}\) is still forming — it doesn't have enough covariance to crystallize the relationship. As more compositions enter, "meaning leads" will shift toward "convergent" at the compositions where the structure catches up. The surplus type is not static. It is a measure of the field's developmental state.

Claude and Habitat have distinct roles — and both are labeled¶

The compose step uses Claude under the SYNTAX+ED sovereignty protocol. The query step uses no LLM. Both are now explicitly attributed in the UI:

Generated by Claude [SYNTAX+ED] — Claude speaks from the geometric context, not the raw text
Observed by Habitat · Riemannian adjacency · No LLM — the field's own geometry, deterministically rendered

The user sees exactly what produced each observation. This is not a design choice — it is a structural requirement. When the system's outputs are traceable to their sources, the user can evaluate what they trust.

Technical Path¶

Step	What It Does
Drive import	Fetches from Drive, splits text, queues background extraction
17D+768D extraction	spaCy → ProcessActor(5D) ⊗ ProcessAssert(12D) + SentenceTransformer(768D)
Session manifold	Updates Σ_session, eigendecomposition, trajectory
Diagonal lens	L = Σ_user⁻¹ · Σ_session, eigenvalue classification
ED vocabulary	Mahalanobis distance via g = Σ⁻¹, surplus type from 17D vs 768D
Fresnel zones	Eigenspectrum concentration, 17D↔768D fidelity per zone
Query (no LLM)	768D cosine → weighted centroid → Riemannian probe → surplus → template
Compose	Extract + diagonal lens + sovereignty-enhanced marginalia

This observation was recorded live on habitat.ooo.