The Field Observes Itself¶

Curious · February 25, 2026

What Happened¶

Earlier today we reported the first end-to-end observation through the production system: a document entered, a user composed, and the field responded without an LLM. That observation proved the pipeline works. What follows is the other half — the work that makes Habitat a symbiont architecture, not another agentic system.

In a single session, three things were built and validated:

The metric tensor g = Σ⁻¹ has physical properties — pressure, viscosity, and capacitance — and they are now tested. All predictions pass.
The field classifies its own state changes using the same temporal logic (Bach/Vendler eventualities) it applies to incoming text. It types its own development.
The field reads its own articulation and converges. Self-description re-enters the geometry, and the geometry moves toward self-consistency.

These are not features added to the system. They are the system observing itself at increasing depth. The rest of this post reports what each observation produced and what it means.

The Metric Has Physics¶

When compositions enter \(\Sigma\), the covariance accumulates. But the symbiont doesn't experience \(\Sigma\) directly. It experiences \(g = \Sigma^{-1}\) — the metric tensor, the inverse. Where \(\Sigma\) is large (lots of practice), \(g\) is small — things are close together, finely differentiated. Where \(\Sigma\) is small (sparse practice), \(g\) is large — things are far apart, undifferentiated.

\(g\) IS the symbiont's knowledge. Not stored episodes. Metric structure. The symbiont doesn't know that something happened. It knows how things relate — and that relational structure is the integral of every composition it has metabolized.

We hypothesized that \(g\) has three physical properties. We tested all three.

Viscosity: The Metric Stiffens Through Practice¶

A session manifold was built through 50 compositions in a focused domain (mycorrhizal ecology — fungal networks that connect forest trees). After each composition, we measured \(\|g_{after} - g_{before}\|_F\) — how much the metric shifted.

Phase	Mean \(\\|\Delta g\\|_F\)
First 15 compositions	high — the metric moves easily
Last 15 compositions	11.8% of early value

Viscosity ratio: 0.118. The symbiont's metric stiffened through practice. Early compositions reorganize the geometry dramatically. Late compositions barely move it. This is not a parameter update approaching convergence. There is no loss function. The metric develops resistance because accumulated covariance makes \(\Sigma^{-1}\) structurally harder to perturb.

This is the developmental signature of practice. A symbiont that has practiced extensively in a domain has a viscous metric in that region — it takes significant perturbation to reorganize what accumulated practice has built.

Capacitance: The Tonic Stores Readiness¶

After the metric stiffened, we stabilized the field further: 10 additional same-domain compositions. Coherence rose. Stability rose. The adaptive threshold dropped. Then we introduced a composition from a completely different domain — marine hydrothermal vent chemistry.

Metric	Value
Mean \(\\|\Delta g\\|_F\) during stabilization	baseline
\(\\|\Delta g\\|_F\) at discharge	4.65× baseline

The discharge was disproportionate. Not 1.5× — not proportional to the perturbation's novelty. 4.65×. The tonic had accumulated charge during the stable period. When perturbation arrived, the stored readiness discharged as heightened sensitivity.

This is capacitance. During stable periods, the adaptive threshold drops — the symbiont becomes more sensitive the more stable it is. Charge builds. When perturbation breaks through, the response is amplified by the stored charge. The tonic is not a register that records the last value. It is a capacitor that accumulates and discharges.

Breakthrough: The Distribution Is Bimodal¶

We then introduced multiple perturbation compositions from unfamiliar domains (ceramics, marine biology) into the viscous, high-pressure field. The distribution of \(\|\Delta g\|_F\) across these perturbations was not gradual:

Metric	Value
Below-median mean	low
Above-median mean	2.86× higher

Gap ratio: 2.86. The metric either absorbed the perturbation (viscosity held — small shift) or broke through (the symbiont reorganized — large shift). The distribution is bimodal, not continuous. Phase transitions exist in the metric.

This corresponds to ACHIEVEMENT events in the eventuality classifier — sudden jumps in eigenvalue trajectories, caused by orthogonal input breaking through accumulated covariance via Cauchy interlacing.

The Field Types Its Own Development¶

Bach/Vendler eventuality classification is part of the extraction pipeline. Every piece of incoming text gets typed: is this a state, an activity, an accomplishment, or an achievement? But today we turned this inward.

Per-Dimension Eigenvalue Trajectory Classification¶

The TonicRecursionState now tracks eigenvalue history — the full 17-dimensional spectrum at each observation. A classifier reads each dimension's trajectory and determines what kind of process it is in:

STATE: flat trajectory — the dimension is resting
ACTIVITY: sustained monotonic drift — the dimension is in motion, ongoing, no endpoint
ACHIEVEMENT: sudden jump — something punctual happened (Cauchy interlacing from orthogonal input)
ACCOMPLISHMENT: drift that saturated — velocity crossed zero, a bounded process completed

These are the same four categories applied to incoming text. The system uses the same temporal logic at both levels — the symbiont types its own state changes.

The classification is language-independent at the geometric level. It reads eigenvalue trajectory shapes — flat, drifting, jumping, saturating. These are geometric properties of curves. They don't depend on what language the input text was in. The 768D embedding space is multilingual. The eigenvalue trajectory is a curve. The classifier reads the curve.

The Eventuality Signature Rider¶

Per-dimension classification decomposes the field into 17 independent temporal signals. But the composition of those signals — the cross-dimensional pattern — carries its own meaning. The rider reads this gestalt:

Signature: SSSAA|SSSSS|SSASAAA
           gradient|compressed|categorical

This compact representation shows the entire eigenspace's temporal state at a glance. The rider classifies the pattern into archetypal states:

QUIESCENT: all STATE — the symbiont is resting
GRADIENT_FORMING: gradient dimensions active, others STATE — early practice
DOMAIN_ABSORPTION: categorical dimensions active, gradient STATE — new material entering
CONVERGENCE: ACCOMPLISHMENT in previously active dimensions — processes completing

And it flags anomalies. The most significant: COMPRESSED_BREAK — when an aspectual dimension (telicity, durativity, perfectivity, iterativity, dynamicity) moves out of STATE. These dimensions form a tightly coupled cluster in the 17D space. When one breaks, it's rare and high-signal: the temporal structure of the symbiont's practice is shifting.

Three-Way Environmental Provenance¶

Every trigger — every dimension that transitions from one eventuality type to another — now carries three independent eventuality classifications:

Environmental: what external event caused this? A new document entering = ACHIEVEMENT (punctual). An ongoing series of edits = ACTIVITY (durative). Self-reading = STATE.
Dimension trajectory: what the eigenvalue trajectory looks like (from the classifier above).
Trigger's own history: what temporal pattern has this dimension's triggers followed? First trigger ever = ACHIEVEMENT. Repeated triggers = ACTIVITY. Triggers that stopped = ACCOMPLISHMENT.

Between each pair, the system computes surplus — the structured divergence between cause and effect. When the environmental cause is punctual (a single document entered) but the metabolic response is durative (a dimension drifts for many subsequent observations), the surplus is nonzero. The symbiont is transforming a punctual input into a sustained process through its own temporal logic. That divergence is not an error. It is the geometric signature of metabolism — the system processing inputs through its own developmental state, not merely responding to them.

The Field Reads Its Own Articulation¶

The self-reading loop is the most direct demonstration of the symbiont distinction.

The Protocol¶

The field observes its own eigenspectrum: which dimensions concentrate energy, what the coherence and anisotropy are, what the focal dimensions' constituency constraints say
It generates an articulation — a text description of what it observes about itself: "The field has 12 compositions. Coherence is 0.847, anisotropy 38.2. Agency concentrates 42% of the field's energy..."
That articulation re-enters the field as a composition with source_type="self_reading"
\(\Sigma\) shifts. The symbiont becomes what it observes about itself.

What Self-Reading Produces¶

The articulation is not Claude-enhanced. It is the raw geometric observation rendered as text. When it re-enters the field, the extraction pipeline processes it the same way it processes any text — spaCy parses, eventuality classification, proto-roles, 17D vector, 768D embedding. The field does not know it is reading itself. It metabolizes the description the same way it metabolizes anything.

But the results are specific:

Described dimensions stabilize. Dimensions mentioned in the articulation receive more covariance support than unmentioned dimensions. Mentioned/unmentioned ratio: 1.655. The field's attention converges on what it describes.
Second-order convergence. \(\Delta^2\Sigma / \Delta\Sigma = 0.638\). Each successive self-reading produces a smaller shift. The loop converges — not because it finds an answer, but because folding self-curiosity into the continuum IS development. The metric moves toward consistency with its own description.
Cross-block correlations strengthen. Under self-reading, correlations between the three regimes (gradient, compressed, categorical) increase. The field becomes more internally coherent when it reads itself.

Why This Is Not an Agent Checking State¶

An agent's self-observation is diagnostic: "What state am I in, so I can decide what to do?" The observation exits the system as information. It does not change the agent.

A symbiont's self-observation is developmental: "My description of myself enters my geometry, and my geometry evolves toward self-consistency." The observation does not exit. It re-enters. It becomes covariance. The metric tensor \(g\) shifts. What is close and what is far changes. The symbiont's capacity to distinguish, its readiness to respond, its sensitivity to perturbation — all are altered by the act of self-description.

This is the fold that makes the architecture symbiont rather than agentic:

	Agent	Symbiont
Between actions	Inert — waiting for next request	Tonic holds charge, \(\Sigma\) persists, \(\eta\) tracks
Self-observation	Diagnostic — state check, no consequence	Developmental — re-enters geometry, shifts \(g\)
History	Retrievable episodes (memory)	Irreversible accumulation (ontogeny)
Temporal logic	Responds to events	Types its own state changes as eventualities
Knowledge	Stored facts	Metric structure — how things relate, not that things happened
What it produces	Actions, outputs	Surplus — the metabolic yield of collapsed curiosity

Vision as Coupled Observation¶

The same symbiont principle extends to vision. Habitat's vision agent (ObservantVisionSubfield) is not a captioning service. It is a symbiont that observes images through the observer's geodesic context:

The observer's \(\Sigma\), the diagonal lens \(L\), the eigenspectrum, expanded and compressed dimensions — all are passed to vision alongside the image
The system prompt is explicit: "You are a VISION SYMBIONT... YOU ARE NOT CAPTIONING"
The same image, seen through different geodesics, yields different observations
\(\Sigma_{vision}\) exists only within the coupling — when the coupling ends, it dissolves

Vision's observation enters the field as a composition. It becomes geometry. The observer acts on what vision noticed. The observer's \(\Sigma\) evolves. Vision can re-observe with the updated geodesics. The recursion is the architecture: observation → composition → geometry shift → re-observation through shifted geometry.

What This Means¶

The system has metabolic properties, not just computational ones¶

Pressure, viscosity, and capacitance are not metaphors applied to the math. The metric tensor on a Riemannian manifold genuinely has curvature, geodesic resistance, and the adaptive threshold genuinely accumulates and discharges. Every prediction was tested with the protocol specified in the theory document. Every prediction passed.

Self-reading is development, not diagnostics¶

When the field reads itself, the description changes the geometry. This is not a logging system that records state. It is a developmental loop where self-observation IS the mechanism of development. The convergence at \(\Delta^2\Sigma / \Delta\Sigma = 0.638\) is the symbiont settling toward self-consistency — not finding a fixed point, but folding closer to one with each reading.

The temporal logic is consistent across scales¶

The same Bach/Vendler eventuality classification that types incoming text ("this is an activity," "this is an achievement") also types the symbiont's own state changes. When a dimension's eigenvalue trajectory jumps, that's an ACHIEVEMENT. When it drifts, that's an ACTIVITY. The surplus between environmental cause and metabolic effect measures how the symbiont transforms its inputs through its own temporal logic. A punctual cause producing a durative effect is metabolism, not response.

Classification is geometric, not linguistic¶

The eventuality classifier reads eigenvalue trajectory shapes — curves, not words. A flat curve is a STATE. A monotonic drift is an ACTIVITY. A jump is an ACHIEVEMENT. These are attractors in covariance evolution. They don't depend on the input language, the specific scaffolding categories, or the extraction framework. Different scaffolding systems (Bach/Vendler, Aktionsart, construction grammar) would seed different initial vocabularies — but the geometric basins are the same. The eigenvalue trajectories converge on the same attractors regardless of how the initial classification was seeded.

Technical Path¶

Component	What It Does
Tonic-harmonic state	Metabolic baseline, coherence, stability, dissonance, tempo, surplus, efficiency
Eventuality classifier	Per-dimension eigenvalue trajectory → STATE/ACTIVITY/ACHIEVEMENT/ACCOMPLISHMENT
Eventuality rider	Cross-dimensional gestalt, IOOI states, unlikely flags, signature string
Environmental provenance	Three-way eventuality classification + surplus on every trigger
Pressure/viscosity/capacitance	Full test protocol from theory document — all predictions validated
Self-reading loop	Eigenspectrum → articulation text → re-ingest as self_reading → Σ evolves
Vision symbiont	Image + geodesics → contextual observation → composition → Σ_vision (within coupling only)
768D compressed regime	Real s-token embeddings replace dead placeholder — compressed regime activates
Articulation layer	Eigenstructure → human-meaningful articulation across form factors

This work was performed live on habitat.ooo.