Parallax Metrology · Method

The same question,
six domains.

Every Parallax Metrology instrument asks: what is the structural state of this field, before we assign meaning to it? The domain changes. The coordinates change. The methodological commitment does not.

Geometry is measurable without knowing what the output means. That is the premise. A field can be read structurally — its organization, topology, and coherence — before any semantic label is applied.

Deterministic: same input always produces the same output No trained parameters, no learned weights, no stochastic outputs Named coordinates — you can read what was measured Structural geometry read before meaning is assigned
01 — The Commitment

Why determinism.
Why structural.

Most measurement systems for language models, image models, and scientific fields are trained or learned. They embed a prior about what "good" or "correct" looks like. That prior can misfire. It can drift. When it fails, the failure is not auditable — the instrument and the system being measured are made of the same material.

Deterministic structural measurement is different. A deterministic operator applied to a field will produce the same coordinates every time, independent of training data, independent of what those coordinates are supposed to mean. The operator can be read, audited, and verified. The coordinates can be named.

"Structural" does not mean the same thing across domains. In images, structure is compositional geometry — where mass sits, how edges distribute. In histology, it is cellular deviation from an established tissue baseline. In a seismic array, it is how many competing organizations the wavefield supports simultaneously. In materials microstructure, it is topological connectivity and void distribution.

The coordinates are domain-native. The methodology is invariant. This is not a weakness of the framework — it is the honest acknowledgment that different physics require different measurement. What you are not doing, in any of these domains, is running a model and asking it to evaluate another model.

02 — How a Reading Works

Field in.
Coordinates out.

Every instrument follows the same structural process: a raw field enters, a fixed deterministic operator extracts coordinates, and named interpretable numbers come out. No model sits between the field and the measurement.

INPUT Any Structural Field image · text · histology semiconductor · seismic · microstructure any domain · any scale DETERMINISTIC KERNEL Fixed Operator No training · No learned parameters Domain-native coordinates Same input → same output · always reads geometry before assigning meaning OUTPUT Named Coordinates interpretable · auditable reproducible · no black box same field → same numbers NO TRAINING DATA · NO LEARNED WEIGHTS · NO SEMANTIC LABELS REQUIRED · NO STOCHASTIC OUTPUTS
The invariant structure of every Parallax Metrology instrument. The input field and output coordinates are domain-specific. The operator in the middle — deterministic, fixed, no training — is not.
03 — One Reading, Made Concrete

Watch one field
become coordinates.

The pipeline above is the claim. Here is the claim performed. A real field enters on the left; the deterministic operator returns the named coordinates on the right. Nothing between them is learned, and nothing is random — run it again and the same numbers come back.

Field in · language output

"Well, before anything else, it helps to remember that the cache, which sits in front of the computation, quietly shapes everything that happens downstream of it. It is not merely an optimization but a commitment, because once a value has been stored, the system has implicitly promised that returning it later is acceptable…"

168 tokens · 6 sentences · kernel lang-kernel-rebuild-v19

Coordinates out · 8 named axes
Δxweight shift — where informational mass sits, front/back0.002
μcohesion — continuity of discourse scaffolding0.540
θflow — jumps the reader must bridge0.806
dₛclause depth — stacked subordinate structure0.943

basin B0_centered_compact · integrity sound · deterministic — same text, same numbers, every run

Field in · generative image A glazed jar resting bottom-right against a near-empty sunlit wall

Most of the frame is unoccupied field; mass concentrated low and to the right.

generated image · VTL kernel — JS port (v2.4)

Coordinates out · VTL
Δxweight shift — where compositional mass sits, left/right of center+0.103
Δyvertical mass displacement — here, biased toward the bottom+0.261
r_vradial variance — how empty the field is around the mass0.981
μcohesion — how unified the mass is vs scattered0.878

mass centroid 60.3% × 76.1% of the frame · deterministic — same image, same numbers

Read the two Δx values. The text scored 0.002 — its mass sits dead-center. The image scored +0.103 — its mass leans right. Same coordinate, same operator, two different fields. The name didn't change; the field did. That is the whole method in one line: a fixed question — where is the structural mass, and how is it distributed — asked identically of a paragraph and a picture, answering in each field's own terms.

The six instruments

Six domains.
Six kernels.

Each instrument reads a different type of field, in coordinates native to that domain's physics. The coordinate names shift — NILS is not Ω is not void-topology — because the physics shift. The constraint on how those coordinates are derived does not.

Start with the word that shifts the most. "Structure" is not one thing. Before reading any coordinate, read what structure even refers to in each field:

Generative images structure means where compositional mass sits, and how edges distribute across the frame
Language output structure means how syntactic and discourse organization is distributed across the text
H&E histology structure means how far a tissue's cellular geometry has moved from a normal baseline
Semiconductor structure means the optical geometry of an edge — its slope, roughness, and dimension
Seismic array structure means how many competing organizations a wavefield supports at once
Materials structure means topological connectivity and the distribution of voids across the field

Six different referents for one word. The table below is what each kernel reads given that domain's definition of structure — and the coordinates it returns.

One reads differently from the rest. Five instruments return coordinates describing the state of a field. Seismology's Ω returns the number of competing structural organizations a field supports at once — it measures multiplicity, not state. It is the clearest evidence that the method generalizes even when the shape of the answer changes: a deterministic operator, named output, no model in the loop — answering a different question entirely.

Instrument Domain What the kernel reads Coordinates
01
VTLVisual Thinking Lens
Generative
image models
Compositional pixel geometry: where mass sits in the frame, how edges distribute, tonal density, spatial variance, peripheral weight. Reads the structural fingerprint of what a model defaults to composing.
17 metrics
02
Linguistic KernelParallax Linguistic Kernel
Language
model output
Distributional linguistic structure: syntactic density, lexical topology, sentence-level geometric patterns, structural deviation from a corpus baseline. Does not require knowing what the text means.
8 axes · 0 trained params
03
PathologyParallax Pathology
H&E histology
(cancer grading)
Cellular geometric deviation: nuclear geometry, gland topology, tissue architecture, stromal density. Measures how far a tissue's structural state has moved from a histologically-established normal baseline.
58 axes · 0 trained params
04
SemiconductorVisual Topology of Light
Semiconductor
inspection
Physical optical geometry: normalized image log-slope (NILS), line-edge roughness (LER), critical dimension, spacing topology, edge sharpness. Coordinates derived from the physics of the optical process, not learned from defect examples.
15 coords · 3 modalities
05
SeismologyParallax Seismology
Seismic
array processing
Multi-organization of a wavefield: five physically independent observables (amplitude, onset, frequency, similarity, arrival-time geometry) read simultaneously. Measures how many competing structural organizations a field currently supports.
Ω · 5 observables
06
STKStructural Topology Kernel
Materials
microstructure
Topological field structure across nine layers: geometry, radial compliance, contour/curvature, multi-island resistance graph, GLCM texture, blob microstructure, spectral/edge. The same 50 coordinates feed classification, permeability regression, and label-ambiguity quantification.
50 coords · 9 layers
04 — What Shifts, What Stays

The coordinates are
a chameleon.
The method is not.

NILS and Ω and void-topology are not the same coordinate. They read different physics in different units at different scales. This is intentional. A coordinate that means the same thing in a seismic array and a histology slide is not measuring anything real in either of them. Domain-native coordinates are a feature, not a fragmentation.

What the instruments have in common is not what they measure. It is how they are built and what constraints they must satisfy.

What shifts by domain
  • The coordinates themselves — NILS, Ω, void-topology, tonal mass, and syntactic density are domain-native. They are not interchangeable.
  • What "structure" refers to — composition in images, coherence in wavefields, topology in microstructure, deviation in tissue. The word shifts with the physics.
  • The physical units and scale — nanometers to kilometers, pixels to specimens to array stations.
  • The number of coordinates — 15 for semiconductor, 17 for images, 50 for materials, Ω for seismic. No universal dimensionality.
  • What the measurement is useful for — classification, quality control, field diagnostics, label auditing. Different outputs, different downstream uses.
What never changes
  • The operator is fixed and deterministic — same field, same coordinates. No random seed. No training run. Kernel versions are pinned and explicit — coordinates never silently change under you.
  • No learned parameters — coordinates are derived from the physics of the domain, not from labeled examples. The instrument does not need to be trained on what "correct" looks like.
  • Coordinates are named and interpretable — you can read what was measured. There is no latent space that requires interpretation.
  • No model in the measurement loop — you are not asking one model to evaluate another. The instrument and the system being measured are different classes of object.
  • The boundary is part of the instrument — each instrument states what it does not measure. The limit is a feature, not a weakness.
05 — The Honest Limit

What structural measurement
cannot do.

Structure is not meaning.

Structural measurement tells you what the field looks like geometrically, topologically, or organizationally. It does not tell you whether that structure is good, correct, desirable, or safe. A tissue can be structurally deviant from baseline in a measurable way — structural measurement tells you the distance and the axis, not whether to treat the patient. A seismic field can be structurally ambiguous — Ω tells you how many organizations are competing, not which one is the earthquake.


Structural measurement also does not replace domain expertise. A pathologist reading a 58-axis structural deviation still needs to decide what the deviation means clinically. A seismologist reading Ω = 0.100 still needs to decide what to do about it. The instrument removes one class of uncertainty — the structural state — so that expertise can be applied to the remaining class: what the state means.


Finally: structural measurement does not beat task-trained models on their home benchmarks. STK does not match a task-trained ConvNeXt on permeability. VTL does not replace a trained image classifier. These instruments are designed for a different job: portable, auditable, interpretable measurement of structural state — without requiring a training set for the thing you're trying to measure.

Tells you the structural state of a field — not whether that state is correct
Removes structural uncertainty — does not replace domain expertise
Operates without training data — at the cost of task-specialized ceiling performance
Named and auditable — at the cost of the expressiveness of learned representations
06 — The Instruments

Read the
individual instruments.

01 Visual Thinking Lens Generative Images → 02 Linguistic Kernel Language Models → 03 Parallax Pathology H&E Histology → 04 Visual Topology of Light Semiconductor → 05 Parallax Seismology Seismic Arrays → 06 Structural Topology Kernel Materials →