Skip to content
HolonomiX

Representation is the control surface.

A proprietary computational substrate. Products are developed in it and surfaced from it.

QTT-native · GPU-native · Physics-native

The same numbers. A different shape.

Section B-B · dense to TT to QTT
Dense one huge block · every value stored
unfold
Tensor Train a chain of small linked pieces
fold by zoom
QTT a finer chain · each core one zoom level

Changing the shape never changes the data. It is the same numbers, organized differently. Information is lost only when the tiniest, least-important detail is deliberately dropped, and the tolerance is set explicitly.

Loss is a choice. Not a side effect.

Every new direction multiplies the size.

A full 3-D volume already holds 16.7 million points. Track it over time and it passes 4.3 billion. This is why high-fidelity physics reaches for GPUs and supercomputers.

256 256 ticks
65,536 256² pixels
16.7M 256³ points
4.3B 256⁴ over time
1,073,741,824 values, dense
≈ 3,840 pieces, QTT
A field with 1.07 billion values, when the structure is real

Does the structure stay simple as you zoom in?

Folding only pays off when the data has real structure: smooth regions, repeating patterns, self-similar shapes, predictable physics. If the answer to the deciding question is yes, the folded form stays tiny no matter how big the data gets.

The boundary is stated plainly: at small sizes folding makes data bigger, and pure noise has nothing to fold. Below the crossover, plain storage wins. The substrate measures the structure and routes honestly rather than forcing one path.

Compute on the pieces. Rebuild only what you need.

An operation, like one physics step, is also a small chain. It runs against the data chain piece by piece; the full object is never materialized. Each step makes the links swell a little, so the chain is tidied back down before the next step.

Run Swell Tidy up
Match the data?How close to the original. The biggest single error. Whether it still obeys the physics. Whether it passes a spot-check.
Answer holds up?Works on new, unseen cases. Keeps mass and energy balanced. The final answer stays right. For science and engineering, this matters more than how small the file is.

Trust you can calculate and validate independently.

The substrate is engineered under the same posture the products ship with: forced in development, surfaced at the client-facing level.

V&VVerification & Validation as a release gate, not a report.
PQCPost-quantum attestation and ledgered provenance on what ships.
Lean4Formal proof where invariants are provable. Measured verification where they are not. The two are never conflated.
IEEE FPFloating-point standards for bit-exact, reproducible numerics.
Evidence doctrine Each product emits ML-DSA-signed, offline-verifiable evidence.

The substrate in production.

HX-ProvenanceSigned receipts and offline-verifiable evidence for registered artifactsDeployed · GCP + AWS
HX-PQC-Encryption Lifecycle PlatformCryptographic inventory to migration to runtime protection to assuranceDeployed · GCP + AWS
HX-AIFactoryTwinGPU-native thermal-fluid simulation with signed evidence per runDeployed · GCP + AWS
HX-MDAO-TPMultidisciplinary design and optimization with a trust and provenance planeDeployed · GCP + AWS
HX-SDPThe AI inference data plane as one signed platformPrivate Appliance