Technology architecture
A private computational core, surfaced as evidence-driven products.
HolonomiX is the technology brand for a QTT-native, GPU-native, physics-native operating-system-grade core. HX-SDP is the enterprise data-plane product. HX-Provenance is the proof product. The Atlas is the classification and governance brain.
Surfaces
Product truth and technology truth are scoped separately.
HX-SDP production benchmarks are scoped to the SVD-latent + SQ8 path. QTT remains the broader HolonomiX technology core and an alternate representation path for specialized callers that already hold TT cores.
Structural intelligence core
HolonomiX
The QTT-native, GPU-native, physics-governed technology core. The substrate beneath the commercial products. Tear down the memory wall. Never go dense.
Flagship structural data platform
HX-SDP
Preserves one representation throughout inference. Serves embeddings, features, retrieval, search, retention, gateway, and observability from one GPU-native runtime.
Separate attestation product
HX-Provenance
Standalone exact retrieval and evidence attestation for regulated, audit-grade workflows. Produces proof-bearing receipts, manifests, hashes, and verifier-ready evidence bundles.
Internal intelligence layer
The Atlas
Fingerprints each workload, classifies spectral structure, and decides how data should be represented before it enters the serving path. Internal intelligence layer of HX-SDP.
HX-SDP hot path
Latent scan first. SQ8 rerank second. No dense materialization in the compute path.
The architecture is simple enough to reason about and rigorous enough to benchmark. The Rank Governor selects rank from the eigenspectrum; the query path works against the compressed artifacts in GPU memory.
01
Input
Dense vectors, features, streams
02
Atlas
classify structure + policy
03
Latent
Z(N,r) + V_T(r,D)
04
SQ8
int8 sidecar rerank
05
Serve
cache · vectors · features · search
G = X.T @ X
λ, V = eigh(G)
r = rank_governor(λ, target_quality=0.95)
Z = X @ V[:, :r]
V_T = V[:, :r].T
SQ8 = quantize_int8_per_row(X)w = V_T @ q
scores = Z @ w
candidates = topk(scores, rerank_k=100)
final = sq8_original_space_rescore(candidates, q)The Atlas
Classification is the operating intelligence layer.
The Atlas records model/corpus behavior and separates two different questions: how compressible the corpus spectrum is, and whether the retrieval workload benefits from rerank.
Build-time
DataVerdict
A_GOVERNABLE, B_COMPRESSIBLE, C_CONDITIONAL, or D_WEAK. This is a property of the corpus eigenspectrum and the governor decision.
Query-time
RetrievalClass
A_ELITE or D_SENSITIVE in the current fp32 calibration atlas. Compressed deployment classes require the next certification phase.
| DataVerdict ↓RetrievalClass → | A_ELITE Native fp32 path | B_RECOVERABLE fp32 with rerank | C_BORDERLINE Calibration in flight | D_SENSITIVE fp32 only, no compress |
|---|---|---|---|---|
A_GOVERNABLE Clean low-rank structure | — | — | — | — |
B_COMPRESSIBLE Structure with guard-rails | calibrated | — | — | — |
C_CONDITIONAL Slice-dependent structure | calibrated | calibrated | — | — |
D_WEAK No exploitable structure | — | calibrated | calibrated | calibrated |
GPU-native execution
The commercial architecture is built for private deployment.
The deployable product surface is hx-gate plus hx-engine with Redis for gate state. The severance track keeps the HX-SDP codebase private, VM-imageable, and separate from broader physics IP.
serving runtime, storage, two-tier query, GPU routes
auth, ACL, billing, audit, onboarding, proxy
rate-limit ZSET, CU billing HASH, reload pub/sub
Read next
Move from architecture to evidence.
The proof and benchmark pages carry the measurement tables, receipt scope, and limitations.