The HolonomiX Atlas
A signed measurement framework for understanding how data behaves before infrastructure decides how to move, compress, rerank, tensorize, or process it.
It is not a leaderboard. It is a deployment behavior map.
text, vision, physics, scientific
3D tensor measurements
physics rows only
Proof, not positioning
Every row is classified under a locked calibration protocol. The Atlas tracks what was measured, what was certified, and what remains pending.
| Evidence layer | Count |
|---|---|
| Classified text, vision, physics, and scientific rows | 71 |
| QTT-native 3D tensor measurements | 6 |
| Text-corpus stability matrix | 7 models × 5 corpora |
| Physics compressed-path certified rows | 5 |
| Signed rollup checkpoints | ML-DSA-65 / FIPS 204 |
Why the Atlas exists
Data is copied, moved, embedded, transformed, indexed, compressed, and recomputed before the system understands what kind of data it is dealing with. That creates waste: unnecessary movement, duplicated compute, overused rerank paths, unsafe compression assumptions, expensive high-dimensional representations, and one-size-fits-all routing decisions. The Atlas changes the order of operations. First, measure the structure. Then choose the execution path.
What the Atlas measures
Can the data be reduced without destroying the structure that matters? The Atlas measures rank, retained-rank fraction, spectral decay, and compression ratio to determine whether a corpus is governable, compressible, conditional, or weak.
Does rerank actually help? The Atlas compares single-pass recall against reranked recall at K=10. Some rows recover strongly. Some are flat. Some get worse.
Rerank is not a default feature. It is a measured policy. Each row is classified RERANK_RECOVERABLE, RERANK_LOW_LIFT, RERANK_FLAT, or RERANK_HARMFUL. A platform should not spend compute on rerank when the data shows it will not help.
Does the behavior survive domain drift? The Atlas measures model behavior across multiple text corpora to determine whether a model is stable, drifting, or insufficiently measured.
Does compression preserve final recall under the actual serving path? The Atlas distinguishes fp32 calibration from compressed-path certification. A row can show strong fp32 rerank utility without being certified for TQ4 deployment.
When is QTT the right tool? The Atlas measures QTT-native 3D tensor fields separately from flattened embedding rows. QTT is not treated as a universal embedding compressor. It is used where the data is natively tensor-structured.
Headline findings
On C4, Qwen3-Embedding-4B at D=2560 lands A_ELITE with RR=0.9958 and G=+2.1 pp. Six of seven measured D=4096 7B-class rows land D_SENSITIVE.
Dimension is not a deployment strategy.
| Rerank policy | Count |
|---|---|
| RERANK_RECOVERABLE | 43 |
| RERANK_FLAT | 15 |
| RERANK_LOW_LIFT | 6 |
| RERANK_HARMFUL | 7 |
A uniform rerank policy is wrong. The Atlas tells the platform when rerank is worth paying for — and when it should stay off.
Five models are STABLE_RECOVERABLE across all five measured text corpora:
NV-Embed-v2 and OpenAI 3-large show drift. Customer data rarely looks exactly like a benchmark corpus.
| Regime | Result |
|---|---|
| MACE-MP-0 large | ρ=0.0007, CR=1365× |
| Smooth PDE fields | CR=16–71× |
| Shock-dominated PDEs | flatter spectra, stronger rerank dependence |
| QTT-native 3D fields | up to 7571× compression |
Different data structures need different execution paths.
QTT underperforms when forced onto arbitrary flattened embeddings. But on native 3D tensor fields, the story changes. Taylor-Green vortex at 128³ reaches 7571× compression at rank 4, and 244× compression at rank 32 with 1.31e-5 relative L2 error.
QTT is not the default embedding path. QTT is the tensor-native path.
Classification coverage
The current Atlas covers only a portion of the verdict×class grid. The visual marks calibrated cells; empty cells are roadmap. fp32 calibration only — compressed deployment is a separate 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 |
Calibration scope
Every text/vision row is fp32 calibration. Compressed certification measures ΔSP, ΔRR, and P before any text or vision row becomes COMPRESSED_CERTIFIED. Five physics rows are already compressed-certified under the locked protocol.
Atlas vs. leaderboard
A leaderboard asks
Which model scored highest?
The Atlas asks
What should the system do with this data?
A model can score well and still be unstable across corpora. A high-dimensional embedding can still be deployment-sensitive. A compressed path can damage single-pass recall but remain safe after rerank. A tensor method can fail on flattened embeddings but excel on native fields. The Atlas is built to expose those distinctions.
HX-SDP integration
HX-SDP uses the Atlas as a structural intelligence layer. The platform should not treat all data the same. It should classify the data first, then decide whether to:
This is the core of structure-aware execution.
The Atlas decision stack
01
Data stream
02
Structural measurement
03
Spectral governability
04
Retrieval recoverability
05
Rerank economics
06
Compression certification
07
Execution policy
The output is not just a score. The output is a decision.
Atlas assessment
For a customer corpus, an Atlas assessment can answer:
Current coverage
| Domain | Coverage |
|---|---|
| General text embeddings | measured |
| Vendor embedding APIs | measured |
| Open-weight embedding models | measured |
| Vision embeddings | measured |
| Scientific text | measured |
| Physics PDE fields | measured |
| Equivariant atomic representations | measured |
| QTT-native 3D fields | measured |
| Text / vision compressed certification | pending |
| Protein language models | planned |
Measurement integrity
Most rows are fp32 calibration rows. Only rows explicitly marked as compressed-certified should be treated as deployment-certified for compressed serving.
Public assets
Public Atlas Table
A cleaned, publishable table of measured rows, compressed-certified rows, stability results, and QTT-native tensor measurements.
Technical Report 001
A technical whitepaper explaining the Atlas method, measured results, and platform implications.
Atlas Assessment
A customer-specific measurement pass that classifies how a real corpus behaves under the HX-SDP structural data platform.
FAQ
No. The Atlas is a deployment behavior map. It measures what the system should do with a model/corpus/data pair, not just which model has the highest score.
Not by default. In fp32 calibration, A_ELITE means native rerank utility is at least +1.0 pp. TQ4 readiness requires compressed-path certification.
A model that behaves well on one corpus may drift on another. The Atlas measures whether model behavior survives domain shift.
QTT is evaluated where tensor structure is real. The Atlas distinguishes between flattened embeddings, where QTT may not be appropriate, and native tensor fields, where QTT can produce very high compression.
HolonomiX uses the Atlas to guide HX-SDP execution policy: reduce, rerank, compress, tensorize, certify, or preserve the exact path based on measured structure.
Before optimization
HolonomiX built the Atlas to measure the structure of data before the platform decides what to do with it. The result is a practical execution map: when to reduce, when to rerank, when to compress, when to tensorize, when to leave the data alone.
Data has structure. The Atlas measures it.