Lu(x)iEdge

LuxiQuant / LuxiEdge commercial product

LuxiQuant
Risk State Engine

LuxiQuant Risk State Engine maintains deterministic, authenticated risk state as a scenario-by-position matrix changes. Risk Runtime v0.4.2 is a LuxiEdge commercial product, available on request.

Customer pricing, market data, scenarios, and methodology remain under customer control. Authoritative rebuilds remain the comparison and reconciliation paths. No public runtime download or public prices.

Request accessInspect the evidenceFree LuxiBook demonstration

LuxiQuant capabilities

One of seven LuxiEdge products.

LuxiQuant Risk State Engine is one of seven LuxiEdge commercial products built on one core engine. Oasis and other proof families do not validate the current Risk State Engine. Lifecycle savings from fewer reruns and less revalidation remain an unmeasured objective.

Inference

Demonstrated / internal measurements

Scoped Qwen2-7B-Instruct comparisons inspect repeatability, reference checks, prompt processing, generation speed, and GPU-board energy against named vLLM and SGLang modes. Normal and deterministic comparator modes remain separate.

These are offline fixed-batch tests, not production serving, live mixed-length traffic, or multi-GPU validation. The benchmarked engine is not supplied by the public demo. Current comparisons have not been third-party validated.

Inference evidence and limitsResearch context

Quantization

Research direction

Quantization raises a separate numerical question: which representation, rounding rules, and execution choices define the result? Any proposed Luxi quantization workload needs its own reference, quality criteria, and reproducibility contract.

No specific quantization quality, compression ratio, performance, or energy result is claimed here. Inference results do not establish a quantization product.

Quantization research boundaryPublic test scope

Numerical primitives

Demonstrated / named inputs

Public GPU-sum artifacts separate repeatability from agreement with a correctly rounded exact total. The exact fixed-order path and the faster chunked path are different results; their correctness and timing cannot be combined.

These artifacts are evidence for the published synthetic input and tested platforms, not universal accuracy, speed, or device support. Public verification does not expose the private kernels.

Numerical primitives evidenceReport and verifier instructions

Risk-state computation

Request access

LuxiQuant Risk State Engine maintains deterministic, authenticated risk state as a scenario-by-position matrix changes. It is available on request.

Customer pricing, market data, scenarios, and methodology remain under the customer’s control.

Authoritative rebuilds remain comparators and reconciliation mechanisms. There is no public runtime download.

State Equivalence ContractAccess scope

Reproducibility evidence

Demonstrated / artifact-specific

Output hashes identify computation bytes. Signed receipts bind a payload to a key. Reference checks test numerical agreement. None is a substitute for the others.

LuxiBook supplies a public option-book demonstration and receipt checks. Public evidence packs supply scope, raw artifacts, and verification instructions. They are not a public release of every measured engine.

LuxiBook downloads and receiptsHow to verify public evidenceReceipt verification boundaries

The handoff is a matrix, not your model

Customers keep their models, market data, pricing logic, scenario generation, trading logic, methodology, and authoritative rebuild process. Where technical integration requires it, Luxi receives canonical little-endian f64 values in scenario-major form.

Maintain and reconcile state

Luxi applies declared updates to the matrix state and compares incremental histories with authoritative fresh rebuilds.

Reproduce the final state

The Luxi State Equivalence Contract requires incremental and fresh-rebuild paths to reach the same Luxi numerical state and authenticated result for the same final matrix.

Measure update and authentication cost

Update shape, CPU state-maintenance work, and optional GPU authentication are measured as distinct costs.

Carry supporting lineage

Declared contracts, hashes, and incremental lineage can be bound in a signed, independently verifiable record. The receipt supports the control layer; it is not a regulatory opinion.

Risk State Engine / evaluation contract

The Luxi State Equivalence Contract

Incremental updates must produce the same Luxi numerical state and authenticated result as a fresh rebuild of the same final matrix. An evaluation checks final-state agreement, reconciliation, and signed receipt verification under the agreed contract. These are available evaluation checks, not completed independent product validation. Fresh rebuilds remain authoritative comparators and reconciliation mechanisms.

Technical scope. The contract applies to Luxi's declared representations, operations, and execution rules. It does not promise bit identity with arbitrary customer or vendor rebuilds or universal permutation independence.

CPU computation, optional GPU authentication

Canonical CPU path

CPU computes hierarchical Neumaier totals, empirical VaR/ES, and tail state.

Optional NVIDIA path

CUDA hashes canonical two-dimensional Merkle tiles on a resident mirror. It does not price, generate scenarios, compute VaR/ES, move arbitrary quant code, or accept a customer-owned CUDA pointer.

On a 1 GB resident state, 10% scatter authentication was about 11.1× faster on H100 NVL than CPU hashing with matching roots. Tiny and column updates stayed on CPU. This is internal H100 NVL evidence; H200 and B200 are unmeasured for Risk Runtime.

That ratio measures tile authentication only, not Risk State Engine or end-to-end application speed.

On CPU, the cost of one update stayed nearly flat as the matrix grew. With 1,000 scenarios and the default layout, the median single-column update took 1.18 ms at 0.4 GB, 1.27 ms at 4 GB, and 1.42 ms at 40 GB, each final state checked against a fresh build. Replaying 73,288 public Deribit BTC and ETH option trades from November 5 and 6, 2024 into a 482,000-position by 1,000-scenario matrix (3.86 GB), the median was 1.20 ms per trade, and all 192 checkpoints matched a fresh rebuild of the same final matrix bit for bit. A smaller 1.45 GB replay on an Apple M1 Pro laptop also matched at all 192 checkpoints.

These are internal measurements on one x86-64 Linux cloud VM (17 vCPUs, CPU only) on October 26, 2026, using the engine code released as v0.4.2. The 0.4 to 40 GB matrices are synthetic; in the replay the trades are real and the account split and opening positions are synthetic. They measure state maintenance, not pricing, scenario generation, trading-path latency, or a customer result.

Evaluate Risk State Engine directly from OCaml

The evaluation package now includes an OCaml library over the existing C ABI. It brings Risk State Engine into an OCaml workflow without exposing customer models or changing the core runtime contract.

Native workflow integration

Use handle lifecycle, Bigarray ingest, v0.3 ingest, incremental apply_cell and apply_column updates, signed receipt access, and accelerator enablement through the same interface. A dune file is included for integration into a dune workspace.

Verified against the runtime

CPU Rust tests: 66/66 on the v0.4.2 evaluation package, on macOS arm64 and Linux x86-64. The OCaml test covers the tiny 4×2 fixture, receipt verification, v0.3 ingest, cell and column updates; the demo matches the C fixture. When CUDA is not built and enabled, accelerator enablement correctly returns unsupported.

Technical note. Production packaging conveniences such as opam distribution, snapshot and batch helpers, CI packaging, and CUDA control from OCaml are still being expanded.

Request access

Risk Runtime v0.4.2 is proprietary and has no public download. The evaluation package includes macOS arm64 and Linux x86-64 builds, and each build records the source commit it was built from. The default evaluation binary has CUDA off unless explicitly rebuilt with CUDA. An evaluation defines the workload, State Equivalence Contract, hardware, comparator, and acceptance criteria, then reports only what that protocol measures.

Evaluate your workloadContact LuxiEdge

Separate access paths

LuxiBook v0.2.1

The downloadable public receipt and determinism demonstration. Its option-pricing CPU/CUDA evidence is not evidence that Risk State Engine numerical aggregation runs on those GPUs.

Downloads and receipt evidence

LuxiSize (part of LuxiBook)

A preserved public offline position-sizing demonstration, included in the free LuxiBook download and outside the institutional Risk State Engine sales path.

LuxiSize download

LuxiEnergy

AI Energy Systems research remains accessible with measured workload and energy boundaries.

Explore LuxiEnergy

Your application

Start with a concrete computation and a decision the evidence should support, not an assumed benefit.

Application questions · Technical evaluation