How Luxi works, in plain English

Luxi has two engines. LuxiQuant is a deterministic numerical engine, ready for pilot. The Luxi Inference Engine runs transformer workloads in stages: prefill and decode. Everything below explains what each engine does, and how energy and determinism are measured against explicitly defined units.

The Two Engines

LuxiQuant Risk State Engine accepts a customer-supplied scenario-by-position numerical matrix and maintains deterministic incremental risk state under named contracts. CPU is canonical for hierarchical totals, empirical VaR/ES, and tail state; optional CUDA authenticates canonical tiles and does not price or generate scenarios. It is available through a paid confidential evaluation. LuxiBook remains the separate downloadable public option-pricing and receipt demonstration.

The Luxi Inference Engine runs transformer workloads - the kind of models behind modern AI assistants - in two stages:

  • Prefill

    Reads the prompt. This stage has an independent measured baseline and a stronger internal matched result, kept clearly separated as research evidence.

  • Decode / Generation

    Produces output tokens. This stage matches the reference next-token result under the defined numerical audit; the next serving-readiness milestone is locking end-to-end quality, throughput, and board energy.

Energy and determinism are always measured against explicitly defined units - for example, GPU-board joules per prefill position, or a byte-identical output hash under a documented configuration. A number without its unit and scope is not a claim we make.

The Problem

Where Floating-Point Reproducibility Breaks Down

The largest reproducibility risks typically arise from order-sensitive reductions, fused versus unfused operations, different math-library implementations, precision changes, compiler optimization choices, backend-specific execution, and uncontrolled concurrency or state.

When those factors are uncontrolled, small rounding differences compound - and the same code with the same input can produce different low-order bits across runs, builds, or machines.

This breaks:

  • Audits

    Without receipts, yesterday's results can't be reproduced - by anyone. LuxiQuant's receipts make reproduction routine.

  • Certifications

    Regulators require predictable behavior.

  • Debugging

    Receipts pin down the exact run, so a failure can be replayed and fixed instead of guessed at.

The Solution

Lu(x)iEdge Makes the Execution Path Reproducible

Lu(x)iEdge controls the factors that cause drift, so supported workloads can be reproduced and compared byte-for-byte under the documented conditions.

01

Defined execution order

Operations follow the documented engine path rather than an uncontrolled reduction order

02

Canonical requests and outputs

Inputs, modes, and output serialization are normalized for reproducible comparison

03

Rust-based systems core

Rust reduces common classes of memory-safety risk

04

signed receipts

Receipts allow byte-level output comparison when the same documented conditions are reproduced

Determinism claims apply to the documented engine version, execution mode, precision, canonical input, and supported/tested platform combination.

Optimization Trade-offs

Why Not Just Disable Optimizations?

Some deterministic engines - physics libraries such as Box2D and Rapier are well-known examples - accept reduced acceleration in exchange for reproducibility. For games, that tradeoff works.

For quant finance pipelines, or safety-critical systems that need both real-time performance and reproducible evidence, giving up acceleration is a hard price to pay. (Luxi Book itself does not do Monte Carlo or VaR - it prices a supplied option book.)

Lu(x)iEdge is designed to preserve defined reproducibility properties while using accelerated CPU and GPU paths. Performance and supported behavior depend on backend, precision, operation, and workload.

Historical / research numerical-engine paths (not Luxi Book)

The public Luxi Book binaries today are macOS ARM64 (CPU), Linux x86_64 (CPU), and Linux x86_64 CUDA. There is no macOS GPU Book binary, and Metal, Vulkan, RISC-V, and WASM are not Luxi Book backends - the table below describes the broader numerical-engine research surface only.

Backend Status
CPU with SIMD (x86-64) Available and tested in the evaluated environments
ARM64 and ARM Neon Internal benchmark: direct-engine, no-service-overhead run with the receipt contract. Independent validation next.
NVIDIA CUDA (H100 tested) Available and tested in the evaluated environments
Vulkan Available but experimental
Apple Metal Implemented and internally exercised through the portable wgpu GPU path on macOS. Same-hardware reproducibility follows the declared GPU path and receipt contract, scoped to the declared path.
Bare-metal-capable / no_std core path Implemented in source; hardware run scheduled.
RISC-V Historical work reported; re-validation on the roadmap.
WASM On the validation roadmap.

Platform evidence note: Edge and constrained ARM systems are deployment environments for the LuxiQuant engine, not a separate product. Implemented means a supported code path exists. Internal benchmark means LuxiEdge has exercised that path in its own testing or demonstration. Historical work means earlier results are reported with re-validation on the roadmap. Validation roadmap means the evidence run is planned.

TestFort numerical-engine footnote (December 2025, not Luxi Book): in the defined H100 evaluation, the tested seven-function workload sustained 286.94 billion aggregate operations per second.

Verification

Compare Outputs Byte-for-Byte

Every Lu(x)iEdge response includes a SHA-256 hash of the output. Store the hash at computation time; re-run the same canonical input later under the same documented engine version, mode, and configuration, and a matching hash confirms the output bytes are identical. A complete reproducibility record should also preserve the canonical input, engine version, execution mode, precision, backend, and configuration - an output hash alone does not prove who ran the computation or which input produced it.

// Luxi Book: price a CSV book, keep the receipt
$ ./luxi-book price --book example_book.csv --out report.csv

# writes report.csv (prices + five Greeks)
# and receipt.json with the run's signed receipt

$ cat receipt.json | grep sha256
"sha256": "4a21b1e708fa5c69..."

# re-run the same book under the same binary and inputs,
# then compare the SHA-256: a match confirms identical output bytes

Efficiency

Race to Idle

For a fixed workload, faster completion can reduce total board energy when the increase in power is smaller than the reduction in execution time. This must be measured for each workload - race-to-idle is not an automatic guarantee of facility or cooling savings.

TestFort numerical-engine footnote (December 2025, not Luxi Book): the tested workload measured 2.35 billion operations per joule under TestFort QA Lab's defined test. TestFort independently evaluated a defined non-linear numerical workload, repeated hash consistency, CPU/GPU output agreement in the tested H100 environment, API load behavior, and sustained GPU execution. That evaluation did not certify every function, every platform, every backend, or every possible expression.
See the measured evidence

Deployment

The engine ships as a single compact binary that runs locally with no cloud dependency. On supported builds it selects an appropriate backend for the host - SIMD on CPU, CUDA on tested NVIDIA GPUs - with minimal configuration. Supported operating systems, GPU APIs, binary size, and startup behavior are documented per release with the published downloads.

The Art of Fugue

We Needed a Benchmark That Plays a Symphony

Most benchmarks test isolated operations. That is like playing a single note on a piano.

The Art of Fugue is a defined reproducibility workload combining several numerical expressions and execution patterns - three concurrent "voices" of conflicting intensity: trigonometric identities, logarithmic decay, and discontinuous transcendentals. Its purpose is to expose drift and compare receipts across documented implementations and environments.

The output is captured as a single SHA-256 hash, so any drift between documented runs shows up as a hash mismatch. The public GitHub repository contains the workload definition and the published receipts.

The Art of Fugue covers the core trigonometric and transcendental function suite. Separately, TestFort QA Lab independently evaluated a defined deterministic numeric workload, including the statistical function suite and GPU endurance testing (December 2025); that evaluation did not certify every function, platform, backend, or expression.