LuxiQuant Risk State Engine

Defined contract / Available evaluation checks

State equivalence and signed verification

The evaluation checks whether incremental updates and a fresh rebuild of the same final matrix reach the same Luxi numerical state and authenticated result. Reconciliation, update cost, and signed receipt verification are checked against the agreed scope.

These are the defined contract and available evaluation checks, not a claim of completed independent product validation.

Internal H100 NVL authentication evidence

Authentication cost on a resident state

On a 1 GB resident state with 10% scatter updates, tile authentication was approximately 11.1× faster than CPU hashing, with matching roots.

This measures authentication only, not risk-engine, pricing, VaR, inference, or end-to-end application speed. Tiny and column updates stayed on CPU.

Internal reproducible evidence

Insurance/catastrophe

Oasis stock defaults produced 10 different hashes in 10 runs across 2 machines; Luxi produced 1.

Extreme-TIV stock differed on 5,851 of 6,006 totals; Luxi had no differences.

Internal reproducible evidence

GPU numerical accuracy

For a sum of 10 million float32 values, stock error was +$14,930.04; Luxi was exact at $0.

Luxi produced the same hash on L4, x86 and Apple.

Cross-platform evidence

Orbital

SGP4 jax100/starlink was byte-identical on macOS arm64, Linux arm64, qemu x86_64 and H100 near-earth.

Internal matched research evidence

AI inference research

Batch-invariant generation at normal-engine generation speed; outperformed the tested deterministic modes.