Evidence register · labeled by source

Every number, its source, and its limits.

This page is the claims register. Independent evaluations, LuxiQuant and LuxiBook evidence, Luxi Inference research, and internal development results are kept in separate groups. Nothing here is merged into one system-wide win.

Independent LuxiQuant and LuxiBook Luxi Inference Internal development

How to read our claims

Each result is labeled by source, workload, hardware, date, and validation status. Version tags such as 99, 100, or 101 are internal milestone tags, not product names.

Independent Evidence

Measured or operated by a third party (TestFort QA Lab) under its own protocol, with dated reports. The strongest evidence class on this page.

Internal Matched Evidence

Internal head-to-head measurement against a named comparator under a stated protocol. Independent validation is noted wherever it exists.

Research Evidence

Internal research-path measurements. Directional evidence about architecture and energy behavior. Not product claims, not production-serving evidence.

Maturity labels

Paid Pilot Open (LuxiQuant / LuxiBook) and Scoped Evaluation Open (H100 prefill) mark commercial engagements that measure your workload under contract.

How to read our numbers. Independent evidence is measured by a third party; internal evidence is measured by us. Prefill results count evaluated prompt positions per second; decode results count generated tokens. Energy is GPU-board energy via NVML on the stated workload. Facility, wall-plug, and PUE-adjusted energy are different quantities we do not claim.

Group 1

Independent evaluations

Two separate TestFort QA Lab reports. The December 2025 report is a seven-function numerical engine. The July 2026 report is packed prefill. Neither is LuxiBook.

Numerical engine (not LuxiBook) · Independent Evidence · TestFort, December 2025

Numerical engine evaluation, TestFort, December 2025

TestFort QA Lab independently evaluated a defined seven-function numerical suite on NVIDIA H100 hardware. This is not option-book pricing and not example_book.csv. It is never applied to LuxiBook's receipt check, and never applied to transformer results.

286.94B

Aggregate operations per second

Tested seven-function workload

331.13B

Peak operations per second

Tested square-root operation

444.4T

Reported operations

One-hour load test

1.47 ms

p95 API latency

Reported 200-user load configuration

Reliability

No failed requests in the specified load test
Average GPU power of approximately 117.2 W during the reported load test

Determinism

Five GPU and five CPU runs produced identical hashes for the tested workload

Scope: TestFort independently evaluated a defined deterministic numeric workload on an NVIDIA H100 SXM. These results apply to the tested workload and configuration and are separate from LuxiEdge's transformer research.

Luxi Inference · Prefill · Independent Evidence

Independent packed-prefill baseline, July 23, 2026

TestFort third-party measurement completed 2026-07-23. Workload: Qwen2-7B-Instruct packed prefill, batch 16, sequence length 128, one NVIDIA H100 80GB, measured in prefill positions. Preserved as dated independent evidence, reported separately from later internal prefill measurements. Full narrative also lives on Research.

3.10%

Lower board J/position vs default vLLM

0.018718 vs 0.019316 J/position

9.15%

Lower board J/position vs batch-invariant vLLM

0.018718 vs 0.020604 J/position

80.60%

Of default vLLM throughput

28,374.7 vs 35,203.1 positions/sec

91.78%

Of batch-invariant vLLM throughput

28,374.7 vs 30,914.3 positions/sec

Configuration	Throughput (positions/sec)	Board J/position	Throughput ratio	Energy ratio
LuxiEdge prefill build	28,374.7	0.018718	-	-
vLLM (default)	35,203.1	0.019316	LuxiEdge 80.60% of vLLM	LuxiEdge 3.10% lower
vLLM (batch-invariant)	30,914.3	0.020604	LuxiEdge 91.78% of vLLM	LuxiEdge 9.15% lower

Independently verified: 3.10% lower board joules per prefill position than default vLLM (9.15% lower than the batch-invariant configuration), at 80.6% of default vLLM throughput on this workload. Lower energy, lower throughput. GPU-board energy is not facility energy.

Scope: packed prefill, Qwen2-7B-Instruct, batch 16, sequence length 128, one H100 80GB. The tested LuxiEdge path had stable pinned-environment fingerprints, zero measured cross-sequence contamination, and zero FlashAttention fallback. The tested build carries the internal milestone tag Version 99. Prefill positions only. Not a full serving or decode comparison.

Report: TestFort validation report, July 2026 (PDF) · Evidence pack

Group 2

LuxiQuant and LuxiBook evidence

LuxiBook is the downloadable option-book pricer. Its cross-box receipt check is an internal measurement covering two books on five machines (August 16, 2026), not a third-party evaluation. The TestFort December 2025 numerical-engine suite above is a different workload and does not cover LuxiBook.

LuxiBook · Internal Evidence · Cross-box receipt check

LuxiBook: two books, one hash each, five GPUs, four architecture generations

LuxiBook is the downloadable deterministic option-book pricer (European Black-Scholes / Black-76 with five Greeks and a signed receipt per run). Every run emits a signed receipt: a SHA-256 fingerprint of the output numbers, sealed with an Ed25519 signature from the install that produced them. Running the published example_book.csv and a 237-row stress book through the published v0.2.1 binaries on August 16, 2026 produced one output hash per book on every machine tested, on the CPU path and through the black_scholes_book_kernel CUDA path. Every run emitted a signed receipt, and all twenty are published alongside the binaries.

Field	Value
Books	`example_book.csv` (10 rows, published) and a 237-row stress book
Output hash - example book	`4a21b1e708fa5c694bf48237df5e5bd3b94599e6273d07986283c6c6b8e3c97a`
Book total - example book	`151.70016507843832`
ATM_CALL price	`10.4505835721856215`
Output hash - stress book	`902667a1070b83bff57ac642cf16779d998b5a954c046c450b154237e9e196e2`
Book total - stress book	`488338590.9188194`
Engine	v0.2.1. The stored receipts record `engine_version 0.2.0` because they were measured before the version string was bumped.
Receipts published	20, each Ed25519-signed, across 5 installs
Receipts folder	downloads/luxibook/evidence/v0.2.1-phi-fix-matrix
Date	August 16, 2026

Machine	Architecture	Compute capability	CPU path	CUDA path
NVIDIA H200	Hopper	9.0	match	match
NVIDIA H100	Hopper	9.0	match	match
NVIDIA A100	Ampere	8.0	match	match
NVIDIA RTX 5090	Blackwell	12.0	match	match
NVIDIA RTX 4090	Ada	8.9	match	match

Every row is one physical machine with its own install key, and each ran both books on both paths - 20 signed receipts in total. Each machine's compute capability is recorded in its receipt, so the four distinct generations are checkable rather than asserted. Apple Silicon has a published binary and was covered by the earlier check, but it is not part of this matrix; there is no macOS GPU binary.

Scope: the identical receipt holds for this book, these boxes, and this kernel. It is not a claim that all GPUs always match, and not a universal cross-platform bit-exact promise. The example_book.csv result you can reproduce yourself with the published binaries and the published book. The stress-book result is reproducible on request, under evaluation, because that book is not published. Both published hashes have also been reproduced on an x86-64 Linux machine that was not part of the test fleet.

Verify a receipt signature yourself - no binary needed.

Each JSON receipt is Ed25519-signed. The payload field is the signed plaintext; the receipt field carries the signature. Verify in three lines of Python using only the standard cryptography package:

# pip install cryptography
import base64, json
from cryptography.hazmat.primitives.asymmetric.ed25519 import Ed25519PublicKey
from cryptography.exceptions import InvalidSignature

with open("receipt.json") as f:
    data = json.load(f)

# Receipt blob layout: 5-byte header + 32-byte pubkey + 64-byte Ed25519 sig
s      = data["receipt"][len("lxq2_"):]
pad    = (-len(s)) % 4
raw    = base64.urlsafe_b64decode(s + "=" * pad)
pubkey = Ed25519PublicKey.from_public_bytes(raw[5:37])
try:
    pubkey.verify(raw[37:], data["payload"].encode())
    print("PASS: signature valid")
except InvalidSignature:
    print("FAIL: signature invalid")

The signer_pubkey field is the Ed25519 public key in hex, and the payload field lists every signed field in plain text, so you can audit what was covered before verifying. Note that the key travels with the receipt, so the check above proves the receipt is internally consistent. To prove it came from a specific machine, pin the key: pass --expect-pubkey <hex> to verify_receipt.py (view on GitHub, published in the LuxiDemo download folder) using a key you already trust from the published evidence.

LuxiBook product page · Download and run the check

Group 3

Luxi Inference research evidence

Transformer prefill and decode research. The independent TestFort prefill baseline is in Group 1 and is not repeated as a new fact here. Later internal matched and decode results stay labeled internal. This is not a production serving replacement. Batch-specific prefill is not end-to-end inference. Narrative: Research.

Luxi Inference · Prefill · Internal Matched Evidence

Internal matched comparison, batch 16

Matched internal comparison on the locked Flash, device-resident FP16 path. Qwen2-7B full-stack prefill, sequence length 128, batch 16, one NVIDIA H100 80GB-class GPU, sequential comparison arms, matched prefill positions. Internal, same-GPU, prefill-stage evidence. August 7, 2026. Separate from the July 23 TestFort baseline.

~1.18x

Of the tested vLLM throughput at batch 16

~41,221 prefill positions per second on the Luxi arm

~12%

Lower board joules per prefill position than the tested vLLM at batch 16

~0.0169 J/position, NVML GPU-board energy

<1%

Throughput variation across measured runs

Stability of the matched comparison

Pass / fail

Determinism contract: dual-run agreement on the measured Luxi path

Run-to-run reported behavior, not a receipt or output-token identity claim

Batch-32 supporting detail: ~43,464 positions per second at ~0.0158 J/position, ~1.19x the tested vLLM at ~14% lower board energy. No matched B72 vLLM arm yet. This ratio is never attached to the batch-72 absolute result below. Independent reproduction is the next step.

Matched evidence pack

Luxi Inference · Decode/Generation · Internal Matched Evidence

Faithful full-model CUDA correctness

Faithful full-model Qwen2-7B CUDA path on NVIDIA H100, measured against CPU and Hugging Face references. Correctness contract: under the defined acceptance protocol, what repeats is the greedy token sequence. A pass/fail gate, not a performance claim.

Full 28-layer

Qwen2-7B on NVIDIA H100

Faithful full-model CUDA path

1.0

CPU/CUDA greedy-token match rate

Current acceptance protocol

1.0

HF/CUDA greedy-token match rate

Current acceptance test

~1.0

Logit cosine similarity

Top-10 overlap 10/10

Current status

Full 28-layer Qwen2-7B CUDA path on NVIDIA H100: running
CPU-versus-CUDA greedy-token match rate: 1.0 under the current acceptance protocol
Hugging Face-versus-CUDA greedy-token match rate: 1.0 under the current test
Logit cosine similarity: approximately 1.0. Top-10 overlap 10/10
Four-request independent CUDA generation smoke test: passed
All decoder mathematics runs on the GPU
Current limitations: token embedding row gathering and final greedy argmax remain host-side
Faithful-path performance, energy optimization, and production-serving validation: ongoing

Correctness results on this faithful path are separate from the TRADE energy benchmark below. Energy and throughput numbers are never carried across the two.

Luxi Inference · Decode/Generation · Research Evidence

Internal prototype · numerical audit complete

True-autoregressive decode, reference agreement achieved

Decode/generation produces output tokens. The GPU path matches the reference next-token result under the defined numerical audit. The next measurement milestone is post-audit end-to-end quality, throughput, and board energy.

Provenance, August 6, 2026: true-AR throughput and board-joules figures published before this date are superseded and are not current product proof. The GPU route was rebuilt to the full reference mathematics.

Luxi Inference · Decode/Generation · Research Evidence

Internal Research · data pack in preparation

Fixed-window verifier research, evaluated positions

An evaluated-position/verifier primitive on the decode research path. Work unit: evaluated positions per second, never generated tokens.

~4,669

Evaluated positions per second

Internal measurement. Owner-supplied. Data pack in preparation

Luxi Inference · Research archive · Research Evidence

Internal TRADE research vs Hugging Face FP16

Separate internal controlled benchmark: TRADE research stack vs Hugging Face Transformers FP16 reference. This is not a production-peer comparison. It uses a different execution path and a different baseline than the matched-vLLM measurements.

Up to 86%

Lower board energy per processed token

~86.3% reduction measured

3.56x

Processed-stack throughput

~32,083 vs ~9,026 stack-tokens/sec

~145 W

LuxiEdge median board power

Reference: ~298 W

H100 NVL

Qwen2-7B-Instruct, seq 128

Sustained-load measurement

Metric	LuxiEdge TRADE research stack	Hugging Face Transformers FP16 reference	Result
Board energy per processed token	~0.00453 J/token	~0.03296 J/token	~86.3% reduction
Processed-stack throughput	~32,083 stack-tokens/sec	~9,026 stack-tokens/sec	~3.56x
Median board power	~145 W	~298 W	~51% lower

NVIDIA H100 NVL. Qwen2-7B-Instruct. Sequence length 128. Sustained-load measurement. Controlled internal H100 research benchmark. Distinct from faithful-chat equivalence, vLLM comparisons, and production-serving evidence.

Research Evidence

Long-context research

Memory scaling for the selectable long-context research path.

Sequence length	Waller state (MB)	Dense scores (MB)	Reduction
1,024	0.52	8.4	16x
4,096	2.1	134	64x
8,192	4.2	537	128x
32,768	16.8	4,295	256x

Memory scales near-linearly vs dense O(N^2). Pack

Group 4

Internal development results

Absolute prefill champion, Llama 3.1 milestone, historical TRADE ladders, and version-lineage packs. Internal unless a third-party report is named.

Luxi Inference · Prefill · Internal Absolute Evidence

Internal absolute prefill champion, dual_gemm at B72

The fastest measured configuration of the prefill path to date, reported as an absolute internal result with no vLLM comparison arm attached. One NVIDIA H100 80GB HBM3; Qwen2-7B-Instruct-class weights; sequence length 128; batch 72; dual-GEMM; Flash attention; device-resident FP16 path; median of five 15-second runs. August 7, 2026.

~44,860 /s

Prefill positions per second

Median of five 15-second runs

~0.01532 J

Board joules per prefill position

NVML GPU-board energy

No matched B72 vLLM arm exists, so the ~1.18x batch-16 ratio is never applied here. The Door-B receipt check passed for independent-process stack-fingerprint equality under the dual-GEMM environment at batches 1, 16, and 32. The B72 throughput loop itself did not emit hashes. This is not a Door-A AUDIT bit-exact B72 result and not a universal determinism claim.

Absolute champion evidence pack

Luxi Inference · Decode/Generation · Internal Matched Evidence

Llama 3.1 internal milestone, independent validation next

Faithful Llama 3.1 inference, internal milestone

Internal Llama 3.1 resident-inference milestone. Active-batch output matched the serial BF16 path on tested correctness gates. Independent validation of performance and energy is the next step.

Metric	Internal result	Scope
Correctness	Active batch = serial	Exact on primary p32 and p128 gates
Cold process/model load	about 190 to 210 seconds	Current development startup path

Internal benchmark, not an MLPerf submission. Energy is GPU-board energy.

Research Evidence · historical

Historical research configurations

Historical research archive, preserved with full methods and raw traces. For current claims, see the sections above.

Earlier public benchmarks (LuxiDemo, 2026-07-11)

Earlier public evidence from a separate TRADE sustain protocol. Open any pack to inspect methods and raw traces.

7B-class TRADE, energy and throughput

Hardware: 1x NVIDIA H100 NVL. 28-layer stack. public pack.

Sequence length	Tokens / sec	Joules / token	Median power (W)	Notes
5	~44	3.560 +/- 0.005	~156	Short-burst reference
32	~221	-	-	Throughput sweep
64	~359	-	-	Throughput sweep
128	~403	0.630 +/- 0.002	~254	Ladder primary (this protocol)
256	~464	~0.604	~280	Longer context
512	~493	-	-	Throughput sweep

Multi-run at sequence length 128, from the earlier public sustain ladder. A different protocol than the Qwen2-7B benchmark above; the two are not directly comparable. Source: evidence/h100-7b-class-TRADE/

Head-to-head with a Flash baseline

12 layers. Sequence 1024. h=768. About 10 s each side.

Metric	TRADE 12L	PyTorch + Flash 12L	Ratio
Median power (W)	177.2	176.4	~1.0x
ms / stack forward	74.32	3.90	TRADE 19.1x slower
Prefill tok/s	13778	262859	PT 19.1x higher
Joules / token	0.0129	0.0007	TRADE 19.2x higher

h100-stack12-H2H . We publish where a standard baseline wins on this shape.

Evidence packs

Open on GitHub. Read methods. Check traces.

Absolute champion

prefill_accel_lock_20260807T233111Z

Internal absolute prefill champion: dual-GEMM at batch 72, ~44,860 positions/s at ~0.01532 board J/position, median of five 15-second runs. Absolute internal result; no matched vLLM arm.

Matched comparison

prefill_freeze_matched_20260807T210749Z

Internal matched comparison at batch 16: ~1.18x the tested vLLM throughput at ~12% lower board J/position, with batch-32 supporting data. Qwen2-7B-Instruct FP16, seq 128, sequential comparison arms.

Independent baseline

h100-qwen2-7b-v99-matched-prefill-2026-07-23

Independent TestFort prefill baseline (July 23, 2026): 28,374.7 positions/s at 0.018718 board J/position. 80.60% of default vLLM throughput at 3.10% lower board energy per position.

Milestone lineage

version-100-h100-gtm

Earlier internal matched prefill evidence from the internal milestone lineage (version tags are internal milestones, not products). Superseded as the current headline by the August 7, 2026 packs above.

Earlier protocol

h100-7b-class-TRADE

Full 28-layer 7B-class TRADE stack from the earlier public sustain ladder. Multi-run throughput + energy (~0.63 J/tok @ seq=128, ~403 tok/s under that historical protocol).

Energy

h100-stack12-TRADE-cuda

Device-resident 12-layer stack energy (GPT-2 width).

Honesty

h100-stack12-H2H

TRADE vs PyTorch Flash. Complete head-to-head, both sides published.

AUDIT

h100-WNSM-free-ride

Null-space free-ride under GPU load.

Memory

h100-LONGCTX-scaling

O(N) vs O(N^2) memory ladder + CUDA 32k.

Micro

h100-BASELINE-vs-geo

Single-layer baseline wedges. Not a 7B throughput claim.

Serve

h100-serve-sustain-2026-07-11

Continuous-batch serve sustain with power traces.

Open LuxiDemo on GitHub

Download

Downloads live on the demo page: LuxiBook, the option-book pricer with a signed receipt on every run, and LuxiRisk, a free secondary position-sizing CLI.

Try LuxiBook LuxiRisk Evaluate Your Workload

The older numerical REST demo builds remain available in LuxiDemo releases.

Two different kinds of hashes appear around these builds: release-file checksums verify that the file you downloaded was not corrupted; evaluation-output receipts attest what a specific computation produced. They are not interchangeable.