Deterministic pricing · measured 2026-08-16 · engine v0.2.1

Five different GPUs.
One identical hash.

LuxiBook priced the same 237-row option book on an H200, an H100, an A100, a 5090 and a 4090 - four NVIDIA architecture generations, CPU path and CUDA path, ten runs - and every run produced the hash below. Download the binary, price the book, and check that you get it too.

NVIDIA H200 sm_90 · avx512 host NVIDIA H100 80GB sm_90 · avx512 host NVIDIA A100-SXM4 sm_80 · avx2 host GeForce RTX 5090 sm_120 · avx512 host GeForce RTX 4090 sm_89 · avx2 host 902667a1070b83bff57ac642cf16779d 998b5a954c046c450b154237e9e196e2 output_vector_sha256 · identical on all ten runs book total 488338590.9188194 · 237 rows · mixed Black-Scholes and Black-76

Do not take our word for it. Take four minutes.

The binary is public, the book is public, and the expected hash is printed above. If your machine disagrees with ours, you have found something we want to know about.

your machine
# 1. get the binary and the book
$ curl -sLO .../downloads/luxibook/luxi-book-linux-x86_64
$ chmod +x luxi-book-linux-x86_64

# 2. price the book
$ ./luxi-book-linux-x86_64 price --book stress_book_a.csv \
      --out out.csv --receipt r.json

  engine_version   0.2.1
  git_sha          02388f778d9017d251bb0fc905a82b89a2e94c9a
  output_hash      902667a1070b83bf...e9e196e2
  book_total       488338590.9188194
  receipt          lxq2_TFhRAgCLCm2h...

# 3. check it
$ ./luxi-book-linux-x86_64 verify --receipt r.json
  [OK ] signature        Ed25519, install b487b0ab3ec0008e
  [OK ] input hash       matches stress_book_a.csv
  [OK ] output hash      matches out.csv
  [OK ] canonical bits   66 f64 little-endian
  [OK ] mode             cpu scalar deterministic
  exit 0

# 4. now try to break it: change one byte of out.csv
$ ./luxi-book-linux-x86_64 verify --receipt r.json
  [FAIL] output hash      does not match out.csv
  exit 1
20/20
Published receipts whose Ed25519 signature verifies
independently checked · 0 invalid
5
Distinct install keys, so you can pin to one specific machine
pin with --expect-pubkey
4
NVIDIA architecture generations that agreed, bit for bit
sm_80 / sm_89 / sm_90 / sm_120

What a receipt proves, and what it does not

A receipt is not a claim that the number is correct. It is a claim about provenance, and the exact wording matters.

The claim: the install holding key K asserted this output vector hash, from this input hash, on this build, on this device, using this instruction-path label - and the assertion was made no earlier than the drand beacon round recorded in the receipt.

Three layers, often confused

SHA-256 Content fingerprint of the input CSV and the output numbers. Answers "are these the exact same numbers?"
Ed25519 The seal, the lxq2_ token. Answers "which install asserted them?"
drand round A public not-before time anchor. Answers "no earlier than when?"

The honest limit

The public key travels inside the receipt. So a checker that reads the key out of the receipt it is checking proves only that the receipt is internally consistent, not that it came from us. Anyone with Python can mint one that passes that test.

Pinning is the defense. Pass --expect-pubkey with a key you already trust from the published evidence, and it becomes a real identity check. We publish five distinct install keys precisely so you can.

How receipts work, in detail

Independently measured: less energy per position

TestFort QA Lab, July 23 2026. Qwen2-7B-Instruct packed prefill, batch 16, sequence length 128, one NVIDIA H100 80GB. Board joules per prefill position, lower is better. This is our inference work, not LuxiBook.

LuxiEdge prefill
0.018718
vLLM (default)
0.019316
vLLM (batch-invariant)
0.020604
And the part most vendors would bury: that came at 80.6% of default vLLM throughput. You buy joules per position here, not positions per second. If your constraint is a power envelope, that is the trade you want. If it is latency, it is not.

Who this is for

You have to prove a number to someone who does not trust you

Model validation and model risk teams, fund administrators and valuation agents, audit and expert-witness work. If your job includes reproducing what was computed last quarter and documenting that you did, a signed receipt is not a feature, it is your deliverable.

You buy watts

Capacity and platform teams measuring inference against a power envelope rather than a latency target. Start with the independently measured prefill numbers above, then bring your own workload and comparator.