papers/lux-blockchain-scaling-laws/sections/06-honesty.tex

\section{Honest Scope Statement}
\label{sec:honesty}

This paper reports what we measured. It does not extrapolate to
hardware we do not own or workloads we did not run. This section
pins the boundary explicitly.

\subsection{What was directly measured}

\begin{itemize}[nosep,leftmargin=2em]
\item NTT throughput sweep on Apple M1 Max: 10 batch sizes $\times$
3 backends (CPU, Metal native, WGSL via wgpu-native 29.0.0). Source:
\texttt{luxcpp/crypto/corona/test/lattice\_ring\_sweep\_bench.cpp}
($\sim$\,260 LoC) commit \texttt{luxcpp/crypto@bb9e9613}. Raw TSV in
\texttt{papers/lux-blockchain-scaling-laws/data/}.

\item ZAP message round-trip: \texttt{luxfi/consensus/bench/
zap\_throughput\_test.go::BenchmarkZAPThroughput} on M1 Max,
\texttt{benchtime=2s}, $5.47$ \textmu s/op.

\item TFHE bootstrap: \texttt{luxfi/fhe/bench\_crdt\_test.go::
BenchmarkBootstrapSingle}, $119.9$ ms/op single-threaded pure-Go.

\item Pulsar reshare $n \in \{21, 64, 128\}$:
\texttt{luxfi/threshold/protocols/lss/lss\_pulsar\_bench\_test.go::
BenchmarkPulsarReshare}, single-process serial, $12.67$/$156.61$/$1212.74$ s
end-to-end.

\item Cross-runtime KAT release gates (Go $\leftrightarrow$ C++):
$991$ assertions across codec / modarith / NTT / sample. All pass
byte-equal.
\end{itemize}

\subsection{What was NOT measured}

\begin{itemize}[nosep,leftmargin=2em]
\item \textbf{NVIDIA H100 / RTX / consumer GPU throughput.} No NVIDIA
hardware on the measurement host. The CUDA driver compiles but live
device dispatch needs CI-runner provisioning. CI workflow IDs
attempted on 2026-05-04 (\texttt{25304130314}, \texttt{25311210772},
\texttt{25313586744}, \texttt{25314213586}) all cancelled or stayed
queued; runner pool was either CPU-only or saturated.

\item \textbf{Block-STM parallel EVM throughput.}
\texttt{luxfi/evm/core/parallel/blockstm.go} ships, with unit tests
passing, but no Go-side \texttt{Benchmark} harness exercises it
under contention. Adding one (with synthetic read/write set
generator + version-conflict distribution) is roughly a day of work
and is queued for v0.3.0.

\item \textbf{DEX matching engine throughput.} The Lux DEX matching
engine lives at \texttt{$\sim$/work/lx/dex} (accessed in this tree
as the symlink \texttt{$\sim$/work/lux/dex}). It has no GPU surface
today; reporting "DEX on GPU" would require either porting the
match path to a compute kernel or measuring the existing CPU path
under realistic order-book load. Neither is in scope of this paper.

\item \textbf{End-to-end consensus + EVM + DEX + FHE pipeline.}
The Lux components ship as separate Go modules with no integrated
bench harness. Construction of one is a v0.3.0 deliverable (it
requires harmonizing goroutine schedules, shared bus, and KAT
gating across four runtimes that today don't share a process).

\item \textbf{precompile crypto package perf.} The 30+ packages in
\texttt{$\sim$/work/lux/precompile} (BLS / curves / hashes / VRF /
ZK / pairing-based) have their own scattered benchmark files. We
did not aggregate or measure them for this paper. The
\texttt{luxfi/math} substrate currently covers params + backend +
codec + modarith + NTT + poly + RNS + sample only — not curves,
not hashes, not pairings.

\item \textbf{WGSL on non-Apple silicon.} Our WGSL number is
wgpu-native dispatching to Metal under the hood on Darwin arm64.
On a Linux x86 host with Vulkan, or a D3D12 host on Windows, the
WGSL crossover may move. We did not measure those.
\end{itemize}

\subsection{Honest "one and only one way" status}

The architectural commitment from LP-107 is "one and only one way to
do everything." The math substrate now has that property for its 8
packages (params / backend / codec / modarith / NTT / poly / RNS /
sample) — every consumer (\texttt{luxfi/pulsar/wire},
\texttt{luxfi/lens/wire}, \texttt{luxfi/fhe/wire},
\texttt{luxfi/lattice/wire}) consumes the same canonical body, and
$991$ cross-runtime KATs gate any drift. The \texttt{luxfi/math}
\texttt{ntt/canonical/} body owns the Lattigo-derived Montgomery NTT;
\texttt{luxfi/lattice/ring/modular\_reduction.go} is a thin shim
over it.

The architectural commitment is NOT yet met for:
\begin{itemize}[nosep,leftmargin=2em]
\item \textbf{Curves.} Each precompile curve package
(BLS12-381, secp256k1, Ed25519, ...) has its own implementation. A
\texttt{luxfi/math/curve} substrate to consolidate them is queued.
\item \textbf{Hashes.} Each precompile hash (BLAKE3, Keccak, Poseidon)
has its own implementation. A \texttt{luxfi/math/hash} substrate is queued.
\item \textbf{ML-DSA / ML-KEM / SLH-DSA.} The pqcrypto packages live
in \texttt{$\sim$/work/lux/precompile/} with their own implementations.
Substrate consolidation queued.
\end{itemize}

These gaps are real. v0.1.x ships the substrate for the lattice/PQ
threshold path because that's the production hot path Lux currently
runs on; the curve / hash / pqcrypto consolidation is a v0.2.0+
deliverable.

\subsection{Reproduce}

\begin{lstlisting}[basicstyle=\small\ttfamily,frame=single,
                   xleftmargin=0pt, framesep=4pt]
# 1. NTT throughput sweep
cd ~/work/luxcpp/crypto
brew install wgpu-native    # macOS -- provides libwgpu_native
cmake -B build -DCRYPTO_BUILD_TESTS=ON \
      -DCRYPTO_ENABLE_METAL=ON -DCRYPTO_ENABLE_WGSL=ON
cmake --build build --target corona_lattice_ring_sweep_bench
cd build && \
  CRYPTO_CORONA_LATTICE_RING_METALLIB=$PWD/lattice_ring.metallib \
  ./corona_lattice_ring_sweep_bench \
  ../corona/test/kat/montgomery_ntt.json

# 2. ZAP throughput
cd ~/work/lux/consensus
GOWORK=off go test -bench=BenchmarkZAPThroughput -benchtime=5s \
  -run='^$' ./bench/

# 3. FHE bootstrap
cd ~/work/lux/fhe
GOWORK=off go test -bench=BenchmarkBootstrapSingle -benchtime=5s \
  -run='^$' ./

# 4. Pulsar reshare
cd ~/work/lux/threshold
GOWORK=off go test -bench='BenchmarkPulsarReshare' -benchtime=1x \
  -timeout=60m -run='^$' ./protocols/lss/...
\end{lstlisting}

Hardware: Apple M1 Max, macOS 14.5, clang 17, Go 1.26.1,
wgpu-native 29.0.0, CMake 4.3.2. The exact M1 Max numbers in
this paper come from a single warm-CPU run on 2026-05-04 evening
post-fuzz-battery; reruns on the same hardware land within $\pm 5\%$
on the per-poly measurements.