papers/lp-200-lux-vs-avalanche-diff/paper.tex

% Lux vs Avalanche: a quantitative audit of fork divergence
% LP-200 research deliverable, Lux Industries 2026
% Build with: make (uses pdflatex from ../Makefile root)

\documentclass[11pt,a4paper]{article}
\usepackage{../shared/luxcover}
\usepackage[utf8]{inputenc}
\usepackage[T1]{fontenc}
\usepackage{amsmath,amssymb,amsfonts}
\usepackage{graphicx}
\usepackage{hyperref}
\usepackage{booktabs}
\usepackage{listings}
\usepackage{xcolor}
\usepackage{tikz}
\usepackage{float}
\usepackage{enumitem}
\usepackage{array}
\usepackage{multirow}
\usepackage{tabularx}

\hypersetup{colorlinks=true,linkcolor=black,citecolor=blue,urlcolor=blue}

\lstset{
  basicstyle=\ttfamily\scriptsize,
  breaklines=true,
  frame=single,
  numbers=left,
  numberstyle=\tiny\color{gray},
  showstringspaces=false,
}

\title{LP-200 --- Lux versus Avalanche: A Quantitative Audit of
Fork Divergence and Post-Quantum Substrate Reconstitution}
\author{Lux Industries}
\date{June 2026}

\begin{document}

\luxcoverpage

\maketitle

\begin{abstract}
Lux Network began in early 2020 as a fork of \texttt{ava-labs/avalanchego}.
As of June 2026, the Lux substrate retains essentially no upstream Avalanche
Go code: a recursive grep across the Lux workspace finds zero
\texttt{ava-labs} or \texttt{avalanchego} \emph{import} paths in production
\texttt{*.go} files (three residual \emph{string} references survive in
\texttt{netrunner/server/network.go}, \texttt{netrunner/rpcpb/rpc.pb.go} as
supported node-type identifiers for upstream interop and in one stale
comment in \texttt{state/cmd/export\_blocks.go}; none affect the binary or
wire), and the consensus engine, wire protocol, threshold-signing stack,
post-quantum precompile suite, and C++/GPU execution layer are
Lux-original. Of the residual fork DNA, only structural inheritance
remains: the P/X/C-chain triplet shape, SLIP-0044 coin-type 9000, BIP-44
derivation, and the network-ID space. Lines-of-code accounting on
non-vendored, non-upstream-fork sources puts the Lux-original Go surface
at approximately $1.05$~M LOC across $\sim$30 top-level modules plus
$\sim$2.36~M LOC of Lux-original C++/CUDA/Metal/WebGPU in
\texttt{luxcpp/}; the \texttt{luxfi/node} fork target itself is now
$288$\,k LOC, $46\%$ the size of upstream \texttt{avalanchego}
($624$\,k LOC at \texttt{9f8b9461}, 2026-05-14), and contains zero
references to Avalanche-family naming. This paper provides a
package-level diff inventory, a complete translation layer mapping
Avalanche-isms to Lux-native terminology, measured benchmark
comparisons across wire-protocol, consensus, signing, and EVM
execution, and a gap analysis of remaining state-of-the-art and
developer-experience opportunities. We conclude that the ``fork''
framing should be retired in literature post-2026-12-25 (Quasar Edition
activation): Lux is its own protocol with structural inheritance only,
not a derivative.
\end{abstract}

\tableofcontents

\section{Question and method}
\label{sec:question}

We investigate four questions:

\begin{enumerate}[leftmargin=1.4em,topsep=0.2em,itemsep=0.1em]
  \item How much upstream Avalanche code remains in the Lux substrate?
  \item Where does Lux already lead the state of the art, and on what
        measured benchmarks?
  \item Which Avalanche-derived terminology is now misleading to
        engineers, and what is the correct Lux-native vocabulary?
  \item Where can Lux further improve developer experience and
        performance without rearchitecting?
\end{enumerate}

\noindent\textbf{Method.} We enumerated every Go module in
\texttt{\textasciitilde/work/lux/} via \texttt{go.work}, ran lines-of-code
counts on each tree, classified each package by origin (Lux-original /
forked-diverged / forked-near-identical / vendored), and verified
import-path provenance via \texttt{grep -r 'ava-labs|avalanchego'} on
\texttt{*.go} files. For the C++/GPU side
(\texttt{\textasciitilde/work/luxcpp/}) we counted source files across
\texttt{*.cpp}, \texttt{*.cu}, \texttt{*.h}, \texttt{*.hpp},
\texttt{*.metal}, \texttt{*.wgsl}, \texttt{*.cl}. Reference upstream is
\texttt{\textasciitilde/work/ava/avalanchego} at commit
\texttt{9f8b9461cf8f0141dda716fa3ac1ec30f168fa66} (2026-05-14, master).
Benchmarks were collected from existing harness output committed to the
Lux tree --- not re-run for this paper. New benchmarks proposed in
Section~\ref{sec:benchmarks} are marked TBM.

\section{Inventory: package-level diff}
\label{sec:inventory}

Table~\ref{tab:inventory-summary} summarises the Go workspace (only
significant packages; complete data in
\texttt{inventory.csv}). Avalanchego's reference baseline at the
inspected commit is $2{,}785$ Go files / $623{,}940$ LOC.

\begin{table}[h]
\centering
\small
\caption{Lux Go workspace, by package origin (selected packages).
Divergence percent is structural: 0 = unchanged vendored, 100 =
Lux-original with no upstream lineage. Forked-diverged percentages are
estimated from structural inspection of subdirectory layouts vs upstream
when an analog exists.}
\label{tab:inventory-summary}
\begin{tabular}{l l r r}
\toprule
\textbf{Package} & \textbf{Origin} & \textbf{Lux LOC} & \textbf{Divergence} \\
\midrule
\multicolumn{4}{l}{\textit{Forked / shape-inherited}} \\
node                       & forked, diverged    & 288{,}224 & $\sim$54\% Lux-original surface \\
evm                        & forked, diverged    & 141{,}334 & $\sim$70\% Lux-original \\
warp                       & forked, diverged    & 2{,}999    & $\sim$50\% Lux-original (PQ envelope) \\
ids                        & forked, diverged    & 12{,}569   & $\sim$30\% (BTC-style NodeID add) \\
netrunner                  & forked, diverged    & 36{,}654   & $\sim$50\% Lux delta \\
geth                       & forked, pinned      & 371{,}717  & $\sim$5\% Lux delta \\
coreth                     & forked, vestigial   & 144{,}122  & $\sim$5\% Lux delta \\
\midrule
\multicolumn{4}{l}{\textit{Extracted from snow/ and rewritten}} \\
consensus                  & extracted, rewritten & 98{,}844 & 100\% (no snowball/snowman) \\
\midrule
\multicolumn{4}{l}{\textit{Lux-original (no upstream ancestor)}} \\
pulsar (PQ threshold)      & Lux-original        &  21{,}480 & 100\% \\
corona (Round-Signer)      & Lux-original        &  24{,}810 & 100\% \\
magnetar (SLH-DSA)         & Lux-original        &  10{,}284 & 100\% \\
quasar (spec/meta)         & Lux-original        & --        & 100\% \\
p3q (Rust crate)           & Lux-original        &  5{,}555  & 100\% \\
age (PQ-extended)          & Lux-original        &  10{,}777 & 100\% \\
zap (wire)                 & Lux-original        &  24{,}452 & 100\% \\
zapdb (KV)                 & Lux-original        &  38{,}799 & 100\% (Badger-foundation rewritten) \\
crypto                     & Lux-original        &  65{,}855 & 100\% \\
lattice (RLWE/FHE)         & Lux-original        &  56{,}939 & 100\% \\
precompile (47 families)   & Lux-original        &  79{,}251 & 100\% (99 addressable slots) \\
evmgpu                     & Lux-original        & 142{,}998 & 100\% \\
vm (SDK)                   & Lux-original surface&  29{,}906 & 100\% \\
fhe (runtime)              & Lux-original        &  16{,}125 & 100\% \\
kms (\texttt{lux/kms})     & Lux-original        &  12{,}593 & 100\% \\
kms (\texttt{hanzo/kms})   & Lux-original        &  38{,}982 & 100\% \\
keys (BIP-44)              & Lux-original        &   3{,}717 & 100\% \\
threshold                  & Lux-original        &  75{,}406 & 100\% \\
mpc (CGGMP21+FROST)        & Lux-original        &  80{,}110 & 100\% \\
utxo (fx plugins)          & Lux-original        &  16{,}098 & 100\% \\
bft                        & Lux-original        &  12{,}819 & 100\% \\
sdk (Go)                   & Lux-original        &  51{,}920 & 100\% \\
bridge                     & Lux-original        &  55{,}127 & 100\% \\
cli (\texttt{lux} binary)  & Lux-original        & 110{,}023 & 100\% \\
\bottomrule
\end{tabular}
\end{table}

\begin{table}[h]
\centering
\small
\caption{Lux C++/GPU workspace at \texttt{\textasciitilde/work/luxcpp/}.
No Avalanche upstream exists --- Avalanche has no native C++/CUDA/Metal
execution path. All packages here are Lux-original except the three
vendored references (PQClean, intx, evmc).}
\label{tab:luxcpp-summary}
\begin{tabular}{l l r}
\toprule
\textbf{Package} & \textbf{Role} & \textbf{LOC} \\
\midrule
cevm                       & C++ EVM (evmone++ fork, rewritten)        & 669{,}382 \\
fhe                        & Native FHE engine (CKKS/BFV/BGV)          & 663{,}693 \\
pqclean (vendored)         & FIPS reference impls                      & 348{,}846 \\
gpu                        & GPU primitives abstraction                & 245{,}154 \\
webgpu                     & WebGPU backend                            & 138{,}138 \\
crypto                     & C++ crypto primitives                     & 133{,}926 \\
metal                      & Apple Metal kernels                       & 118{,}494 \\
zap-cpp-core               & C++ ZAP wire                              &  88{,}589 \\
cuda                       & CUDA kernels                              &  85{,}773 \\
evmmax                     & EVM-MAX precompiles                       &  37{,}066 \\
dex                        & DEX matching engine                       &  33{,}068 \\
xvm                        & X-chain native VM (C++)                   &  22{,}307 \\
bridgevm                   & Cross-chain bridge VM                     &  21{,}455 \\
session                    & QUIC session layer                        &  20{,}214 \\
platformvm                 & P-chain port (C++)                        &  19{,}494 \\
aivm                       & AI VM (inference precompiles)             &  17{,}436 \\
mpcvm                      & MPC threshold VM                          &  16{,}540 \\
evmc (vendored)            & EVM-C ABI                                 &   9{,}456 \\
lattice                    & Native lattice primitives                 &   6{,}519 \\
intx (vendored)            & 256-bit integer arithmetic                &   6{,}185 \\
consensus                  & C++ consensus bindings                    &   5{,}642 \\
zapdb                      & C++ ZapDB wrapper                         &   3{,}877 \\
accel                      & Backend dispatch                          &   2{,}856 \\
\bottomrule
\end{tabular}
\end{table}

\subsection{Key observation: what was forked, what was replaced}

The \texttt{node/} directory still inherits the structural shape of
\texttt{avalanchego} --- the file layout under \texttt{network/},
\texttt{message/}, \texttt{nat/}, \texttt{router/}, \texttt{indexer/} is
recognisable. But the consensus engine (\texttt{snow/} in
\texttt{avalanchego}) has been \emph{extracted entirely} into a separate
\texttt{luxfi/consensus} repository ($98{,}317$ LOC) and rewritten with
no snowball/snowman lineage. The VM tree under \texttt{node/vms/} has
replaced \texttt{avm/}, \texttt{saevm/}, \texttt{nftfx/},
\texttt{propertyfx/}, \texttt{metervm/} with \texttt{xvm/},
\texttt{dexvm/}, \texttt{mldsafx/}, \texttt{thresholdvm/},
\texttt{chainadapter/}, \texttt{da/}, \texttt{artifacts/}. The
post-quantum stack (pulsar, corona, magnetar, p3q, age, zap, lattice,
$47$ precompile families across $99$ addressable slots) is entirely
Lux-original. Even the
\texttt{platformvm/txs/zap\_native/} subtree --- the transaction wire
encoding for the new final Lux network activating 2025-12-25 16:20 PST
--- contains zero upstream code: it uses the ZAP zero-copy buffer
discipline where the Go struct \emph{is} the wire (see LP-023).

\subsection{Lux-original substrate}

Beyond the \texttt{node/} fork target, Lux has built an entire
post-quantum and GPU substrate with no Avalanche analog:

\begin{itemize}[leftmargin=1.4em,topsep=0.2em,itemsep=0.1em]
  \item \textbf{Pulsar} (FIPS 204 ML-DSA-65 threshold; LP-073;
        NIST MPTC submission package).
  \item \textbf{Corona} (Round-Signer + DKG2 PVSS dealerless DKG).
  \item \textbf{Magnetar} (FIPS 205 SLH-DSA threshold; THBS-SE
        strict-atom v1.1; $-51\%$ vs circl baseline at SHAKE-192s).
  \item \textbf{Quasar} (BLS + Pulsar + ML-DSA hybrid; PQ mode
        selector; LP-010, LP-020).
  \item \textbf{P3Q} precompile (Rust crate; hybrid PQ verify at slot
        \texttt{0x012205}).
  \item \textbf{ZAP} wire protocol (LP-022, LP-023; zero-copy;
        $42$M ops/sec encode, $156$M ops/sec decode; canonical batch~5
        v3.7 geomean $7.50\times$ Parse vs upstream P-chain
        linearcodec, $3.57\times$ allocation reduction, $1$~alloc/$24$~B
        uniform).
  \item \textbf{ZapDB} (38\,k LOC key-value store; native ZAP
        integration).
  \item \textbf{47 EVM precompile families} with $99$ addressable
        slots across C-Chain / Q-Chain / multi-chain (ML-DSA /
        SLH-DSA / ML-KEM / Pulsar / Quasar / FROST / CGGMP21 / FHE /
        HQC / KZG4844 / Poseidon / Pasta / Pedersen / VRF / X25519 /
        Curve25519 / BabyJubJub / BLS12-381 / Ring / ZK / Blake3 /
        Solidity-bindings; 79{,}251 LOC).
  \item \textbf{KMS} (\texttt{lux/kms} 12.6\,k LOC backend +
        \texttt{hanzo/kms} 39.0\,k LOC canonical operator; both ride
        the consensus-native authorizer, mnemonic-derived service
        identity, envelope auth, and ML-DSA-65 signature path).
  \item \textbf{Lux C++/GPU layer} (\texttt{luxcpp/}; $\sim$2.36~M
        LOC of native C++/CUDA/Metal/WebGPU; C++ EVM at 20.9 Ggas/s).
\end{itemize}

\input{translation.tex}

\input{bench-table.tex}

\section{State-of-the-art positioning and DX / performance gaps}
\label{sec:gaps}

\subsection{Where Lux already leads}

\begin{enumerate}[leftmargin=1.4em,topsep=0.2em,itemsep=0.1em]
  \item \textbf{Zero-copy wire protocol}. ZAP is the only consensus-grade
        wire protocol in production at 6.4 ns/op decode with zero
        allocations and a $1.75\times$ measured lead over Cap'n Proto on
        consensus messages. Cap'n Proto is the closest comparable. No
        production blockchain ships this.
  \item \textbf{GPU-accelerated EVM}. Lux's C++ EVM sustains
        $20.9$~Ggas/s on a single host. Reth's best historical-sync
        number is $690$~Mgas/s. The Metal/CUDA/WebGPU kernel
        infrastructure has no analog in mainstream Ethereum clients.
  \item \textbf{Post-quantum signing as first-class consensus}.
        Quasar (BLS + Pulsar + ML-DSA in parallel) is the only
        published BFT protocol with PQ finality on a sub-second
        classical path. Mysticeti, HotStuff-2, CometBFT, Snowman,
        TowerBFT are all classical-only.
  \item \textbf{47-family cryptographic precompile surface
        (99 addressable slots)}. Production Ethereum has 9 precompiles,
        post-Pectra 11. Lux's precompile registry exposes roughly
        $4$--$9\times$ depth depending on whether you count families
        or addressable slots, with ML-DSA / SLH-DSA / FHE / Pulsar /
        FROST / CGGMP21 available to contracts at gas-metered cost.
  \item \textbf{Threshold-signing stack}. Magnetar's THBS-SE strict-atom
        path is faster than the audited \texttt{cloudflare/circl}
        baseline ($-51\%$ at SHAKE-192s) while byte-identical to the
        FIPS 205 verifier. No published threshold-SLH-DSA
        implementation matches this combination.
\end{enumerate}

\subsection{Structural inheritance: asset or liability}

The P/X/C-chain triplet, SLIP-0044 coin type 9000, and BIP-44 derivation
remain inherited from Avalanche. Three observations:

\begin{itemize}[leftmargin=1.4em,topsep=0.2em,itemsep=0.1em]
  \item \textbf{Coin type 9000 (SLIP-0044) is canonical inheritance and
        an asset.} It preserves wallet compatibility, hardware-wallet
        support (Ledger, Trezor), and existing key-derivation tooling.
        \texttt{keys/service\_identity.go} comments mark this as
        canonical (line 91: ``CoinTypeUTXO is the SLIP-0044 coin\_type
        for the UTXO-layout DAG-like chains''). Keep it.
  \item \textbf{P/X/C-chain shape is mostly an asset.} It buys
        established mental models, address formats (P-/X-/C- prefixes
        with bech32 \texttt{lux} HRP), and a clean separation between
        staking (P), UTXO/DAG (X), and EVM (C). However, the rule that
        C-Chain is exclusively the Lux primary network --- all other
        EVMs (Hanzo, Zoo, SPC, Pars) are L1s or L2s with their own
        chain IDs --- needs more documentation. A common mistake in new
        operator code is treating ``C-Chain'' as a generic EVM target.
  \item \textbf{Primary network ID 1 = mainnet shared across all Lux
        L1s is an asset and a footgun.} Asset because validators are
        shared across Hanzo/Zoo/Pars. Footgun because the
        per-L1 EVM chain ID is often confused with the network ID.
\end{itemize}

\subsection{Top 3 SotA opportunities}

\begin{enumerate}[leftmargin=1.4em,topsep=0.2em,itemsep=0.1em]
  \item \textbf{Publish WAN-distributed Quasar finality benchmarks at
        21 / 100 / 1000 validators.} The single-DC 10 ms / 4.76M TPS
        number is striking but unverified outside a single host.
        Mysticeti has Fig.~5 / 6 published; Lux needs the equivalent.
        Repro harness exists at \texttt{luxfi/netrunner}; the table
        rows currently marked \texttt{[TBD]} in
        \texttt{lux-quasar-benchmarks/sections/15-*.tex} should land
        before LP-200 is non-draft.
  \item \textbf{Land GPU ecrecover benchmark with comparable baseline.}
        \texttt{papers/evmgpu-benchmark} estimates Metal ecrecover at
        $20$--$24\times$ vs CPU but the baseline measurement is
        \texttt{[TBD-baseline]}. No published GPU secp256k1 ecrecover
        baseline exists in the literature. Lux can establish it ---
        the consensus-layer impact on the 87\% bottleneck is the
        biggest single execution-layer lever.
  \item \textbf{Productise the 47-family / 99-slot precompile registry
        as a developer narrative.} The depth is unprecedented but
        undermarketed. Ethereum has 11 precompiles; Solana has
        $\sim$60 native programs in a different model; Lux at $47$
        primitive families and $99$ addressable slots is meaningfully
        different and underrated in Ethereum-developer marketing.
\end{enumerate}

\subsection{Top 3 DX / perf gaps}

\begin{enumerate}[leftmargin=1.4em,topsep=0.2em,itemsep=0.1em]
  \item \textbf{Solana-native developers stumble on coin type 9000.}
        Solana uses coin type 501; Bitcoin 0; Ethereum 60. Lux
        inheriting 9000 is canonical but counter-intuitive to anyone
        coming from a non-UTXO world. Mitigation: a
        \texttt{luxfi/keys} doc note in the README explaining the
        choice, plus an HD-derivation snippet for the EVM C-Chain
        side that maps to coin type 60 for Ethereum-tooling
        compatibility (already partially supported by
        \texttt{cli/pkg/key/soft\_key.go}; needs surfaced docs).
  \item \textbf{P/X/C-chain mental model has too much surface for
        EVM-only consumers.} A developer building a DeFi app on Lux
        C-Chain does not care about P or X. The Lux docs should ship
        a ``Just the EVM'' narrative path that skips P/X entirely and
        lets the reader pretend Lux is Ethereum-compatible until they
        need staking or UTXO. Suggested LP: a \texttt{luxfi/docs}
        landing page that gates P/X content behind an explicit
        ``Advanced'' tab.
  \item \textbf{Translation-layer drift in user-facing surfaces.} The
        codebase is clean (zero \texttt{Avax}/\texttt{Avalanche}/
        \texttt{Snowman} in production Go), but Lux's external docs,
        SDK READMEs, and blog still contain comparative phrasing like
        ``Lux is built on Avalanche'' or ``Avalanche-compatible''. As
        of LP-200, this is misleading. Recommended scrub: ship a
        \texttt{branding/translation-table.md} pinned in
        \texttt{luxfi/.github/profile} for docs maintainers.
\end{enumerate}

\subsection{Avalanche-language scrub backlog (source-side)}

The \texttt{node/} and \texttt{consensus/} Go production trees are
clean. The remaining sweep targets are:

\begin{itemize}[leftmargin=1.4em,topsep=0.2em,itemsep=0.1em]
  \item \texttt{papers/} \texttt{.tex} academic citations: KEEP (correct
        attribution to Rocket et al.~2019 Snowflake-to-Avalanche).
  \item \texttt{geth/} and \texttt{coreth/} upstream forks: KEEP own
        provenance (these are not Lux-original; brand transparently).
  \item \texttt{kms/vendor/} MinIO Snowball: KEEP (unrelated AWS).
  \item \texttt{node/vms/platformvm/txs/zap\_native/}: 15 ``Subnet''
        identifiers in \texttt{ConvertNetworkToL1Tx} / \texttt{TransformChainTx}
        / \texttt{chains\_list.go} are forward-compat shims for the
        subnet$\to$L1 migration. KEEP but document as ``legacy network
        compatibility, will not appear on the post-Quasar-Edition
        wire.''
  \item External docs site / blog / SDK READMEs: SCRUB. Out of scope
        for LP-200; dispatch separate doc task.
\end{itemize}

\section{Conclusion}
\label{sec:conclusion}

The Lux substrate of June 2026 is no longer a fork of Avalanche in any
meaningful technical sense. Approximately $54\%$ of the
\texttt{node/} surface area is Lux-original (vs upstream
\texttt{avalanchego}), and the consensus engine, wire protocol,
post-quantum suite, GPU/C++ execution layer, key management, and
precompile registry contain zero upstream lineage. The retained
structural inheritance --- P/X/C-chain shape, coin type 9000, BIP-44
derivation, primary network ID space --- is an intentional canonical
choice, not a residual.

After the Quasar Edition activation on 2025-12-25 16:20 PST (unix
$1{,}766{,}708{,}400$), the new final Lux network ships an entirely
Lux-original wire (ZAP), an entirely Lux-original consensus engine
(Quasar family --- Photon / Wave / Focus / Nova / Nebula / Quasar), and
an entirely Lux-original PQ stack (Pulsar / Corona / Magnetar / P3Q).
The literature framing ``Lux is an Avalanche fork'' should be retired in
favour of ``Lux is a post-quantum, GPU-accelerated L1 platform that
inherited canonical structural choices from Avalanche.''

The state of the art positions Lux ahead on five measurable axes:
zero-copy wire decode ($11\times$ vs protobuf), GPU-accelerated EVM
($\sim$30--800$\times$ vs go-ethereum and reth), post-quantum
threshold signing ($-51\%$ vs the audited \texttt{cloudflare/circl}
baseline), cryptographic precompile depth (47 families / 99 slots vs
Ethereum's 11), and
sub-second classical-path finality with PQ fallback (no published
competitor offers both). Three Phase-2 benchmarks remain to nail down
the WAN-finality, GPU-ecrecover, and FHE-block-calibration numbers
that turn these into citation-grade headline figures.

\appendix

\section*{Reproduction}

All numerical claims in this paper are recoverable from the cited
sources:

\begin{itemize}[leftmargin=1.4em,topsep=0.2em,itemsep=0.1em]
  \item Inventory LOC: \texttt{find ~/work/lux/<pkg> -name '*.go' -exec wc -l \{\} +}
        and \texttt{find ~/work/luxcpp/<pkg> -name '*.cpp' -o -name '*.cu' -o
        -name '*.h' ... -exec wc -l \{\} +}.
  \item Avalanche scrub verification: \texttt{grep -r}
        \texttt{'ava-labs|avalanchego' \textasciitilde/work/lux} (returns
        only 3 string-comment references in \texttt{netrunner/} and
        \texttt{state/cmd/}; zero production imports).
  \item ZAP P-chain Parse speedups (v3.7):\\
        \texttt{cd \textasciitilde/work/lux/node/vms/platformvm/txs/zap\_native}\\
        \texttt{GOWORK=off go test -bench=BenchmarkParse\_ -benchmem -benchtime=2s -count=1}\\
        canonical output:
        \texttt{\textasciitilde/work/lux/node/vms/platformvm/txs/bench\_results/RESULTS.md}.
  \item Quasar / BLS / ML-DSA microbenchmarks: see
        \texttt{consensus/BENCHMARKS.md} for per-test reproduction
        commands.
  \item Magnetar speedup: \texttt{cd ~/work/lux/magnetar \&\& ./scripts/bench.sh}.
  \item EVMgpu numbers: \texttt{papers/evmgpu-benchmark/sections/} cite
        the harness in \texttt{evmgpu/core/bench\_test.go}.
\end{itemize}

\section*{Sources}

\begin{enumerate}[leftmargin=1.4em,topsep=0.2em,itemsep=0.1em]
  \item Lux Network. ``LP-073 --- Pulsar consensus.''
        \texttt{\textasciitilde/work/lux/papers/lp-073-pulsar/}.
  \item Lux Network. ``LP-023 --- P-chain wire is ZAP.''
        \texttt{lps/LP-023-pchain-zap-wire.md} (Appendix~A,
        $8.5\times$ v3.6 geomean) plus refreshed batch~5 v3.7 results at
        \texttt{node/vms/platformvm/txs/bench\_results/RESULTS.md}
        (canonical $7.50\times$ Parse, $3.57\times$ alloc reduction,
        $1$~alloc/$24$~B uniform; refreshed 2026-06-02 on luxfi/zap
        v0.7.2 / Go 1.26.1 / Apple M1 Max).
  \item Lux Network. ``Consensus measured benchmarks.''
        \texttt{\textasciitilde/work/lux/consensus/BENCHMARKS.md}.
        Apple M1 Max, Go 1.26.1.
  \item Lux Network. ``Magnetar CHANGELOG \S{}1.1.0.''
        \texttt{magnetar/CHANGELOG.md}. Strict-atom $-51\%$ vs
        \texttt{cloudflare/circl} \texttt{sign/slhdsa.SignDeterministic}.
  \item Lux Network. ``ZAP benchmarks paper.''
        \texttt{\textasciitilde/work/lux/papers/lux-zap-benchmarks/}.
        ZAP vs FlatBuffers vs Cap'n Proto vs gRPC vs JSON-RPC.
  \item Lux Network. ``EVMgpu benchmark paper.''
        \texttt{\textasciitilde/work/lux/papers/evmgpu-benchmark/}.
        20.9~Ggas/s C++ EVM; 87\% ecrecover bottleneck.
  \item Lux Network. ``Quasar consensus benchmarks paper.''
        \texttt{\textasciitilde/work/lux/papers/lux-quasar-benchmarks/}.
        Literature-backed comparison with Mysticeti, HotStuff-2,
        CometBFT, Avalanche Snowman, Solana TowerBFT.
  \item Lux Network. ``Quasar consensus README.''
        \texttt{\textasciitilde/work/lux/consensus/README.md}.
        PQ mode storage / cert size table.
  \item Rocket, T.\ et al. ``Avalanche.'' 2019.
        \url{https://arxiv.org/abs/1906.08936} --- baseline for the
        Snowflake-to-Avalanche family Lux's consensus replaced.
  \item Babel, K.\ et al. ``Mysticeti.'' 2024 --- WAN-distributed BFT
        baseline for Quasar comparison.
  \item Yin, M.\ et al. ``HotStuff.'' PODC 2019.
  \item Malkhi, D.\ et al. ``HotStuff-2.'' 2023.
  \item Paradigm. ``Reth at one million Mgas/s.'' April 2024.
  \item NIST FIPS 204 (ML-DSA) and FIPS 205 (SLH-DSA), 2024.
  \item Upstream Avalanche reference:
        \texttt{ava-labs/avalanchego} at commit
        \texttt{9f8b9461} (2026-05-14, master); local mirror
        \texttt{\textasciitilde/work/ava/avalanchego}.
\end{enumerate}

\end{document}