papers/hanzo-4-0-launch/hanzo-4-0-launch.tex

\documentclass[11pt,letterpaper]{article}
\usepackage{../shared/hanzocover}

\usepackage[utf8]{inputenc}
\usepackage[T1]{fontenc}
\usepackage{amsmath,amssymb,amsfonts,amsthm}
\usepackage{mathtools}
\usepackage{graphicx}
\usepackage{booktabs}
\usepackage{hyperref}
\usepackage{enumitem}
\usepackage{algorithm}
\usepackage{algorithmic}
\usepackage{xcolor}
\usepackage{listings}
\usepackage{tikz}
\usepackage{float}
\usepackage{array}
\usepackage{tabularx}
\usepackage{multirow}
\input{../shared/lstlang}

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

\lstset{
  basicstyle=\ttfamily\scriptsize,
  breaklines=true,
  frame=single,
  numbers=left,
  numberstyle=\tiny\color{gray},
  keywordstyle=\color{blue!70!black}\bfseries,
  commentstyle=\color{green!50!black}\itshape,
  stringstyle=\color{red!60!black},
  showstringspaces=false,
}

\newtheorem{theorem}{Theorem}[section]
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{proposition}[theorem]{Proposition}
\newtheorem{corollary}[theorem]{Corollary}
\newtheorem{definition}{Definition}[section]
\newtheorem{invariant}{Invariant}[section]

\title{Hanzo 4.0 --- GPU-Native AI Compute Network with DEX+EVM+FHE Triumvirate\\[6pt]
\large Native AI-compute DEX (HMM) with Liquidity Protocol\\
Hamiltonian LLMs at L1 \quad GPU-Residency from chain to model}

\author{
Hanzo AI Research\thanks{[email protected]} \and Antje Worring\thanks{[email protected]}\\
\textit{Hanzo Industries, Inc.} \quad Techstars '17\\
\texttt{[email protected]}
}

\date{Spec freeze: February~7, 2026 \quad|\quad Activation: February~14, 2026}

\begin{document}
\hanzocoverpage
\maketitle

\begin{abstract}
On 14 February 2026 the Hanzo network activated its 4.0 epoch.
The network became a sovereign Layer~1 running the canonical
Lux-family triumvirate (DEX + EVM + FHE) on a GPU-native Quasar
4.0 stack, with Hamiltonian LLMs as a first-class L1 service
and Hanzo Market Maker (HMM) as the native DEX for AI compute.
This paper documents the launch as a retrospective: the four-version
chronology (1.0 initial AI compute network, 2.0 Hamiltonian LLMs and
the AI Coin, 3.0 full post-quantum, 4.0 GPU-native), the AI Chain
Virtual Machine (AIVM) opcode surface, the HMM cross-listing with
Lux D-Chain and Zoo DEX through the Liquidity Protocol common API,
the GPU-Residency invariant from chain to model, and the on-mainnet
receipts captured at activation block height $h_0 = 36{,}963{,}000$.
\end{abstract}

\tableofcontents
\newpage

% ============================================================================
\section{Chronology}
\label{sec:chronology}
% ============================================================================

Hanzo, founded by the Techstars '17 batch, is now four major versions
deep. Each successor strictly subsumes its predecessor; nothing in the
1.0 line was deprecated.

\begin{table}[H]
\centering
\small
\begin{tabularx}{\textwidth}{lllX}
\toprule
\textbf{Version} & \textbf{Active from} & \textbf{Theme} & \textbf{Defining changes} \\
\midrule
1.0 & 2018--2023 & Initial AI compute network & Centralised inference and
training infrastructure with off-chain billing. \\
2.0 & 2024 & Hamiltonian LLMs + AI Coin & On-chain accounting for inference, the AI Coin native asset, the first version of the HMM DEX. \\
3.0 & 2025 & Full PQ & ML-DSA mandatory for validators; Hanzo settles
under Lux Quasar 3.0 (LP-105) cert lanes; FHE inference research preview. \\
4.0 & 2026-02-14 & GPU-native + native DEX with Liquidity Protocol &
Sovereign L1 with the canonical 9-chain template (LP-134), AIVM
GPU-resident, HMM cross-listed via Liquidity Protocol, FHE inference promoted to production
($\texttt{hanzo-fhe-inference}$). \\
\bottomrule
\end{tabularx}
\caption{Hanzo major-version timeline. The 4.0 boundary co-activates
with the Lux 4.0 boundary so that the family hard-fork is atomic.}
\end{table}

% ============================================================================
\section{The Triumvirate on Hanzo}
\label{sec:triumvirate}
% ============================================================================

The \emph{triumvirate} is the three-pillar set of DEX + EVM + FHE that
defines a complete financial-rail-resident L1. Hanzo 4.0 is the
canonical \emph{triumvirate network}: its chain set \emph{is} exactly
those three subsystems --- DEX (HMM), EVM (Cancun-equivalent C-Chain
template), and FHE (confidential AI inference) --- and nothing else.
Hanzo's EVM is plain EVM; AIVM is the dedicated A-Chain VM at the
Lux primary network (LP-134, attestation + AI provenance), which
Hanzo accesses \emph{natively} (\S\ref{sec:lux-native-access}), not
as a Hanzo-local clone.

The other Lux-family networks carry the financial-rail core too, but
as a subset of broader chain sets: Lux 4.0 runs the full 9-chain
topology (LP-134); Zoo 4.0 adds per-LLM chains, conservation chains,
and L1$\to$L4 tenancy; Pars 2.0 adds P/X/S-Chain plus
Persian-diaspora precompiles. Hanzo is the focused triumvirate where
every chain is the financial rail; everything non-financial is
serviced by Lux primary natively.

Hanzo 4.0 also serves as the \textbf{Rust canonical node} of the family:
\texttt{\textasciitilde/work/hanzo/cli} (Rust) and the
\texttt{hanzo-libs} workspace under \texttt{\textasciitilde/work/hanzo/node}
are the reference Rust implementations the other networks integrate
against.

\subsection{Native Lux primary-chain access}
\label{sec:lux-native-access}

Hanzo, like every Lux-family L1, natively uses the Lux primary
network's chains --- P, C, X, Q, Z, A, B, M, F --- without bridging.
"Native" means same Quasar cert lanes, same subject-binding, same
finality, no wrapped assets, no cross-chain hop tax. Hanzo
specifically reaches into Lux primary for:

\begin{itemize}[leftmargin=*]
\item \textbf{A-Chain (AIVM)} for attestation, model-registry
      anchors, and AI-provenance receipts. The 8 AIVM precompiles
      (\texttt{0x0a01}--\texttt{0x0a08}) live on \emph{Lux} A-Chain;
      Hanzo's HMM marketplace records AI-compute attestations there
      directly.
\item \textbf{Z-Chain (ZVM)} for ZK rollups when Hanzo customers need
      proof-aggregated batches.
\item \textbf{M-Chain (MVM)} for any MPC ceremony Hanzo's HMM
      requires (CGGMP21, FROST, Corona-general).
\item \textbf{B-Chain (BVM)} for canonical cross-ecosystem bridging.
\item \textbf{F-Chain (FVM)} when Hanzo's local FHE service needs to
      delegate to Lux primary's FHE substrate.
\end{itemize}

Hanzo does not duplicate any of those chains locally. The boundary
between Hanzo's three chains and Lux's primary nine is cert-lane
subject-binding, not a bridge.

\subsection{DEX --- Hanzo Market Maker (HMM)}

HMM is the native AI-compute exchange, originally introduced in HIP-8
(2024) and graduated to L1 status at the 4.0 boundary. It quotes,
matches, and settles five compute primitives:

\begin{enumerate}[leftmargin=*]
\item \textbf{Inference seconds.} GPU-second futures by model class
(LLM-7B, LLM-70B, LLM-405B, ViT-B, etc.), priced per quality tier.
\item \textbf{Training slots.} Fixed-window training capacity on
specific GPU classes, with deterministic data-locality routing.
\item \textbf{GPU leases.} Bare-metal capacity by hour, expressed
through the AIVM \texttt{GPU\_LEASE} opcode (\texttt{0x0a05}) on
Lux A-Chain, accessed natively from Hanzo's EVM.
\item \textbf{Embedding throughput.} Tokens/sec for retrieval-augmented
generation pipelines.
\item \textbf{Active-inference budgets.} Hamiltonian LLM compute
budgets (cf.\ HIP-7).
\end{enumerate}

The price function for inference seconds in pool $p$ at time $t$ is the
constant-product invariant adapted for capacity decay:
\begin{equation}
x_t \cdot y_t = k_p \cdot \exp\!\left(-\lambda_p \cdot (t - t_0)\right),
\label{eq:hmm-decay}
\end{equation}
where $x_t$ is unsold compute capacity, $y_t$ is the AI-coin
reserve, $k_p$ is the genesis liquidity, and $\lambda_p$ is the
capacity-decay rate calibrated per pool to express that compute time
not used by $t$ is forever lost. Equation~\ref{eq:hmm-decay} is solved
on-GPU within the EVM execution path with native AIVM hooks; round-trip latency for a swap is
under one millisecond.

\paragraph{Liquidity Protocol cross-listing.} HMM registered with the
Liquidity Protocol common settlement API (LP-310) at $h_0 + 12$. The
router exposes HMM pools to Lux D-Chain and Zoo DEX, allowing a
single transaction to draw inference capacity quoted on Hanzo against
collateral held on Lux or Zoo.

\subsection{EVM --- C-Chain template, GPU-native}

Hanzo's EVM is plain Cancun-equivalent EVM (the C-Chain template
from LP-134), GPU-native via cevm v0.41+. AIVM is the dedicated
A-Chain VM at the Lux primary network and is \emph{not} a Hanzo
local; Hanzo uses Lux A-Chain (AIVM) natively for AI-provenance
opcodes. The eight AIVM precompiles (\texttt{0x0a01}--\texttt{0x0a08})
live on Lux A-Chain. From Hanzo's EVM, calls into AIVM look like any
other native primary-chain access (no bridge, same cert lanes, same
subject-binding):

\begin{table}[H]
\centering
\small
\begin{tabular}{lll}
\toprule
\textbf{Address} & \textbf{Function} & \textbf{Notes} \\
\midrule
\texttt{0x0a01} & \textsc{HMM\_LOOKUP} & Resolve model\_root $\to$ metadata \\
\texttt{0x0a02} & \textsc{INFER\_REQUEST} & Submit inference job to market \\
\texttt{0x0a03} & \textsc{INFER\_VERIFY} & Verify TEE quote + receipt \\
\texttt{0x0a04} & \textsc{TRAIN\_BOUNTY} & Post training bounty \\
\texttt{0x0a05} & \textsc{GPU\_LEASE} & Lease GPU pool capacity \\
\texttt{0x0a06} & \textsc{AGENT\_AUTH} & Authorize agent action under capability \\
\texttt{0x0a07} & \textsc{F\_CALL} & Cross-chain call to F-Chain (FHE) \\
\texttt{0x0a08} & \textsc{M\_CUSTODY} & Cross-chain call to M-Chain (MPC) \\
\bottomrule
\end{tabular}
\caption{AIVM precompile surface (Lux A-Chain, accessed natively by Hanzo). All GPU-resident at the 4.0 boundary.}
\end{table}

GPU-residency is enforced by the same QuasarSTM 4.0 transactional
memory boundaries used on Lux:

\begin{invariant}[GPU residency, LP-137]
For every transaction $\tau$ executed at height $h \ge h_0$, the
canonical execution path runs on a GPU device $g$ with all input,
intermediate, and output state resident on $g$'s HBM for the duration
of $\tau$. Any host-RAM round-trip during $\tau$ is a violation.
\end{invariant}

The AIVM precompiles \texttt{0x0a02} (\textsc{INFER\_REQUEST}) and
\texttt{0x0a03} (\textsc{INFER\_VERIFY}) are direct-call paths into
QuasarGPU's inference kernels --- the same hardware that performs cert
aggregation also serves the model. Co-location halves end-to-end
latency for AI-bound contracts.

\subsection{FHE --- Confidential AI Inference}

Confidential AI inference is documented in
\texttt{hanzo-fhe-inference}. The 4.0 launch promotes that paper from
research preview to production. Three application patterns activate at
$h_0$:

\begin{enumerate}[leftmargin=*]
\item \textbf{Encrypted prompt inference.} Clients submit CKKS-encrypted
prompts; Hanzo validators run the model over ciphertexts; only the
client (holder of the secret key) can decrypt. Hanzo never sees the
plaintext.
\item \textbf{Encrypted decision-tree inference.} TFHE-encrypted feature
vectors traversed against TFHE-encrypted decision boundaries; useful
for medical and financial classification.
\item \textbf{Encrypted federated aggregation.} Replaces the secure-aggregator
trust assumption in $K$-party fine-tuning with homomorphic gradient
summation.
\end{enumerate}

The FHE substrate is reached through the AIVM \texttt{F\_CALL}
precompile (\texttt{0x0a07}). The bridge to Lux F-Chain (LP-013 v2)
is the canonical fallback for clients that prefer Lux-rooted threshold
decryption.

% ============================================================================
\section{\$AI Coin Tokenomics + GPU Mining (PoAI)}
\label{sec:tokenomics}
% ============================================================================

The AI Coin (ticker \texttt{AI}) is the native asset of the Hanzo L1.
Tokenomics are documented in detail in \texttt{hanzo-tokenomics}; the
4.0 boundary preserves all parameters and adds two emission streams.

\begin{itemize}[leftmargin=*]
\item \textbf{HMM fee routing.} $0.30\%$ of HMM swap volume routes to
the canonical Lux-family split: $50\%$ liquidity providers, $30\%$
emission stabiliser, $15\%$ ve-AI lockers, $5\%$ DAO treasury.
\item \textbf{FHE relay rewards.} Capped at $0.5\%$ annualised AI
supply, paid to validators that perform FHE bootstrapping and ciphertext
maintenance. The cap sits inside the previously-allocated ecosystem
bucket; \emph{no new emission}.
\end{itemize}

\paragraph{Proof-of-AI (PoAI) GPU mining.} Hanzo's emission curve is
funded by Proof-of-AI: validators must prove they served verifiable
inference work (TEE quote + receipt, verified via
\texttt{INFER\_VERIFY}) to earn block rewards above the floor.
PoAI is a strict superset of proof-of-stake; validators with no
inference receipts still earn the floor.

\begin{theorem}[Supply preservation]
The 4.0 emission additions preserve the fixed total supply of $10^9$
AI coins. The FHE relay reward cap and HMM fee routing changes are
funded entirely from previously-allocated buckets in the
\texttt{hanzo-tokenomics} schedule.
\end{theorem}
\begin{proof}
Mechanical. The 2.0-era schedule allocates $30\%$ of total supply
($3 \times 10^8$) to ecosystem incentives. The FHE reward cap is
$0.5\%$ of $10^9$ for at most 8 years $= 4 \times 10^7$. The HMM
liquidity-bootstrap vault is $5 \times 10^7$. Both fit within
$3 \times 10^8$.
\end{proof}

% ============================================================================
\section{HMM and ASO Protocols}
\label{sec:hmm-aso}
% ============================================================================

HMM is paired with ASO (Active Sequence Orchestration), the
inference-routing protocol that selects which validator serves each
request. ASO and HMM together form the Hanzo compute marketplace.

\subsection{ASO routing}

For an inference request $r$ with budget $b$ and quality target $q$,
ASO selects the validator $v^*$ minimising the objective
\begin{equation}
v^* = \arg\min_{v \in V} \;
\alpha_1 \cdot \mathrm{lat}_v(r)
+ \alpha_2 \cdot \mathrm{cost}_v(r)
+ \alpha_3 \cdot \mathrm{qualitygap}_v(r, q)
+ \alpha_4 \cdot \mathrm{distance}_v(r),
\label{eq:aso-route}
\end{equation}
subject to $\mathrm{cost}_{v^*}(r) \le b$. The weights
$\alpha_1\dots\alpha_4$ are governance parameters; default values are
fixed at $(0.4, 0.3, 0.2, 0.1)$ at $h_0$.

\subsection{HMM-ASO settlement}

A served inference produces a TEE-signed receipt
$\rho = (r,\, v^*,\, t_{\text{start}},\, t_{\text{end}},\,
\mathrm{cost})$. The HMM contract verifies $\rho$ via the
\texttt{INFER\_VERIFY} precompile and atomically:
\begin{enumerate}[leftmargin=*]
\item Debits the buyer's pool position by $\mathrm{cost}$.
\item Credits the validator's settlement account.
\item Updates the pool's capacity decay (\S~\ref{sec:triumvirate}, eq.~\ref{eq:hmm-decay}).
\item Emits a Liquidity-Protocol event so cross-listed pools on Lux
and Zoo see the price update.
\end{enumerate}

The cycle finishes inside a single block.

% ============================================================================
\section{Bridge to Lux, Zoo, Pars}
\label{sec:bridges}
% ============================================================================

\begin{table}[H]
\centering
\small
\begin{tabular}{lllr}
\toprule
\textbf{Counterparty} & \textbf{Mechanism} & \textbf{Proof} & \textbf{Settlement} \\
\midrule
Lux & A-Chain attestation (\texttt{lux-achain-attestation}) & ML-DSA-87 & 24 s \\
Zoo & Liquidity Protocol router & ML-DSA-87 & 30 s \\
Pars & Liquidity Protocol router & ML-DSA-87 & 30 s \\
Ethereum & Optimistic light client & ECDSA + ML-DSA-87 & 7 d \\
\bottomrule
\end{tabular}
\caption{Hanzo 4.0 cross-chain bridges. All bridge proofs are
post-quantum at the signature layer.}
\end{table}

% ============================================================================
\section{Activation Receipts}
\label{sec:receipts}
% ============================================================================

\begin{table}[H]
\centering
\small
\begin{tabular}{lll}
\toprule
\textbf{Event} & \textbf{Height} & \textbf{Notes} \\
\midrule
Quasar 4.0 epoch boundary & $h_0$ & cert lanes 1--4 \\
QuasarSTM 4.0 activation & $h_0$ & STM root rotated \\
GPU-Residency invariant on & $h_0 + 1$ & violations $= 0$ \\
HMM genesis pools & $h_0 + 6$ & 12 inference, 5 training, 8 GPU-lease \\
Liquidity Protocol register & $h_0 + 12$ & router id \texttt{lp.hanzo.router} \\
First $AI \to LUX$ HMM swap & $h_0 + 47$ & cross-listed via Lux D-Chain \\
First $AI \to ZOO$ HMM swap & $h_0 + 81$ & cross-listed via Zoo DEX \\
First confidential inference & $h_0 + 122$ & CKKS-encrypted prompt \\
F-Chain bridge live & $h_0 + 220$ & via M-Chain custody \\
First active-inference job & $h_0 + 350$ & Hamiltonian LLM, HIP-7 path \\
\bottomrule
\end{tabular}
\caption{Selected mainnet receipts from Hanzo 4.0 activation.}
\end{table}

\subsection{Hardware footprint at $h_0$}

\begin{table}[H]
\centering
\small
\begin{tabular}{lr}
\toprule
\textbf{Indicator} & \textbf{Value at $h_0 + 24h$} \\
\midrule
Active validators & $312$ \\
Aggregate GPU FLOPs (FP16) & $4.2 \cdot 10^{18}$ \\
Inference TPS sustained & $2{,}100{,}000$ \\
Median inference latency (LLM-7B) & $42$ ms \\
Median HMM swap latency & $0.91$ ms \\
\bottomrule
\end{tabular}
\caption{Production hardware footprint observed in the 24h after
$h_0$.}
\end{table}

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

Hanzo 4.0 closes the loop on what 1.0 began in 2018: a verifiable,
liquid market for AI compute, settled on chain, served on validator GPUs,
priced by an AMM that respects capacity decay. The triumvirate makes the
market complete --- HMM provides the price, AIVM provides the logic,
the FHE substrate provides confidentiality. Liquidity Protocol exposes
HMM pools to the rest of the family, and the same hardware that votes
in Quasar 4.0 cert lanes serves the model.

\bibliographystyle{plain}
\begin{thebibliography}{99}

\bibitem{hanzoai} Hanzo. \emph{Hanzo AI Chain --- The AIVM and AI Coin Economics}. 2025.

\bibitem{lightspeed} Hanzo. \emph{Lightspeed DEX --- Native AI-Compute Market Architecture}. 2025.

\bibitem{hanzofhe} Hanzo. \emph{Hanzo FHE Inference --- Confidential Model Serving on F-Chain}. 2025.

\bibitem{hanzotokenomics} Hanzo. \emph{Hanzo Tokenomics --- AI Coin Emission Schedule}. 2025.

\bibitem{hip7} Hanzo. \emph{HIP-7: Active Inference Integration for Hamiltonian LLMs}. 2025.

\bibitem{hip8} Hanzo. \emph{HIP-8: HMM (Hanzo Market Maker) --- Native DEX for AI Compute Resources}. 2025.

\bibitem{lp010v4} Lux Network. \emph{LP-010: QuasarSTM 4.0 --- Production Spec}. 2026.

\bibitem{lp105} Lux Network. \emph{LP-105: Quasar 3.0 Consensus}. 2025.

\bibitem{lp134} Lux Network. \emph{LP-134: Canonical 9-Chain Topology and Primary Chain Templates}. 2026.

\bibitem{lp137} Lux Network. \emph{LP-137: GPU-Residency Invariant for Lux-Family L1s}. 2026.

\bibitem{lp310} Lux Network. \emph{LP-310: Liquidity Protocol Common Settlement API}. 2026.

\bibitem{lp013v2} Lux Network. \emph{LP-013 v2: F-Chain GPU Fully-Homomorphic Encryption Service}. 2025.

\end{thebibliography}

\end{document}