mirror of
https://github.com/hanzoai/papers.git
synced 2026-10-29 07:34:31 +00:00
162 lines
7.2 KiB
TeX
162 lines
7.2 KiB
TeX
\section{Architecture}
|
|
\label{sec:arch}
|
|
|
|
\subsection{AI Chain is the Lux A-Chain}
|
|
|
|
Hanzo AI Chain is the public identity of the Lux A-Chain when it operates the
|
|
open AI protocol. There is not a Hanzo proof chain beside A-Chain. A single
|
|
canonical state machine owns job admission, proof policy, challenge state, work
|
|
nullifiers, reward authorization, and receipt roots.
|
|
|
|
\begin{figure}[h]
|
|
\centering
|
|
\resizebox{\linewidth}{!}{%
|
|
\begin{tikzpicture}[
|
|
node distance=0.8cm and 0.85cm,
|
|
box/.style={rectangle, draw, minimum width=2.3cm, minimum height=0.7cm,
|
|
font=\small, rounded corners=1pt},
|
|
arrow/.style={->, thick}
|
|
]
|
|
\node[box] (job) {Demand-backed AI job};
|
|
\node[box, right=of job] (exec) {CPU/GPU execution};
|
|
\node[box, right=of exec] (a) {A-Chain PoAI};
|
|
\node[box, right=of a] (z) {Z-Chain P3Q};
|
|
\node[box, right=of z] (receipt) {Settled receipt};
|
|
\node[box, below=of z] (lux) {Lux Quasar + PQ policy};
|
|
\node[box, below=of receipt] (dest) {Bound destination};
|
|
\draw[arrow] (job) -- (exec);
|
|
\draw[arrow] (exec) -- node[above,font=\scriptsize]{commit + proof} (a);
|
|
\draw[arrow] (a) -- (z);
|
|
\draw[arrow] (z) -- (receipt);
|
|
\draw[arrow] (lux) -- (z);
|
|
\draw[arrow] (receipt) -- (dest);
|
|
\end{tikzpicture}
|
|
}
|
|
\caption{PoAI validates useful work on A-Chain; P3Q settles it on Z-Chain; Lux
|
|
supplies PQ-rooted finality; destinations consume receipts rather than raw work.}
|
|
\label{fig:arch}
|
|
\end{figure}
|
|
|
|
\subsection{PoAI useful-work state machine}
|
|
|
|
Each job progresses through
|
|
|
|
\begin{equation}
|
|
\begin{aligned}
|
|
\textsc{Requested}&\rightarrow\textsc{Committed}
|
|
\rightarrow\textsc{Verifying/Challengeable}\\
|
|
&\rightarrow\textsc{Finalized}\;|\;\textsc{Rejected}\;|\;\textsc{Expired}.
|
|
\end{aligned}
|
|
\end{equation}
|
|
|
|
A-Chain fixes the execution and proof policy before commitment. It derives any
|
|
random challenge afterward, resolves the policy to final or rejected, checks
|
|
metering and reward allowance, consumes the global nullifier, and creates the
|
|
destination-bound receipt atomically. An optimistic result cannot become
|
|
issuance-final while its challenge window remains open.
|
|
|
|
\subsection{Z-Chain settlement and Lux base-layer finality}
|
|
|
|
Z-Chain is the P3Q proof-compression and settlement rollup for finalized A-Chain
|
|
receipt transitions. Its intended proof system is a Plonky3-derived,
|
|
pairing-free STARK/FRI stack. Z-Chain does not admit jobs, determine usefulness,
|
|
own the mining nullifier, or authorize subsidy; its proof must bind exactly to
|
|
the finalized A-to-Z receipt transition.
|
|
|
|
Lux Quasar orders and finalizes A-Chain and Z-Chain blocks. Exported roots must be
|
|
authenticated under the validator and post-quantum policy that is actually
|
|
activated. The security statement therefore names its assumptions: validator
|
|
set, threshold, key lifecycle, downgrade prevention, and relay verification.
|
|
Selecting ML-DSA in a registry does not by itself produce a threshold finality
|
|
certificate.
|
|
|
|
Pulsar and Corona explore threshold lattice signatures for consensus and
|
|
custody. P3Q is the intended Z-Chain settlement substrate, but the exact PoAI
|
|
AIR, recursion policy, state-transition binding, and verifier remain subject to
|
|
protocol, audit, integration, and activation gates. This paper does not treat a
|
|
repository or precompile address as deployment evidence.
|
|
|
|
\subsection{Deterministic execution substrate}
|
|
|
|
The public proof-bearing path admits a bounded deterministic graph whose
|
|
operations define:
|
|
|
|
\begin{itemize}[leftmargin=1.2em]
|
|
\item integer weights and activations, normally \texttt{int8};
|
|
\item wider accumulators with explicit overflow or saturation behavior;
|
|
\item canonical tensor layout, serialization, reduction, and rounding;
|
|
\item deterministic routing, normalization, sampling, and tie-breaking;
|
|
\item content-addressed model, input, output, runtime, and graph roots.
|
|
\end{itemize}
|
|
|
|
Execution may occur on a CPU, GPU, NPU, or other backend that produces the same
|
|
committed bytes. Accelerators improve throughput; they are not a consensus
|
|
requirement for every validator.
|
|
|
|
\subsection{Execution profile and canonical transcript}
|
|
|
|
Determinism is fixed by a versioned \emph{kernel profile}, not by naming a
|
|
particular device. The profile commits the admitted operators, tensor layout,
|
|
integer widths, accumulator bounds, overflow behavior, reduction order,
|
|
fixed-point scales, rounding, routing tie-breaks, and byte encoding. A cloud
|
|
provider may run optimized CPU, GPU, or accelerator kernels, but those kernels
|
|
are PoAI-compatible only when they implement the same profile exactly.
|
|
|
|
Every execution emits one \emph{canonical transcript}. Its header binds the
|
|
A-Chain job and policy epoch, kernel-profile version, model, runtime, input,
|
|
declared graph, output, metering, provider, destination, and nonce. Its ordered
|
|
body binds every typed operation to its input and output commitments. The root
|
|
of this content-addressed serialization is the object challenged by A-Chain and,
|
|
after PoAI finality, settled by P3Q on Z-Chain. The kernels define the function;
|
|
the transcript identifies and commits the particular paid invocation.
|
|
|
|
CPU/GPU parity is a conformance gate for a profile release, not a mining
|
|
primitive. Golden vectors must cover kernel outputs, operation leaves,
|
|
transcript roots, challenge derivation, and openings. A backend that diverges is
|
|
not admitted; consensus does not introduce floating-point tolerance to
|
|
accommodate it.
|
|
|
|
\subsection{Evidence backends}
|
|
|
|
The job selects one activated evidence profile:
|
|
|
|
\begin{table}[h]
|
|
\centering
|
|
\small
|
|
\begin{tabular}{p{0.23\linewidth}p{0.67\linewidth}}
|
|
\toprule
|
|
\textbf{Profile} & \textbf{Role} \\
|
|
\midrule
|
|
Exact replay & Canonical CPU or verifier-set recomputation for small jobs. \\
|
|
Optimistic integer & Transcript commitment, data availability, watcher
|
|
re-execution, and objective challenge openings. \\
|
|
Succinct P3Q & Plonky3-derived Z-Chain settlement only after the exact PoAI AIR,
|
|
A-to-Z binding, recursion policy, and verifier are audited and activated. \\
|
|
Confidential compute & Optional vendor attestation for private prompts, weights,
|
|
or outputs, combined with explicit TCB and revocation policy. \\
|
|
Subjective quorum & Attributable judgment that may earn demand fees but cannot
|
|
alone authorize mining subsidy. \\
|
|
\bottomrule
|
|
\end{tabular}
|
|
\caption{Evidence profiles are selected by job policy; none creates another
|
|
mining authority outside A-Chain.}
|
|
\end{table}
|
|
|
|
\subsection{Canonical protocol objects}
|
|
|
|
The minimum consensus objects are \texttt{Mining\allowbreak Job},
|
|
\texttt{Execution\allowbreak Commitment}, \texttt{Challenge},
|
|
\texttt{Proof\allowbreak Finality}, \texttt{Work\allowbreak Nullifier}, and
|
|
\texttt{Mining\allowbreak Receipt}. Public APIs are submission
|
|
and query surfaces for these objects. An in-memory task server, GPU scheduler,
|
|
or destination smart contract is not the canonical mining state machine.
|
|
|
|
\subsection{Implementation status}
|
|
|
|
The exact-integer arithmetic core, Go/Rust vectors, Solidity witness fixtures,
|
|
and A-Chain receipt components exist as implementation evidence. The production
|
|
system still requires a complete Zen model transcript, data-availability path,
|
|
finalized-not-pending proof gate, verifiable metering, token-policy upgrade, PQ
|
|
root export, Z-Chain settlement, and exactly-once destination consumption. Mining remains disabled
|
|
until the launch gates in \S\ref{sec:deploy} pass.
|