papers/hanzo-ai-chain/sections/01-introduction.tex

\section{Introduction}
\label{sec:intro}

\subsection{Why an AI Chain}

The economics of artificial intelligence are concentrating. A small number of
hyperscalers control the GPUs that train frontier models, the model weights
themselves, the APIs through which those models are sold, and the price at
which inference is offered. This concentration produces three structural
problems for everyone who is not a hyperscaler. First, \emph{access}:
researchers, agents, and applications cannot build on weights they cannot
inspect or fine-tune. Second, \emph{pricing}: there is no spot market for
inference, only published price lists. Third, \emph{provenance}: model
outputs arrive without a verifiable record of which model produced them, on
what hardware, at what version. A user must trust the API.

The Hanzo AI Chain replaces those three trust assumptions with verifiable
on-chain primitives. Models are content-addressed and versioned on chain.
Inference is sold through an open market in which any operator may list
capacity and any caller may submit jobs. Each job receipt is a Quasar
certificate (LP-105) that attests, in a single 192-byte SNARK plus a
32-byte public-input hash, both that the consensus quorum signed off on
the result and that the executing GPU produced a TEE quote pinning the
exact model commitment, hardware identity, and runtime version.

\subsection{Decentralized AI as a Three-Sided Market}

The AI Chain organizes the global compute market as three interlocking
sides:

\begin{itemize}[leftmargin=1.2em]
  \item \textbf{Model owners} publish weights, license terms, evaluation
        results, and an inference price function. The chain stores
        commitments, not weights; large weight blobs live in content-
        addressed storage and are referenced by their root hash.
  \item \textbf{Compute operators} stake AI to register GPU capacity,
        produce TEE attestations on every job, and earn a share of the
        per-token fee plus a block reward.
  \item \textbf{Callers} (humans, applications, agents) submit jobs,
        pay in AI, and receive a Quasar-signed receipt that proves which
        model produced which output on which hardware.
\end{itemize}

These three sides are mediated by a fourth component, the \emph{agent
registry}, which gives autonomous Hanzo Agents
(\texttt{\textasciitilde/work/hanzo/agents}) a persistent on-chain identity,
an explicit capability declaration, and a reputation that follows them
across jobs.

\subsection{Position vs.\ Centralized AI Labs}

A centralized AI lab is, from a buyer's standpoint, an opaque oracle: input
in, output out, no proof of anything in between. The AI Chain offers the
opposite primitive. Every output is accompanied by:

\begin{enumerate}[leftmargin=1.4em]
  \item A \emph{model commitment} (\S\ref{sec:hmm}) pinning the exact
        weights and architecture that produced the output;
  \item A \emph{TEE attestation quote} (LP-134) pinning the hardware
        identity and runtime version;
  \item A \emph{QuasarCert} (LP-105) anchoring both of the above into the
        canonical chain history with three independent post-quantum
        verification lanes.
\end{enumerate}

This is the same security envelope the Lux network uses for value
settlement, applied to AI computation. It is not an additional system;
it is a re-framing of A-Chain as the AI Virtual Machine, with a native
coin (AI), a registry of models (HMMs), and a market structure designed
for compute rather than transfer.

\subsection{Contributions and Outline}

\begin{enumerate}[leftmargin=1.4em]
  \item The AI Chain architecture (\S\ref{sec:arch}): A-Chain reframed
        as the AIVM, settled by Lux Quasar 3.0 (LP-105) and executed on
        QuasarGPU (LP-132) with QuasarSTM lane scheduling (LP-010).
  \item The AI native coin (\S\ref{sec:aicoin}): block reward, fee burn,
        staking yields, and operator economics for the Hanzo Network.
  \item The HMM registry (\S\ref{sec:hmm}): on-chain provenance,
        versioning, license, benchmark, and price commitments for every
        Hanzo model.
  \item Three composable compute markets (\S\ref{sec:markets}): per-token
        inference, training bounties, and GPU pool capacity.
  \item The agent registry (\S\ref{sec:agents}): persistent identity,
        capability declaration, and on-chain reputation for
        autonomous agents.
  \item Cross-chain integration (\S\ref{sec:xchain}) with Lux C-Chain
        (EVM), F-Chain (FHE inference), M-Chain (threshold custody),
        and B-Chain (bridges) per LP-134.
  \item Confidential inference (\S\ref{sec:privacy}) and three-lane
        post-quantum security (\S\ref{sec:security}).
  \item Deployment plan (\S\ref{sec:deploy}): activation 25 December 2025
        on the same Quasar 3.0 boundary that activates LP-010, LP-105,
        LP-132, LP-134.
\end{enumerate}