papers/hanzo-ai-chain/sections/04-hmm-registry.tex

\section{Hanzo Models (HMMs)}
\label{sec:hmm}

\subsection{What Lives On Chain}

A Hanzo Model is a content-addressed, versioned bundle of weights,
metadata, and license. The chain does \emph{not} store weight blobs;
those live in content-addressed storage (Hanzo object storage,
\texttt{hanzo/storage}, or any compatible CAS) and are referenced by
their root hash. The chain stores a small fixed-size record per model
version that commits to everything a caller, an operator, or an auditor
needs to verify a job:

\begin{table}[h]
\centering
\small
\begin{tabular}{lp{6.0cm}}
\toprule
\textbf{Field} & \textbf{Definition} \\
\midrule
\texttt{model\_root}      & Merkle root over the weight tensors at
                            version $V$. Content-addressed; pinning the
                            root pins the weights bit-exactly. \\
\texttt{benchmark\_root}  & Merkle root over a canonical evaluation
                            harness output (e.g.\ MMLU, HELM, internal
                            evals). Submitted by the model owner;
                            challengeable on chain. \\
\texttt{license\_root}    & Hash of the license document (commercial,
                            research, or open under explicit terms),
                            with a machine-readable summary. \\
\texttt{inference\_price} & A price function $p(\cdot)$ in AI per
                            output token (or per call, per byte for
                            non-text models). May be a constant, a
                            piecewise function, or a reference to an
                            on-chain pricing oracle. \\
\texttt{training\_history}& Linked list of parent model\_roots, each
                            edge annotated with the training dataset
                            commitment, the trainer identity, and the
                            block height at which training completed. \\
\texttt{owner}            & Address authorized to publish new versions
                            and to update license / price; may be a
                            multisig or a DAO. \\
\bottomrule
\end{tabular}
\caption{HMM registry record. Six fields per model version. The chain
holds the commitments; weights and bench artifacts live in CAS.}
\label{tab:hmm-record}
\end{table}

\subsection{Versioning and Provenance}

Each new version of a model is a new record with a fresh
\texttt{model\_root} and a \texttt{training\_history} edge pointing to
the parent. This produces a directed acyclic graph rooted at the genesis
of each model family. Anyone can walk the graph to see the full
provenance: the original pretraining run, every fine-tune, every
distillation, and the dataset commitment used at each step.

\begin{figure}[h]
\centering
\begin{tikzpicture}[
    node distance=0.55cm and 1.1cm,
    box/.style={rectangle, draw, minimum width=1.9cm, minimum height=0.5cm,
                font=\footnotesize, rounded corners=1pt},
    arrow/.style={->, thick, gray}
]
\node[box] (v0) {hmm:zen-1.0};
\node[box, right=of v0] (v1) {hmm:zen-1.1};
\node[box, right=of v1] (v2) {hmm:zen-2.0};
\node[box, below right=of v0] (ft) {hmm:zen-1.0-medic};
\node[box, below=of v2] (q) {hmm:zen-2.0-q4};
\draw[arrow] (v0) -- (v1) node[midway,above,font=\tiny]{ft +1.2T tok};
\draw[arrow] (v1) -- (v2) node[midway,above,font=\tiny]{full retrain};
\draw[arrow] (v0) -- (ft) node[midway,below,font=\tiny]{specialize};
\draw[arrow] (v2) -- (q) node[midway,right,font=\tiny]{quantize};
\end{tikzpicture}
\caption{Provenance DAG. Each edge is a chain transaction recording the
parent model\_root, the training dataset commitment, and the trainer
identity.}
\end{figure}

\subsection{License Enforcement}

The \texttt{license\_root} commits to a license document with three
machine-readable fields: \emph{usage} (commercial / research / open),
\emph{redistribution} (allowed / restricted / prohibited), and
\emph{derivative} (allowed / restricted / prohibited). Compute operators
are required to read and respect the license at job time; the
\texttt{INFER\_REQUEST} precompile (\S\ref{sec:arch}) refuses to schedule
a job if the caller has not posted a matching license attestation.

\subsection{Price Function}

\texttt{inference\_price} is a first-class on-chain object. The simplest
form is a constant \texttt{uint256} in AI per output token. More
expressive forms are supported:

\begin{itemize}[leftmargin=1.2em]
  \item \emph{Piecewise tiered}: a price ladder that lowers per-token
        cost above batch thresholds.
  \item \emph{Time-of-day}: a function over block timestamp, used to
        clear demand at peak.
  \item \emph{Oracle-backed}: a reference to an on-chain price oracle
        whose own update mechanism is visible on chain.
\end{itemize}

The model owner sets the function and may update it; updates are
rate-limited (one update per epoch by default) to prevent surprise
price moves between job submission and execution.

\subsection{Benchmark Commitments and Challenges}

A model owner publishes a \texttt{benchmark\_root} when registering a
version. Any address may post a \emph{challenge}: a deposit-bonded
claim that the published benchmark output does not match the actual
output of the published \texttt{model\_root} on the canonical eval
harness. Challenges are resolved under an activated A-Chain proof policy:
deterministic replay or an optimistic integer transcript for public jobs, with
optional TEE evidence when the evaluation is confidential. If the
challenge succeeds, the challenger collects the model owner's
benchmark bond; if it fails, the challenger forfeits theirs. This
matches the dispute pattern used elsewhere in Lux for attestation
challenges (LP-134 §6).

\subsection{Candidate genesis model registry}

The proposed genesis state may seed the following Hanzo model families after
their exact roots, licenses, execution policies, and proof profiles are frozen:

\begin{itemize}[leftmargin=1.2em]
  \item \texttt{hmm:zen-text} (text-to-text foundation model);
  \item \texttt{hmm:zen-vision} (multimodal vision-language);
  \item \texttt{hmm:zen-audio} (speech and audio);
  \item \texttt{hmm:zen-code} (code-specialized);
  \item \texttt{hmm:zen-embed} (embedding model).
\end{itemize}

Registry activation is separate from ordinary gateway availability. A model is
PoAI-mineable only after its deterministic execution and proof policy pass the
deployment gates in \S\ref{sec:deploy}.