papers/zoo-mpc-custody/sections/04-model-weight-custody.tex

\section{Model Weight Custody}

\label{sec:model-custody}

\subsection{Model NFT Standard}

On Zoo Network, trained AI models are represented as Model NFTs (ERC-721 tokens) with the following metadata:

\begin{lstlisting}[basicstyle=\ttfamily\small,caption=Model NFT Metadata]
{
  "name": "Snow Leopard Classifier v3",
  "modelHash": "0xabc...def",
  "architecture": "ResNet-50",
  "trainingDataHash": "0x123...456",
  "accuracy": 0.947,
  "custodyKey": "zoo-mpc-0x789...",
  "threshold": [2, 3],
  "roles": ["researcher", "dao", "escrow"],
  "deploymentPolicy": "requires_dao_vote"
}
\end{lstlisting}

\subsection{Custody Operations}

The MPC key associated with a Model NFT authorizes four operations:

\begin{enumerate}
    \item \textbf{Deploy}: Make the model available for inference on Zoo Network. Requires researcher + DAO (2-of-3).

    \item \textbf{Fine-tune}: Authorize fine-tuning with new data. Requires researcher + DAO (2-of-3).

    \item \textbf{Transfer}: Transfer model ownership to another party. Requires researcher + DAO (2-of-3) or DAO + escrow (after 48h delay).

    \item \textbf{Retire}: Permanently disable the model. Requires any 2-of-3.
\end{enumerate}

\subsection{Governance Gate}

Model deployment goes through a governance gate:

\begin{algorithm}[t]
\caption{Model Deployment with MPC Governance}
\label{alg:deploy}
\begin{algorithmic}[1]
\STATE Researcher submits deployment proposal on-chain
\STATE DAO reviews model safety report (bias audit, accuracy, risk assessment)
\STATE DAO votes on deployment (ZOO token-weighted, 72h voting period)
\IF{vote passes ($> 50\%$ quorum, $> 66\%$ approval)}
    \STATE DAO activates its key share
    \STATE Researcher activates its key share
    \STATE MPC produces deployment signature
    \STATE Model deployed to Zoo inference network
\ELSE
    \STATE Deployment rejected; model remains locked
\ENDIF
\end{algorithmic}
\end{algorithm}