papers/hanzo-dso/sections/04-dso-protocol-specification.tex

\section{DSO Protocol Specification}

\label{sec:protocol}

\subsection{Roles}

The DSO protocol defines five roles:

\begin{enumerate}[leftmargin=1.5em]
    \item \textbf{Agents:} Run local ASO, extract experiential priors, submit
          to registry. Each agent has a keypair $(pk_i, sk_i)$ and stake $S_i$.
    \item \textbf{Aggregators:} Compute Byzantine-robust aggregations of
          submitted priors. May be a subset of agents or dedicated nodes.
    \item \textbf{Validators:} Verify PoAI attestations, vote on quality scores,
          execute slashing for malicious submissions.
    \item \textbf{Storage Providers:} Pin prior data on IPFS/Arweave, earn
          propagation rebates.
    \item \textbf{Registry:} Smart contract managing entries, bonds, quality
          votes, and aggregation requests.
\end{enumerate}

\subsection{Protocol Phases}

\begin{algorithm}[H]
\caption{DSO Protocol: Full Lifecycle}
\label{alg:dso-lifecycle}
\begin{algorithmic}[1]
\Require Agent set $\mathcal{A} = \{a_1, \ldots, a_n\}$, task distribution $\mathcal{D}$
\Ensure Aggregated prior bank $\mathcal{E}_{\text{agg}}$

\State \texttt{// Phase 1: Local Prior Generation}
\For{each agent $a_i \in \mathcal{A}$ \textbf{in parallel}}
    \State Sample task $x \sim \mathcal{D}$
    \State $\mathcal{E}_i \gets \text{ASO}(x, \pi_\theta)$
    \Comment{Local TF-GRPO + PoE}
\EndFor

\State \texttt{// Phase 2: Submission}
\For{each agent $a_i$ with priors $\mathcal{E}_i$}
    \For{each prior $E \in \mathcal{E}_i$}
        \State $\texttt{cid} \gets \text{IPFS.Pin}(\text{Serialize}(E))$
        \State $\texttt{root} \gets \text{MerkleRoot}(E)$
        \State $\texttt{id} \gets \text{Registry.Submit}(\texttt{root}, \texttt{cid}, q_E, D)$
        \Comment{Bond $D$ required}
    \EndFor
\EndFor

\State \texttt{// Phase 3: Quality Verification}
\For{each validator $v_j$}
    \State Sample subset of submissions $\mathcal{S}_j$
    \For{each $\texttt{id} \in \mathcal{S}_j$}
        \State $E \gets \text{IPFS.Fetch}(\texttt{cid}_{\texttt{id}})$
        \State Verify Merkle proof
        \State $q' \gets \text{EvaluateQuality}(E, \text{holdout tasks})$
        \State $\text{Registry.VoteQuality}(\texttt{id}, q')$
    \EndFor
\EndFor

\State \texttt{// Phase 4: Aggregation}
\State $\texttt{ids} \gets \text{Registry.GetEntriesBySchema}(\texttt{schema})$
\State Filter: $\texttt{ids} \gets \{\texttt{id} : q_{\texttt{id}} \geq q_{\min}\}$
\State $\mathcal{E}_{\text{agg}} \gets \text{ByzantineAggregate}(\texttt{ids})$
\Comment{Section~\ref{sec:aggregation}}

\State \texttt{// Phase 5: Gossip Propagation}
\For{each node $n_k$}
    \State Subscribe to registry events
    \State Fetch and verify high-quality priors
    \State Merge into local LanceDB
    \Comment{Section~\ref{sec:gossip}}
\EndFor

\State \Return $\mathcal{E}_{\text{agg}}$
\end{algorithmic}
\end{algorithm}

\subsection{Schema Decomposition}

Aggregation requires a common decomposition schema $\mathcal{S}$ so that priors
from different agents can be compared element-wise.

\begin{definition}[Schema]
A schema $\mathcal{S} = \{s_1, \ldots, s_K\}$ partitions the token-level
expert factor space into $K$ bins. Each bin $s_k$ corresponds to either:
\begin{enumerate}[leftmargin=1.5em]
    \item A \textbf{token bin:} A subset of vocabulary tokens grouped by
          semantic similarity (e.g., all Python keywords, all arithmetic
          operators).
    \item An \textbf{embedding centroid:} A region in the task embedding space
          identified by $k$-means clustering.
    \item A \textbf{context type:} A category of generation contexts (e.g.,
          function signature, loop body, error handling).
\end{enumerate}
\end{definition}

Each submitted prior $E_i$ is decomposed under the schema:
\begin{equation}
    E_i = \{\Delta_i^{(s)}\}_{s \in \mathcal{S}},
\end{equation}
where $\Delta_i^{(s)} \in \{-\alpha_i, 0, +\alpha_i\}^{|s|}$ is the 1-bit
quantized factor restricted to schema element $s$.

% ============================================================================
% 5. BYZANTINE-ROBUST AGGREGATION
% ============================================================================