papers/hanzo-federated-agents/sections/03-system-architecture.tex

\section{System Architecture}

\label{sec:architecture}

The FAI system consists of three components running at each agent node:
(i) the local node with experience library and GRPO optimizer,
(ii) the network layer for P2P broadcast and receive, and
(iii) the aggregation layer implementing DeltaSoup.

\subsection{Local Node}

\paragraph{SemanticMemoryManager.}
Each node maintains a \emph{local experience library} $\mathcal{L}$ as a
vector database (LanceDB~\cite{lancedb2024}). Each entry stores:
\begin{itemize}[leftmargin=1.5em]
    \item The pattern text (natural language, $\leq 32$ words).
    \item A $d$-dimensional embedding $\bm{e}_m \in \mathbb{R}^d$ (default $d = 768$).
    \item Quality score $q_m \in (0,1)$, domain tag, and provenance metadata.
    \item A 1-bit quantized expert factor for PoE decoding.
\end{itemize}
The library supports approximate nearest-neighbor queries over embeddings,
enabling retrieval of relevant experiences for new tasks in $O(\log M)$ time
where $M = |\mathcal{L}|$.

\paragraph{LocalDSOOptimizer.}
The optimizer runs in the background after each agent session:
\begin{enumerate}[leftmargin=1.5em]
    \item Collect tool-use telemetry from the session (tool calls, outcomes,
          reward signals).
    \item Construct agent trajectories grouping related tool calls.
    \item Run GRPO on the trajectory group to extract semantic advantages.
    \item Compress resulting experience embeddings via BitDelta.
    \item Store in $\mathcal{L}$ and enqueue for DSO broadcast.
\end{enumerate}

\paragraph{Tool Telemetry Pipeline.}
The telemetry pipeline intercepts tool calls at the harness layer and records:
\begin{equation}
    \tau_t = \left(\text{tool}_t, \; \text{input}_t, \; \text{output}_t, \; r_t\right),
\end{equation}
where $r_t \in \mathbb{R}$ is a task-specific reward signal (e.g., test pass
rate, user rating, completion confirmation). A trajectory is a sequence
$\tau = (\tau_1, \ldots, \tau_T)$ associated with a single task.

\subsection{Network Layer}

\paragraph{P2P Gossip Protocol.}
Agents communicate via an epidemic gossip protocol~\cite{demers1987epidemic}.
Each node maintains a peer list of size $k_{\text{peers}}$ (default 8) and
broadcasts compressed experience batches at configurable intervals
(default every 60 seconds or after accumulating $B_{\min} = 10$ new experiences).

\paragraph{Node Discovery.}
Initial peers are bootstrapped from:
\begin{itemize}[leftmargin=1.5em]
    \item A configurable seed list (static bootstrap nodes).
    \item The DSO registry smart contract~\cite{hanzo2026dso} (on-chain
          peer advertisement).
    \item mDNS for local-network discovery in development settings.
\end{itemize}
Each node advertises its address by posting a signed advertisement to the
DSO registry. Advertisements expire after 24 hours and must be renewed.

\paragraph{Hash-Based Integrity.}
Each broadcast batch is accompanied by a SHA-256 hash over its canonical
serialization. Receiving nodes verify the hash before processing; batches
with invalid hashes are silently dropped and the sending node's reputation
score is decremented.

\subsection{Compressed Experience Format}

We define the \texttt{CompressedBatch} wire format as a JSON-serializable
protocol message:

\begin{verbatim}
{
  "node_id":   string,          // SHA-256 of node public key
  "timestamp": number,          // Unix milliseconds
  "experiences": {
    "<exp_id>": {
      "text":       string,     // Pattern description (<= 32 words)
      "signs":      int[],      // Packed 1-bit sign vectors (uint32 words)
      "scale":      float,      // Per-group L1 scale factor
      "shape":      int[],      // [num_groups, group_size]
      "confidence": float,      // Quality score in (0, 1)
      "domain":     string      // Domain tag (e.g. "coding", "writing")
    }
  },
  "hash":      string           // SHA-256 of canonical serialization
}
\end{verbatim}

\begin{definition}[Canonical Serialization]
The canonical serialization of a \texttt{CompressedBatch} is the
deterministic JSON serialization with keys sorted lexicographically and
no whitespace. The \texttt{hash} field is excluded during hash computation
and appended afterward.
\end{definition}

For $M$ experiences each with embedding dimension $d = 768$ using group size
$g = 128$, the compressed batch size is approximately:
\begin{equation}
    |B_{\text{compressed}}| = M \cdot \left(\frac{d}{8} + \frac{d}{g} \cdot 4 + 64\right) \; \text{bytes},
\end{equation}
where $d/8$ bytes stores the sign bits, $d/g \cdot 4$ bytes stores float32
scales, and 64 bytes stores metadata. For $M = 100$: ${\approx}29\,\text{KB}$
versus ${\approx}307\,\text{KB}$ uncompressed---a $10.6\times$ reduction.

% ============================================================================
% 4. BITDELTA EXPERIENCE COMPRESSION
% ============================================================================