papers/zoo-experience-ledger/sections/03-experience-format-specification.tex

\section{Experience Format Specification}

\label{sec:format}

\subsection{Semantic Experience Definition}

\begin{definition}[Semantic Experience]
A semantic experience $e$ is a tuple:
\begin{equation}
e = (t, d, c, v, m, \sigma)
\end{equation}
where:
\begin{itemize}
    \item $t \in \Sigma^*$: Natural language text (max 256 tokens)
    \item $d \in \mathcal{D}$: Domain taxonomy path (e.g., ``math.geometry.intersection'')
    \item $c \in [0,1]$: Confidence score derived from validation
    \item $v \in \mathbb{R}^{768}$: Semantic embedding vector
    \item $m$: Metadata (timestamp, author, provenance)
    \item $\sigma$: Digital signature (optional)
\end{itemize}
\end{definition}

\subsection{Canonical Serialization}

Experiences are serialized using CBOR (Concise Binary Object Representation) with deterministic encoding:

\begin{lstlisting}[language=json,caption=Experience CBOR Schema]
{
  "version": 1,
  "type": "experience/v1",
  "content": {
    "text": "When solving geometry problems...",
    "domain": ["math", "geometry", "intersection"],
    "confidence": 0.87
  },
  "embedding": <768-dim float16 array>,
  "metadata": {
    "created": 1634567890,
    "author": "did:key:z6Mk...",
    "source": "training_run_42"
  },
  "signature": <optional Ed25519 signature>
}
\end{lstlisting}

\subsection{Content Addressing}

The experience CID is computed as:

\begin{equation}
\text{CID}(e) = \text{CIDv1}(\text{dag-cbor}, \text{sha256}(\text{CBOR}(e)))
\end{equation}

This ensures:
\begin{itemize}
    \item \textbf{Determinism}: Same experience always produces same CID
    \item \textbf{Integrity}: Any modification changes the CID
    \item \textbf{Self-certification}: CID proves content authenticity
\end{itemize}

\subsection{Constraints and Validation}

Valid experiences must satisfy:

\begin{enumerate}
    \item \textbf{Length}: $|\text{tokens}(t)| \leq 256$
    \item \textbf{Structure}: Text must follow ``When [context], [action]'' pattern
    \item \textbf{Generality}: No problem-specific constants
    \item \textbf{Embedding Consistency}: $\|v - \text{embed}(t)\|_2 < \epsilon$
    \item \textbf{Signature Validity}: If $\sigma$ present, must verify against author key
\end{enumerate}