\section{Corpus and Method}
\label{sec:method}

\subsection{Corpus construction}

Internet-Drafts were collected from the IETF Datatracker. Candidate documents
were identified by keyword and topic matching for AI-agent and autonomous-system
terminology (e.g.\ ``agent'', ``autonomous'', ``LLM'', ``inference''), then
filtered to remove false positives---documents in which such terms occur in an
unrelated sense (``user agent'', ``autonomous system'' in the BGP sense). Of
597 candidate IETF documents, 73 ($12.2\%$) were flagged as false positives and
excluded, yielding a clean corpus of \textbf{524 IETF Internet-Drafts}.

We restrict this survey to IETF Internet-Drafts. The collection pipeline also
ingests documents from other standards bodies (ISO, ITU, ETSI, NIST, W3C), but
those exhibit heterogeneous and frequently missing publication dates and a
different document model; mixing them would compromise the temporal and
authorship analyses. All 524 IETF drafts carry ISO-8601 submission timestamps,
so the IETF-only scope is also the cleanest with respect to date quality. The
corpus snapshot used throughout is \texttt{data/drafts.db} as of 2026-05-23.

\subsection{Thematic classification}

Each draft is assigned to one or more of eleven thematic categories
(Table~\ref{tab:categories}) by an LLM-assisted rating pipeline. The classifier
is prompted with the draft's title, metadata, and abstract (truncated to the
first 2000 characters), and returns a JSON object containing the category
assignment and five ordinal $1$--$5$ quality dimensions (novelty, maturity,
overlap, momentum, relevance). We emphasise two properties of this pipeline
that bear directly on interpretation:

\begin{itemize}[leftmargin=1.4em,topsep=2pt,itemsep=1pt]
  \item \textbf{Abstract-only.} Ratings are derived from the abstract, not the
        full document text. This keeps the pipeline cheap and uniform but means
        classifications reflect how a draft \emph{presents} itself, not a full
        reading of its mechanics.
  \item \textbf{Model-derived.} Categories and scores are produced by a large
        language model (Claude Sonnet, \texttt{claude-sonnet-4-20250514}).
        Section~\ref{sec:reliability} quantifies how reproducible these labels
        are; the headline result is that we trust the categories but not the
        ordinal scores.
\end{itemize}

\subsection{Semantic embeddings}

For the redundancy analysis (Section~\ref{sec:redundancy}) each draft is
embedded into a 768-dimensional vector using the \texttt{nomic-embed-text}
model run locally via Ollama. Embedding coverage is complete (524/524). We
compute cosine similarity over all $\binom{524}{2}=137{,}026$ document pairs.

\subsection{Reproducibility}

The pipeline, queries, and the two-model re-rating used for the reliability
analysis are released as scripts (\texttt{scripts/survey-phase0.py},
\texttt{scripts/rerate-intercoder.py}, \texttt{scripts/survey-kappa.py}). The
re-rating itself was executed through the Anthropic Batch API at a total cost of
USD~2.41. Raw model outputs are released as line-delimited JSON.