\section{Reflexive analysis of the WP29 production process}
\label{sec:reflexive}

This paper was itself produced under a mandate of the kind it studies, and the mandate requires that the production process be reported as an instrumented case study, kept separate from the confirmatory evidence. We do so here from the append-only phase log (\texttt{reports/phase\_status.md}), the amendment ledger, the intervention ledger and the git history. Nothing in this section enters the statistical analysis.

\subsection{Who did what}

The researcher prepared the mandate outside the session and pasted it into a Claude Code session (client 2.1.314, model \texttt{claude-fable-5-1}) with a four-word instruction in French to carry out the project. From that point until the delivery of the manuscript, no human intervened in the session: every design decision, line of code, task, oracle, run, capture and sentence was produced by the builder session or by the sub-sessions it delegated to. The builder delegated thirteen pieces of work to fresh general-purpose sub-sessions: the corpus coding, the literature review (84 verified references), one task family each for F1--F6 (one development and four confirmatory tasks per family, with hidden oracles and validation sets), the three transfer mini-projects, the drafting of the related-work and corpus sections of this paper, and three adversarial reviews (systems and isolation; statistics and evaluators; manuscript claims), whose findings and the builder's responses are recorded in the review files of the repository. The builder wrote the harness, the protocols, the smoke task, the analysis pipeline and the remaining sections, ran every experiment, and reviewed the sub-sessions' deliverables by executing their validation sets rather than by reading their reports alone.

The researcher's preparation time for the mandate is unknown and is reported as unknown. The builder session's own effort is observable in two units: wall-clock time from the first command (2026-10-29 02:38 UTC) to the final build, and the token usage of the builder and its sub-sessions as reported by the client; both are given in Table~\ref{tab:effort}.

\begin{table}[htbp]\centering\small
\caption{Human and builder effort with its measurement basis, and inference cost by split. Unknown and unmeasured quantities are shown as such.}
\label{tab:effort}
\input{../results/tables/tab_effort}
\end{table} These are costs of the method, not interventions in the experiment, and they dwarf the inference cost of the benchmark runs themselves.

\subsection{What the process looked like from the inside}

The log shows a sequence that resembles the protocol imposed on arm B more than the open instruction of arm A, which is unsurprising since the mandate prescribes phases and gates. Three features are worth recording because they are the kind of thing an agent's own summary would omit.

First, infrastructure failures preceded every stage of live experimentation and were caught by the harness's own checks rather than by inspection of outputs: the first smoke attempt failed on all four runs before any inference (an unexpanded tilde in a quoted remote path made the workspace-manifest verification fail, and a shared SQLite connection used from several threads returned a spurious missing-row error); the pilot exposed that the final evaluator was being invoked on a workspace without the task inputs, so that four correct source-grounded answers were scored as fabrications. Each failure is in the ledger with its original and corrected receipts (amendment A-002). An executor that had declared success at the first green exit code would have reported a benchmark in which arm-independent evaluator bugs masqueraded as scientific failures.

Second, the two subagents that authored F2 and F6 independently chose the same public series and the same regression for one confirmatory task each (Okun's law on quarterly FRED data, 1960--2019). The overlap was detected at registry merge time and reported rather than silently removed, because removing a task after seeing the suite but before seeing results is a choice that an audit should be able to inspect.

Third, visual evidence was harder to obtain honestly than numerical evidence. Native window capture of the browser produced the browser chrome without the page; two such captures were recorded as failures in the screenshot manifest and replaced by composited region captures, and the first capture of a ``live'' terminal matched the builder's own session window by title. All of these artefacts are retained with their labels. The temptation that the mandate warns against, namely to render an HTML imitation of a terminal, never arose as a technical necessity, but it is easy to see how it would arise under time pressure.

\subsection{What this case study does and does not show}

It shows that a mandate of this form can be executed end to end by the client under study, with every intermediate artefact hashed and every deviation logged. It does not show that the execution was autonomous in the scientific sense: the mandate was written by a human who knew what a credible study looks like, the task families and the hypotheses were given, and the same organisation controlled the builder, the tasks and the executors. It is one instrumented case, not an observation of the benchmark, and we draw no inference from it beyond the descriptive account above.