% ---------------------------------------------------------------------------
% Disclosure and research-integrity statement (Agent H). Facts from CLAUDE.md
% (provenance note), docs/PHASE_0_DISCOVERY.md, docs/PHASE_5_AUDIT.md,
% protocol/amendments.csv, README.md.
% ---------------------------------------------------------------------------
\section*{Disclosure and research-integrity statement}\label{sec:disclosure}
\addcontentsline{toc}{section}{Disclosure and research-integrity statement}

\paragraph{How the study was produced.} This study was designed, built, executed, analysed and written by Claude Code agents acting under a research mandate given by the principal investigator on 2 November 2026. The controller session that held the Research Director role, drafted the preregistration, orchestrated every phase and launched the sub-sessions ran on \texttt{claude-fable-5-1} in Claude Code 2.1.316; the scenario generators (Agent B) and the scenario validators (Agent C) were fresh Claude Code sub-sessions launched from the controller, one per scenario and per role, with no shared conversation; the autonomous workers (Agent D) were independent \texttt{claude --bare} processes on four executor nodes running Claude Code 2.1.314, using \texttt{claude-opus-5-5} in the pilot and \texttt{claude-sonnet-5-5} and \texttt{claude-fable-5-1} in the model sensitivities; the blind judges (Agent F) were one draw of \texttt{claude-opus-5-5} and two of \texttt{claude-sonnet-5-5} per run; the deterministic evaluators (Agent E) are code; the analysis scripts (Agent G) and the manuscript sections (Agent H) were produced by further sub-sessions working from the validated data and the analysis outputs, and the reproducibility audit (Agent I) by a fresh sub-session with access to all records. A single \texttt{claude-haiku-5-5} session was used once, to verify the repaired recovery mechanism (amendment A-012). Model and client identifiers are recorded in every session's result event and in the reproducibility package.
% docs/PHASE_0_DISCOVERY.md §0.1–0.3; protocol/amendments.csv A-004, A-005, A-012; RESEARCH_PLAN.md §7

\paragraph{The principal investigator's role.} The principal investigator set the mandate, provided the computing infrastructure and the API credentials, and bears responsibility for the work; the executing agents' decisions on matters the mandate left open are logged as amendments A-001 to A-012 and are pending the principal investigator's confirmation where so marked. The principal investigator's review of the manuscript, the benchmark and the records is pending at the time this version was produced, and the independent human expert review of deliverables that the mandate calls for has not been performed (amendment A-005).

\paragraph{Human participants.} No human participant was involved. The human arms were designed, with a consent protocol and an ethics plan, and not executed (amendment A-002); no human-performance figure is used, and every human cost in the paper is an assumption labelled as such.

\paragraph{Spending.} Total API spend was 261.89 US dollars against an envelope of 800, read from the client's cost reports and recorded in a ledger per run; the 150 pilot worker sessions account for 124.43 dollars, the judges for about 0.23 dollars per run.
% docs/PHASE_5_AUDIT.md §0, §4; analysis/derived/estimates_pilot.json total_cost_usd

\paragraph{The benchmark is self-authored.} All scenarios, oracles, rubrics and validation reports were produced by agents of the same family as the workers, under a specification written by the controller session. The circularity this creates is discussed in Section~\ref{sec:implications} and is not resolved in this version.

\paragraph{Data and code availability.} The repository contains the orchestration and evaluation code, the public scenario bundles, the preregistration with its freeze manifest and amendment log, the raw event streams of every session, the validated Parquet datasets, the analysis scripts and the tables and figures they produce. The hidden oracles are kept outside the repository, on the controller only, and are not published; the hash of every oracle is recorded in the freeze manifest so that any change after the freeze is detectable. The raw event streams are excluded from the public repository by its ignore rules and retained on the controller. API credentials are never stored in the repository.

\paragraph{Conflicts of interest.} The agents that produced this study and the models they evaluate are products of the same vendor, which is also the provider of the API through which the study was paid for; no other relationship, funding or interest is declared. The sub-sessions that wrote the sections reporting results had access to the analysis outputs and not to the runs; the sections describing the theory, the benchmark, the design and the infrastructure were written from the protocol and the code, without access to any run.