1\section{End-to-end transfer and visual case studies}2\label{sec:transfer}34\subsection{Three held-out mini-projects}56The transfer study runs three bounded research projects that are larger than the benchmark tasks and were authored after the benchmark suite by a separate sub-session, with a larger envelope (12.00 USD, 60 minutes, 150 turns per session): T1, a real-time forecasting project on US retail-sales growth from ten ALFRED vintages with release-lagged predictors, three baselines, a Diebold--Mariano test and a short paper; T2, a simulation-methodology project comparing three cluster-robust confidence intervals for a mean across a $3 \times 3$ grid of cluster counts and intra-class correlations with 2,000 replications, Monte Carlo errors and failure regimes; and T3, an evidence synthesis of the effect of generative-AI assistance on task-completion time from five CC~BY primary studies with heterogeneous outcome definitions, requiring harmonisation, pooling with heterogeneity statistics and an appropriately limited conclusion (two of five studies poolable; the honest synthesis is an interval that includes zero). Each project was run under every arm twice (\nTransferRuns{} runs). In the two arm-D runs of T1 the reviewer never ran, because the harness launches the reviewer only after the controller gate passes and the gate never passed (see below); for T1, arm D therefore collapsed to arm C. Table~\ref{tab:transfer} gives the case-level results.78\begin{table}[htbp]\centering\small9\caption{Transfer mini-projects: every run (project $\times$ arm $\times$ repeat) with status, scientific success, criteria passed, violations, fresh-environment reproducibility, inference cost and wall time.}10\label{tab:transfer}11\input{../results/tables/tab_transfer}12\end{table}1314All \nTransferEval{} runs finished within the envelope and \nTransferSuccess{} were scientifically successful; the exception is one arm-C run of T2 whose claims ledger asserted a universal statement about coverage from the simulation grid alone (an invalid-inference violation under the project's predeclared rule). As in the benchmark, the arms do not separate on outcomes; they separate on cost, from \transCostA{} USD per run under A to \transCostD{} under D, with median wall time \wallTransferMed{} minutes.1516\subsection{A natural gate failure, and what it shows}1718The transfer study produced the only natural controller gate failures of the whole study, and they are instructive because they were the gate's fault. The forecasting project's public gate specification enabled the generic \texttt{forecasts\_csv\_chronology} check, which requires every forecast's target date to be strictly later than its origin date. Under real-time vintages this is wrong: at a month-end origin the most recent published observation lags by one or two months, so the one-step-ahead target is a month that has already ended but whose value is not yet released, and its first-of-month date precedes the origin date. All eight T1 deliverables, in every arm, therefore ``failed'' this check on 40 of 80 rows. In arms A and B the failure was recorded post hoc and had no consequence. In arms C and D the executors had already run the gate themselves during their initial session, seen the failure, and written in their reports that passing it would require false dates, which they declined to write; the controller then returned the failure three times, but because of the launcher defect each returned turn carried one word, and the resumed sessions replied that the message was cut off and stopped (two or three turns, 0.01--0.03 USD each). The four C and D runs of T1 thus ended with \transTOneSessCD{} sessions each and a mean cost of \transTOneCostCD{} USD against \transTOneCostAB{} for A and B (the difference is the protocol session, not the loop), and all passed the hidden evaluator, whose chronology check accounts for the release lag. With the corrected launcher (post-fix runs RT0025 under arm C and RT0026 under arm D, Table~\ref{tab:postfix}), the executor received the full gate output three times and refused each time to alter the dates, stating that passing ``would require false dates in \texttt{out/forecasts.csv}'' while the other 64 checks passed; both runs passed the hidden evaluator, and in the arm-D run the reviewer again never ran because the gate never passed. This is false blocking by a mis-specified control (secondary question~4 of the mandate): the gate consumed budget and bought nothing, and the only reason it did no scientific harm is that the executor declined to ``fix'' correct output. It is also a reminder that an oracle-free runtime check is only as good as the contract it encodes; the generic check was adequate for the benchmark's final-vintage forecasting tasks and wrong for the real-time one.1920\subsection{Authentic visual case study}2122Figure~\ref{fig:case} assembles four captured surfaces from the screenshot manifest, taken from different runs: a native terminal window tailing the live event stream of a confirmatory executor session, a labelled stress-test rejection by the gate on a copy of a real deliverable, the evidence dashboard during the main benchmark, and the post-run page of a pilot run with an executor and a reviewer session. Each panel carries its capture mode, its UTC time and the identifiers it documents; the full gallery with hashes is in Appendix~\ref{app:gallery}. Two kinds of image are deliberately absent: no capture was synthesised or reconstructed, and the two Safari window-level captures that rendered only the browser chrome are listed in the manifest as failures. One post-hoc capture of the natural gate-failure run page could not be taken because the operating-system session was locked at the time; the receipts themselves are in the state store.2324\begin{figure}[htbp]\centering25\includegraphics[width=0.48\textwidth]{../evidence/public_screenshots/S007_native_terminal_live_RC0009.png}\hfill26\includegraphics[width=0.48\textwidth]{../evidence/public_screenshots/S009_native_terminal_gate_rejection_stress.png}\\[2pt]27\includegraphics[width=0.48\textwidth]{../evidence/public_screenshots/S008_dashboard_live_main.png}\hfill28\includegraphics[width=0.48\textwidth]{../evidence/public_screenshots/S006_viewer_run_RP0014.png}29\caption{Authentic captures (native \texttt{screencapture}; no synthetic imagery). Top left: Terminal window tailing the live stream of confirmatory run RC0009 (F2-conf-02, arm C) on M3U96a (S007, native\_live, 04:06 UTC). Top right: the gate program rejecting a copy of RP0006's deliverables after two defects were planted by the controller (S009, native\_live, labelled stress test). Bottom left: the evidence dashboard during the main benchmark, 38 runs evaluated and 7 running (S008, dashboard\_live, 04:06 UTC). Bottom right: run page of RP0014 (F6-dev-01, arm D) with its executor and reviewer sessions, gate receipts, review and evaluation receipt (S006, viewer\_postrun). Hashes and selection rules: Appendix~\ref{app:gallery}.}30\label{fig:case}31\end{figure}32