\paragraph{At equal output budget.}\label{sec:swe-equal} The DGM's initial and reported best agents allow at
most 4{,}096 output tokens in each tool-calling request, and the textbook agent allows 16{,}000. We raised this limit of
both DGM agents to 16{,}000 and changed nothing else (other model calls, such as the best agent's summaries of its history, keep 4{,}096 in all three agents), and in the third round ran the
three agents in one batch on $\wN$ of the fresh tasks, drawn at random, with half the earlier parallelism. The textbook
agent resolves $\wMovesRate\%$ of the tasks at \$\wMovesUsd{} per task run, the initial agent resolves $\wInitRate\%$
at \$\wInitUsd{}, and the best agent resolves $\wBestRate\%$ at \$\wBestUsd{}. The textbook agent leads the best agent by
$\wMovesBestAbs$ points [$\wMovesBestAbslo$, $\wMovesBestAbshi$] (McNemar $p=\wMovesBestP$), so the registered
non-inferiority test passes and the lead is significant, and the best agent is $\wBestInitAbs$ points
[$\wBestInitAbslo$, $\wBestInitAbshi$] below its own initial agent ($p=\wBestInitP$). The initial agent is
$\wMovesInitAbs$ points [$\wMovesInitAbslo$, $\wMovesInitAbshi$] below the textbook agent, which is not significant
($p=\wMovesInitP$).

The best agent still ends \wBestEarly{} of its \wN{} runs early, compared with \wInitEarly{} for the initial agent and
\wMovesEarly{} for the textbook agent, although the machine was far less loaded than in the earlier batch and memory
pressure does not explain the endings (Appendix~\ref{app:dev}, items 13 and 14). With the patches of early-ended runs
kept, the textbook agent still leads the best agent by $\wsMovesBestAbs$ points [$\wsMovesBestAbslo$,
$\wsMovesBestAbshi$], and the best agent is $\wsBestInitAbs$ points [$\wsBestInitAbslo$, $\wsBestInitAbshi$] below its
initial agent, which is not significant ($p=\wsBestInitP$).