\begin{table}[t]\centering\footnotesize
\caption{Pre-registered hypotheses of the third round (H11--H15), the estimate with its 95\% interval and the
verdict. Every contrast is within one batch.}\label{tab:hyp3}
\setlength{\tabcolsep}{3pt}
\begin{tabular}{>{\raggedright\arraybackslash}p{0.30\linewidth}>{\raggedright\arraybackslash}p{0.47\linewidth}>{\raggedright\arraybackslash}p{0.16\linewidth}}\toprule
hypothesis (as registered) & estimate & verdict \\ \midrule
H11 at equal output budget, seed $+$ textbook moves $\geq$ DGM best agent $-3$ (SWE-bench) & as run
$\wMovesBest$ [$\wMovesBestlo$, $\wMovesBesthi$]; early-ended patches kept $\wsMovesBest$ [$\wsMovesBestlo$,
$\wsMovesBesthi$] & supported (also superior) \\
H12 at equal output budget, DGM best agent vs its initial agent (expected: no significant gain) & as run $\wBestInit$
[$\wBestInitlo$, $\wBestInithi$]; patches kept $\wsBestInit$ [$\wsBestInitlo$, $\wsBestInithi$] & no gain; worse as
run \\
H13 (descriptive) textbook agent vs initial agent with the raised limit & $\wMovesInit$ [$\wMovesInitlo$,
$\wMovesInithi$] & -- \\
H14 clean random-diagnosis loops $\geq$ all DGM loops $-3$ (mean fresh gain) & $\rcHfourteen$ [$\rcHfourteenlo$,
$\rcHfourteenhi$] & \rcHfourteenVerdict \\
H15 log-free redesigns: $\geq 4$ of 5 raise the cap; mean gain $\geq$ log-based $-3$ & \lfRaised{} of 5;
$\lfDiff$ [$\lfDifflo$, $\lfDiffhi$] & supported \\ \bottomrule
\end{tabular}\end{table}