\begin{table}[t]
\caption{Diagnostic evidence from the small-loss analysis and its role in the review argument.}\label{tab:smallloss_evidence_summary}
\small
\setlength{\tabcolsep}{3pt}
\begin{tabular}{@{}L{0.22\textwidth}L{0.30\textwidth}L{0.36\textwidth}@{}}
\toprule
Evidence & What it shows & Review implication \\
\midrule
PUM and LSVR diagnostics & Early loss separability is linked to functional perturbation and feature-spectrum structure & Small-loss should be treated as a representation-dependent proxy, not a direct label-correctness oracle \\
Toy structural intervention & Making noisy samples structurally easy increases clean/noisy loss overlap & The proxy can fail when noisy targets are coherent rather than random \\
SILN transition matrices & Structured semantic corruptions differ from uniform random flips & Benchmark design should test corruption structure, not only noise rate \\
Loss distribution and $\Delta/\sigma_0$ & SILN reduces normalized clean/noisy loss separation under CE and RoLR & Loss-only selectors face an information bottleneck when distributions overlap \\
GMM splitter accuracy & Clean samples in the overlap region are often assigned to the noisy component & Evaluations should report hard-clean retention and clean/noisy detection behavior \\
Confidence and architecture checks & The overlap persists beyond one posterior model or one training architecture & Stress tests should check proxy robustness across diagnostics and model choices \\
\botrule
\end{tabular}
\end{table}