\begin{table}[t]
\caption{Training-dynamics diagnostics for noisy supervision.}\label{tab:dynamics_diagnostics}
\small
\setlength{\tabcolsep}{3pt}
\begin{tabular}{@{}L{0.18\textwidth}L{0.27\textwidth}L{0.25\textwidth}L{0.20\textwidth}@{}}
\toprule
Diagnostic & What it can reveal & Typical false signal & Useful reporting target \\
\midrule
Loss trajectory & Whether an example becomes easy before or after memorization & Tail, boundary, or rare clean samples may remain high-loss & Class-aware clean/noisy separability over epochs \\
Forgetting or transition events & Whether predictions are unstable under repeated training updates & Ambiguous examples may be unstable without being mislabeled & Frequency and timing of clean-to-noisy role changes \\
Neighborhood consistency & Whether local feature geometry supports the observed label & Biased representations can make wrong neighborhoods look coherent & kNN purity, reverse-neighbor support, and hard-clean retention \\
Prototype or centroid drift & Whether class anchors move toward corrupted supervision & Long-tailed classes can drift because of sparse support, not only noise & Per-class prototype stability and tail-class behavior \\
Gradient agreement & Whether samples reinforce or conflict with shared optimization directions & Easy mislabeled examples can agree with dominant shortcuts & Agreement by class, difficulty, and suspected-noise group \\
Calibration or entropy drift & Whether confidence tracks uncertainty rather than memorized labels & Overconfident pseudo-labels can mimic reliable targets & Expected calibration error, coverage, and uncertainty on suspected noise \\
\botrule
\end{tabular}
\end{table}