\begin{table}[t]
\caption{Assumption-aware method-selection guide.}\label{tab:method_selection_guide}
\tiny
\setlength{\tabcolsep}{2pt}
\renewcommand{\arraystretch}{0.95}
\begin{tabular}{@{}L{0.17\textwidth}L{0.22\textwidth}L{0.22\textwidth}L{0.19\textwidth}L{0.12\textwidth}@{}}
\toprule
Diagnosis & Observable evidence to check & Pipeline emphasis & Metrics and reporting & Failure if ignored \\
\midrule
Closed-set class-conditional noise & Credible noise-rate prior, clean anchors, or validation labels; fixed label space & Transition correction, robust losses, and calibrated reweighting & State noise-rate/anchor assumptions; report sensitivity to $T$ & Incorrect transition priors dominate \\
Instance-dependent or boundary noise & Loss trajectories, feature neighborhoods, repeated labels, or clean validation signals & Time-aware isolation plus target refinement and geometry checks & Report detection precision/recall and hard-clean retention & Hard or tail samples are removed as noise \\
Open-set or open-world contamination & OOD scores, prototype distance, unknown-class evidence, or mixed-label-space metadata & Membership detection before label correction; multi-prototype separation & Report clean/ID-noisy/OOD separation, AUROC, and rejection behavior & Unknown samples are forced into known classes \\
Fine-grained or long-tailed data & Class hierarchy, class frequency, semantic similarity, and tail-class loss patterns & Class-aware selection, contrastive/prototype preservation, and calibration & Report per-class, tail, and ambiguity-sensitive metrics & Subtle distinctions are erased \\
Weak, LLM-generated, or preference labels & Prompt/model provenance, repeated agreement, annotator reliability, or preference uncertainty & Process-specific uncertainty modeling and selective refinement & Report provenance, calibration, and uncertainty over targets & Pseudo-labels become overconfident \\
Graph, federated, or foundation-model settings & Client-level reliability, neighborhood propagation, pre-training source, or domain shift & Multi-level reliability: sample, client, neighborhood, model prior, and task & Report client/graph/transfer metrics and pipeline ablations & Noise is confused with structure or shift \\
\botrule
\end{tabular}
\renewcommand{\arraystretch}{1}
\end{table}