Online misinformation research in NLP has expanded rapidly, including approaches that model affective signals such as sentiment, discrete emotions, and emotion dynamics. However, the Information Disorder framework distinguishes misinformation, disinformation, and malinformation along dimensions of intention, harm, and contextual dependence, which are rarely operationalised in current datasets, tasks, and evaluation protocols. We provide a systematic mapping of 82 studies at the intersection of Information Disorder and emotion-aware NLP (51 model papers, 7 dataset papers, 24 survey/theory papers). Across empirical works (58), veracity-centric supervision dominates (72.4% binary labels), while explicit intention and harm variables appear in only 1.7% each. Evaluation relies mostly on random splits (79.3%), limiting robustness to source and temporal shifts. Emotion is represented in 43.1% of model papers, mostly as static features, with emotion dynamics and audience emotion rare. Based on these findings, we propose an operational taxonomy aligned with Information Disorder and a benchmark blueprint specifying tasks, annotation variables, split strategies, and evaluation protocols to support theory-grounded, comparable progress.

Bias-detection benchmarks routinely report accuracies above 90%, but a high number can reflect where a sentence came from as much as whether it contains a stereotype or an unfair claim about a group. We treat this as an evaluation problem and propose an integrated auditing protocol that makes source confounding measurable: an LLM-assisted label audit with cross-family verification, pair-aware splitting with role tracking, source-aware evaluation with a source-only baseline on the actual test split, counterfactual discrimination, and an artifact analysis of edited text. We apply it to a three-source corpus of 10,304 sentences that we audit and release. The audit reduces the global source-only majority baseline from 95.0% to 55.4%, and the same protocol shows that the real-only test split still carries a 77.6% source-only baseline, so confounding is reduced and measured but far from eliminated. Against that baseline, fine-tuned BERT and GPT-2 gain roughly 18 pp and outperform attention-derived models; four augmentation variants that look ordered in the mean fall within seed variance under paired testing; and a real-vs-edited probe separates Gemini counterfactuals from real text, although the signal disappears once the gold label is held fixed and has no measurable accuracy effect. Applied to CrowS-Pairs, the label audit flags 33.2% of the supposedly neutral members as still containing a group-level stereotype. The main contribution is the protocol, which exposes confounding that standard pipelines would report as results.