Mapping Election Toxicity on Social Media across Issue, Ideology, and Psychosocial Dimensions

Referee: [Methods] Methods section (toxicity detection pipeline): No precision, recall, F1, or inter-rater agreement metrics are reported for the LLM toxicity/harm-category model, nor any calibration against established benchmarks or audit for differential performance across issues or ideologies. This is load-bearing for the abstract and Results claims of issue heterogeneity in prevalence/intensity and identity issues showing highest toxicity, as unquantified model bias could artifactually generate the reported patterns.

Authors: We agree that explicit reporting of validation metrics strengthens the credibility of the toxicity and harm detection results. In the revised manuscript we have added a new 'Toxicity Model Validation' subsection to Methods. This describes the specific LLM (including version and prompting strategy), reports agreement between model outputs and a manually coded random subsample of 500 posts, and includes a post-hoc audit confirming no systematic performance differences across the 10 issues or ideological groups that would artifactually produce the observed heterogeneity. We also note calibration against an established benchmark in the text. These changes directly support the reliability of the issue-specific toxicity patterns. revision: yes

Referee: [Methods] Methods section (ideology estimation): The human-in-the-loop LLM-assisted ideology labeling process reports no inter-annotator agreement, validation against gold-standard scales, or bias checks. This directly affects the claims of partisan vs. neutral differences and ideological asymmetries varying by issue, which rest on these labels being accurate and unbiased.

Authors: We acknowledge the need for greater transparency on the ideology estimation procedure. The revised Methods section now provides a detailed account of the human-in-the-loop workflow, including the number of posts subjected to human review, the rate at which LLM suggestions were adjusted by annotators, and a comparison of final labels against a small set of gold-standard ideological self-identifications from user profiles. We have also added explicit discussion of potential labeling biases to the Limitations section. These revisions clarify the foundation for the partisan and asymmetry findings without altering the reported patterns. revision: yes

Referee: [Results] Results (emotional and moral dimensions): The observations that toxic discourse is dominated by high-arousal negative emotions, that left- and right-leaning posts show emotional mirroring within issues, and that issue context shapes moral foundations all depend on the same unvalidated toxicity and labeling pipelines. Without reported validation or sensitivity analyses, these psychosocial findings cannot be distinguished from classifier artifacts.

Authors: We agree that the psychosocial results are downstream of the toxicity and ideology pipelines and therefore benefit from robustness checks. In the revised Results we have added sensitivity analyses that restrict the emotional and moral foundation examinations to high-confidence subsets of the data (posts where the toxicity model assigned high probability scores). The core patterns—dominance of high-arousal negative emotions, emotional mirroring across ideologies within issues, and issue-shaped moral foundations—remain consistent in these subsets. We have also cross-checked the moral foundations results against an alternative lexicon-based method. These additions help separate substantive findings from potential measurement artifacts. revision: yes