{"id":"dbed4d95-760a-450e-a90b-3bbdc4eddbea","arxiv_id":"2608.04120","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Trump-related Bluesky discourse clusters into 14 themes, negative sentiment rose over time, and attention shifts often align with executive orders.","lead":"This paper analyzes 38.5 million Bluesky posts about Trump, using an AI-assisted pipeline to identify 14 discussion themes and 19 executive-order categories, and finds negative sentiment growing over time. It offers one of the largest public datasets of Bluesky political discourse and a reusable workflow for tracking how online attention reacts to policy events.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theme shares rest on only 34.6% of cleaned posts that clustered; unclustered posts may differ systematically, biasing every theme-level claim.","rationale":"The reader's weakest assumption identifies the same load-bearing issue: theme shares and temporal patterns are computed only from the 10.46M clustered posts, with no test of whether the remaining ~19.7M unclustered posts are missing at random. This is indeed the most consequential threat to the paper's central claims, because the three headline findings—theme distribution, theme-level sentiment, and issue-attention dynamics—all depend on the clustered subset. The overall sentiment trend (Figure 2) is computed on the full cleaned corpus and is therefore less vulnerable, but the theme-level claims are not. The paper includes some supporting evidence for the pipeline (silhouette scores, trustworthiness, human validation), which is commendable, but none of that addresses the selection into clusters. The proposed check is feasible because the dataset and code are public, and it directly quantifies how sensitive the headline percentages are to cluster coverage. Since this concern matches the reader's conditional verdict rather than overturning it, the verdict remains CONDITIONAL; no new objection rises to rejection because the authors may well be able to show robustness with additional analysis.","tokens_in":8733,"tokens_out":2132,"duration_ms":23514,"concrete_test":"Re-run the microtopic pipeline with relaxed HDBSCAN hyperparameters (e.g., lower min_cluster_size or higher epsilon) so that at least 80% of the 30.2M cleaned posts are assigned to clusters, then recompute the 14 theme shares in Figure 3 (left) and the monthly theme shares in Figure 4. If any of the 14 theme percentages shifts by more than 5 percentage points, or if theme rankings change, the clustered subset is not representative and the thematic conclusions are conditional on clusterability. If the shares are stable, the concern is substantially mitigated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.2 reports that 10,456,083 of roughly 30.2 million cleaned English posts were assigned to microtopics, i.e., about 34.6% coverage. All theme-level results—Figure 3's left-panel percentages, the per-theme sentiment composition, and Figure 4's monthly theme shares—are computed from this clustered subset only, with the remaining ~19.7 million posts treated as unclustered observations. The paper provides no comparison between clustered and unclustered posts on any observable feature, so the missingness could easily be non-ignorable: HDBSCAN's noise label is not random, and posts that fail to form dense clusters may be shorter, more generic, more pro-Trump, less event-focused, or drawn from different user populations. If unclustered posts differ systematically, every central descriptive claim about thematic composition and issue attention is biased, not just imprecise. This is the weakest load-bearing assumption because the paper's headline contributions (14 interpretable themes, persistent versus episodic attention, sentiment composition by theme) all pass through this subset. The paper's honest limitation statements do not address this selection step.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a large-scale computational analysis of Trump-related Bluesky posts, reporting 38.5 million collected posts from January 2019 to February 2026, reduced to 30.2 million cleaned English-language posts. The authors use an LLM-assisted clustering pipeline with human validation to identify 4,770 microtopics and 14 overarching discourse themes, and separately code 258 executive orders into 19 thematic categories. They report sentiment proportions from VADER (47.0% negative, 17.3% neutral, 35.7% positive), describe theme-level sentiment composition, analyze monthly theme dynamics, and apply structural break detection to identify shifts in attention. The core claims are that Bluesky discourse is simultaneously persistent and event-sensitive, and that negative sentiment grows over time, particularly after January 2025.","tokens_in":1338,"tokens_out":1397,"duration_ms":51491,"significance":"If the results hold, this would be a valuable descriptive resource for studying political discourse on an increasingly important platform, and the public release of the dataset and pipeline is a concrete contribution. The paper's strengths include the scale of the corpus, the explicit human validation of the thematic codebook, the transparent pipeline description, and the acknowledgment of several limitations. However, the central theme-level findings rest on a clustered subset covering only about one-third of the cleaned corpus, and the paper does not establish that this subset is representative. This selection issue is load-bearing for the theme distributions, theme-level sentiment composition, and structural break timing claims, so the current version requires substantive additional analysis rather than copyediting.","major_comments":[{"comment":"The theme distribution and all theme-level percentages are computed from the 10,456,083 posts assigned to microtopics, which is 34.6% of the 30.2 million cleaned posts. The paper offers no comparison between clustered and unclustered posts on length, sentiment, author characteristics, or temporal coverage, so the implicit assumption that unclustered posts are missing at random is unsupported. HDBSCAN's noise label is systematic rather than random, and the unclustered majority could plausibly be shorter, more generic, more or less event-focused, or drawn from different user populations. This would bias every theme share, the per-theme sentiment composition in Figure 3, and the monthly theme volumes in Figure 4. The authors should report coverage by temporal partition, compare clustered and unclustered posts on observable features, and provide a sensitivity analysis (for example, reweighting or keyword-based assignment of unclustered posts) to demonstrate that the 14-theme characterization is robust to the missing 65.4%.","section":"Section 4.2, Figures 3 and 4"},{"comment":"The structural break analysis constructs daily post-volume time series for each Bluesky theme from the same clustered subset. If the fraction of posts assigned to microtopics varies across time—which is plausible given the four temporal partitions and the platform's rapid growth—the estimated break dates may reflect changes in clustering coverage rather than changes in public discourse. For example, Figure 5's breaks in immigration discourse could coincide with periods when more posts happened to form dense clusters. The authors should plot the proportion of posts assigned to microtopics over time and rerun break detection on coverage-adjusted series, or on full-corpus series such as keyword-based volumes or sentiment totals, before attributing the breaks to executive orders and media coverage.","section":"Section 4.4, Figure 5"},{"comment":"The keyword lexicon used for retrieval includes 'Republican' and 'maga' without any precision or recall validation. Because the abstract and research questions describe the corpus as 'Trump-related,' posts retrieved solely on the keyword 'Republican' may concern other political races, non-Trump candidates, or generic partisan topics, which would blur both the thematic and sentiment results. A small hand-coded precision sample, reported per keyword, would establish that the retrieval definition matches the intended construct. This is a load-bearing validity issue for the dataset's interpretation as Trump-specific discourse.","section":"Section 3.1, keyword lexicon"}],"minor_comments":[{"comment":"The concession that VADER 'does not distinguish supportive from opposing sentiment' is only in the conclusion; it should also appear in Section 4.1 and be reflected in the abstract, since the claim of 'increasing proportion of negative sentiment' conflates criticism of Trump with negative affect more generally.","section":"Section 5, Conclusion"},{"comment":"The 17 seed codes were derived from a topic model of the microtopic summaries, which are themselves products of the same clustering pipeline. This creates an anchoring risk in the subsequent coding step; the paper should acknowledge this explicitly and describe how the human reviewers guarded against simply reproducing the seed structure.","section":"Section 3.1, Overarching Theme Generation"},{"comment":"The left panel's percentages should state their denominator explicitly (theme-assigned posts rather than all cleaned posts), and the caption should clarify that sentiment composition is also computed only for clustered posts.","section":"Figure 3"},{"comment":"The column labeled 'Part.' refers to temporal partitions A–D, but the caption does not define the date ranges of these partitions. Adding the ranges would make the table interpretable without consulting the supplementary material.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The principal concern is the unaddressed selection in the clustering step, which affects the paper's central descriptive claims. I would ask for a dedicated robustness section comparing clustered and unclustered posts and showing coverage over time. The paper is otherwise within scope for a computational social science or human-computer interaction venue, and the public dataset is a definite asset."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline result is real: a 38.5M-post Trump-related Bluesky corpus with a transparent LLM-assisted thematic pipeline, public code, and human validation. That is genuinely useful. The 14 Bluesky themes and 19 EO categories are a new empirical mapping, and the temporal analysis is presented as exploratory rather than over-claimed. The paper also states its own limits in the conclusion, which I weighed.\n\nThe stress-test concern lands. All theme shares, sentiment-by-theme compositions, and monthly attention figures come from the 10,456,083 posts that clustered into microtopics, roughly 35% of the cleaned 30.2M corpus. HDBSCAN assigns noise for reasons that are usually correlated with text properties, so the unclustered 65% could look very different. The paper gives no comparison. That does not sink a descriptive study, but it means the specific percentages in Figure 3 and the theme trends in Figure 4 are not established for the corpus as a whole. The fix is straightforward: compare clustered vs unclustered posts on date, length, sentiment, author features, and simple keyword signals, then re-weight or explicitly qualify. I would make that a required revision.\n\nTwo other points, in order of importance. VADER sentiment is, as the authors concede, unable to distinguish supportive from opposing speech, so 'sentiment' is doing more work than the measure supports. And the structural break detection is applied to raw daily counts without detrending, which is risky given the platform's massive growth; the immigration example in Figure 5 is suggestive, not authoritative. The seed-code anchoring is real but minor, since the main clustering and sentiment measures are external to that step.\n\nMy overall read: the artifact is solid, but several descriptive claims outrun the evidence. The dataset and pipeline deserve serious refereeing. I would accept for peer review with the cluster-coverage issue as the main condition. I would cite the paper for the dataset and pipeline description, and I would bring it to a methods-focused reading group to talk about what coverage means in clustering-based text analysis.","headline":"A genuinely useful large Bluesky dataset and mostly transparent pipeline, but the theme-level percentages rest on the clustered third of the corpus and the paper has not shown that third is representative.","tokens_in":633,"tokens_out":936,"would_cite":true,"duration_ms":27853,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that 30.2 million English-language Bluesky posts about Trump can be organized into 14 interpretable discourse themes, and that the same hybrid coding of 258 executive orders yields 19 thematic categories, with issue…","keywords":["Bluesky","executive orders","political discourse","sentiment analysis","topic modeling","large language models","computational thematic analysis","social media"],"falsifier":"Run the same embedding, clustering, and theme-assignment pipeline on a random sample of the roughly 19.7 million posts that were not assigned to microtopics; if those posts show a different sentiment split or different dominant topics than the 10.5 million clustered posts, the reported theme shares and sentiment trends would have to be revised.","tokens_in":8552,"feed_emoji":"📊","tokens_out":7536,"duration_ms":74336,"temperature":0.7,"pith_summary":"This paper tries to show that the English-language Trump-related conversation on Bluesky from January 2019 to February 2026 is not shapeless noise: after collecting 38.5 million posts and cleaning to 30.2 million, it identifies 14 interpretable discourse themes such as executive power, partisan conflict, national security, immigration, and media narratives. A parallel coding of all 258 executive orders from Trump's second term yields 19 thematic categories, letting the authors compare what the administration formalized in policy with what the public discussed online. The paper further claims that attention to these themes is both persistent and event-sensitive, with issue volumes shifting abruptly around executive actions, and that the share of negative sentiment grew over time, widening the negative-positive gap from about 2 percent in August 2024 to roughly 15 percent by January 2026. If true, the analysis offers one of the largest public corpora of Trump-related Bluesky posts and a reusable human-LLM pipeline for mapping online political discourse onto policy events.","feed_headline":"Trump-era Bluesky discourse falls into 14 themes","feed_subtitle":"A public 30.2-million-post dataset links executive orders to 19 categories and shows negative sentiment gaining.","key_machinery":"The load-bearing mechanism is the microtopic discovery and theme-generation pipeline: each post is embedded, dimension-reduced, density-clustered, and each resulting microtopic is titled and described by a large language model, then reviewed by human researchers. The microtopics are converted into initial codes by an LLM using a seed codebook, and researchers iteratively merge and refine these codes into the final 14 overarching themes. A separate but analogous hybrid coding, seeded by a policy-agendas codebook, maps the 258 executive orders into 19 categories. The structural break analysis of daily post-volume time series is the tool that turns the theme annotations into a temporal claim about event-sensitive attention.","core_discovery":"The central claim, stated on the paper's own terms, is that Trump-related discourse on Bluesky has a discoverable thematic anatomy: 4,770 fine-grained microtopic clusters produced by an LLM-assisted clustering pipeline, validated by human review, organize 10.5 million posts into 14 overarching themes, with executive power and partisan conflict the most prominent. The same hybrid coding of the 258 executive orders signed between January 20, 2025 and May 1, 2026 produces 19 categories, dominated by administrative task forces, border and crime enforcement, foreign policy, and national security. The paper also claims that monthly theme volumes reveal persistent core topics alongside episodic spikes, and that a structural break analysis of immigration discourse places breakpoints near executive orders such as the April 2025 criminal-alien order and the August 2025 D.C. crime-emergency order. Sentiment analysis with a lexicon-based classifier assigns 47 percent negative, 17 percent neutral, and 36 percent positive sentiment overall and documents a widening negative-positive gap over time.","pith_inferences":["The same microtopic pipeline could be applied to non-English posts or to other platforms, since the method itself is not language-bound and the authors explicitly chose to restrict to English.","The event-sensitivity finding suggests a predictive design: executive-order signing dates could be used to forecast abrupt increases in issue-specific attention on Bluesky, though the paper only describes the relationship retrospectively.","The roughly 19.7 million unclustered posts could be analyzed with a cheaper classifier to test whether the 10.5 million clustered posts are representative; if they differ, every theme-share estimate would need revision.","Applying structural break detection to individual microtopics, rather than only to grouped themes, might pinpoint which specific events trigger attention shifts at a finer granularity than the paper reports."],"forward_implications":["The public 30.2-million-post corpus and 14-theme annotation let other researchers track issue attention on Bluesky without recollecting the data from scratch.","Monthly theme volumes can be aligned with future executive actions to detect attention spikes, making the approach a template for real-time policy-discourse monitoring.","The 19 executive-order categories allow direct comparisons between policy emphasis and public discussion emphasis, revealing which policy areas receive disproportionate online attention.","The structural break result suggests immigration discourse is episodic, switching abruptly around discrete policy events rather than drifting gradually.","The widening negative sentiment gap, from about 2 percent in August 2024 to 15 percent by January 2026, indicates that critical reactions became increasingly dominant in Trump-related Bluesky conversation."],"supporting_citations":[{"why":"Supplies the microtopic discovery approach and the prior characterization of Bluesky as politically polarized, which the pipeline builds on.","marker":"[11]"},{"why":"Supplies the thematic-analysis principles that structure the human-AI code grouping into overarching themes.","marker":"[4]"},{"why":"Establishes the LLM-assisted thematic analysis procedure with human validation that the paper extends to this corpus.","marker":"[12]"},{"why":"Documents Bluesky's polarized and left-leaning user composition, which the sentiment interpretation relies on.","marker":"[10]"},{"why":"Shows that Trump-related social-media opinion tracked national polling aggregates, motivating the temporal attention analysis.","marker":"[3]"},{"why":"Demonstrates LLM-based analysis of 2024 U.S. election discourse on Telegram responding to major events, supporting the event-driven interpretation.","marker":"[9]"}],"fun_headline_variants":["38.5M Bluesky posts reveal 14 Trump themes, 19 EO categories","Bluesky's Trump-era chatter: 14 themes, 19 EO topics","LLM clustering maps 38.5M Bluesky posts to 14 themes","Negative sentiment grows in Bluesky Trump discourse, data show","Bluesky echoes: 14 Trump themes, 19 EO categories, rising negativity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The analysis treats the 10.5 million posts that clustering assigned to identifiable microtopics as standing in for the full 30.2 million-post corpus, yet it never checks whether the roughly 19.7 million unclustered posts differ systematically.","fun_headline_variants_meta":{"raw":{"variants":["38.5M Bluesky posts reveal 14 Trump themes, 19 EO categories","Bluesky's Trump-era chatter: 14 themes, 19 EO topics","LLM clustering maps 38.5M Bluesky posts to 14 themes","Negative sentiment grows in Bluesky Trump discourse, data show","Bluesky echoes: 14 Trump themes, 19 EO categories, rising negativity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000762,"raw_usage":{"total_tokens":3397,"prompt_tokens":972,"completion_tokens":2425,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":588,"completion_tokens_details":{"reasoning_tokens":2322}},"tokens_in":588,"tokens_out":2425,"duration_ms":17122,"temperature":1.0,"reasoning_tokens":2322,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T00:28:12.462214+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same embedding, clustering, and theme-assignment pipeline on a random sample of the roughly 19.7 million posts that were not assigned to microtopics; if those posts show a different sentiment split or different dominant topics than the 10.5 million clustered posts, the reported theme shares and sentiment trends would have to be revised.","supporting_citations":[{"cited_title":"Qualita- tive research in psychology3(2), 77–101 (2006).https://doi.org/10.1191/ 1478088706qp063oa","cited_arxiv_id":null,"evidence_quote":"Supplies the thematic-analysis principles that structure the human-AI code grouping into overarching themes."},{"cited_title":"Scientific reports 8(1), 8673 (2018).https://doi.org/10.1038/s41598-018-26951-y","cited_arxiv_id":null,"evidence_quote":"Shows that Trump-related social-media opinion tracked national polling aggregates, motivating the temporal attention analysis."},{"cited_title":"Social Network Analy- sisandMining15(1), 91(2025).https://doi.org/10.1007/s13278-025-01504-0","cited_arxiv_id":null,"evidence_quote":"Demonstrates LLM-based analysis of 2024 U.S. election discourse on Telegram responding to major events, supporting the event-driven interpretation."}],"review_version":1}