{"id":"8aa19903-9107-4293-b44d-58cc6563062d","arxiv_id":"2607.14924","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"SwimComposer enables authoring narrative visualization in motion for swimming videos via an automated data pipeline and a multi-view, timeline-based tool, evaluated with 9 participants.","lead":"SwimComposer is a new tool that lets content creators add moving data visualizations to swimming videos—like speed and leader gaps—to tell a race story from start to finish. It automatically prepares race data, learns narrative patterns from broadcast footage, and lets authors arrange views and transitions on a timeline.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pipeline accuracy unvalidated: if tracking/transcription errors propagate to live data and insights, authored videos may mislead viewers, undermining the effectiveness claim.","rationale":"The reader's weakest assumption—that the automated data-preparation pipeline produces sufficiently accurate structured data—is indeed the most load-bearing concern. The user study measures whether participants can author coherent narratives, but those narratives are only as truthful as the underlying data. If the pipeline's positions, speeds, gaps, or insights are wrong, then even a perfectly usable authoring tool could produce misleading content, directly contradicting the spirit of the central claim. The paper's own Sec. 7 acknowledges the risk of misleading information, yet never validates the data that feeds the tool.\n\nI considered alternative concerns: the small sample size (n=9), lack of a baseline condition, and reliance on participants from the authors' networks. These limit generalizability but do not invalidate a technology-probe contribution; a probe's purpose is to reveal benefits and challenges, not to prove superiority. The pipeline accuracy concern is more fundamental because it threatens the validity of every authored artifact in the study. If the data were visibly wrong, positive Likert ratings about 'visual storytelling' would be difficult to interpret—participants might be satisfied with a tool that produces polished but false stories.\n\nThe 98.15% coverage statistic is a taxonomy-completeness measure, not an accuracy measure. Coverage says that the predefined insight categories can label nearly all coder-identified key moments; it does not say that the LLM correctly extracts those events from commentary. WhisperX transcription errors, especially with names and accents, could shift event timestamps or attributes; GPT-5.4 could hallucinate or misattribute a lead change. Tracking through SAM3 in a bird's-eye video may be robust, but the paper provides no error bars or failure cases. Without such validation, the central claim remains conditional.\n\nThe proposed concrete test would settle the concern: manually annotate positions and key moments for a sample of races, compare against pipeline output, and compute quantitative error metrics. If the errors are small, the concern is resolved and the conditional acceptance can proceed. If the errors are large, the paper should report accuracy limitations and possibly re-analyze the user study with corrected data. Either way, this is a check the authors should run before claiming that SwimComposer enables effective and truthful narrative authoring.","tokens_in":20067,"tokens_out":3770,"duration_ms":36739,"concrete_test":"Use the study's race (women's 100 m butterfly final) plus 3–5 additional races from the validation set. Manually annotate swimmer positions every 5 frames from the bird's-eye video (or use known lane-center ground truth); compare with SAM3 tracking output and compute median/95th-percentile pixel error. Convert position error to derived speed and gap-to-leader error and compare with official 50 m split times. Separately, take the original commentary audio, have two independent annotators label key moments (lead change, record watch, result, etc.), and compare their labels with the pipeline's extracted insights, reporting precision/recall per insight type. If median speed error exceeds ~0.05 m/s or per-insight F1 falls below 0.8, the pipeline's data is not demonstrably truthful; the study results should be re-analyzed with corrected data or the claim qualified accordingly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that SwimComposer enables effective narrative authoring and storytelling—depends on the trustworthiness of the data participants put into their stories. That data comes entirely from the automatic multimodal pipeline in Sec. 3: swimmer trajectories and derived speed/acceleration/gaps from SAM3 tracking, event insights from WhisperX transcription plus GPT-5.4 extraction, and athlete/record metadata from official sources. The paper reports no ground-truth validation for any of these components. The only quantitative check (Sec. 3.3) is a 98.15% coverage rate of coder-identified key moments by the proposed insight types on 15 races; that measures taxonomy completeness, not whether the LLM correctly identified lead changes, record watches, or other events in each race. Nor is there any evaluation of tracking accuracy or of the derived live-data values. The user study (Sec. 6) used pipeline output for the women's 100 m butterfly final, so any tracking errors (e.g., lane misassignment, jitter-induced speed spikes) or mis-transcribed commentary directly propagate into participants' authored videos. If a participant displays a 'gap to leader' that is off by several tenths of a second, the resulting narrative is misleading—not merely imperfect in style. Sec. 7 even acknowledges that visualizations can 'convey misleading information' if not correctly coupled, but it does not test whether the pipeline data itself is correct. Thus the effectiveness claim rests on an unverified assumption that the pipeline produces sufficiently accurate structured data.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SwimComposer, a technology probe for authoring narrative visualizations in motion in swimming videos. It contributes an automatic multimodal pipeline (Sec. 3) that turns a race video, commentary audio, and a competition name into structured live data, event insights, athlete information, and records; an observational analysis of professional broadcasts (Sec. 4) that motivates views, transitions, and event-driven emphasis; and a user study (Sec. 6) with nine experienced content creators/designers who authored a full-race narrative video. The central claim is that SwimComposer generally enables effective narrative authoring and storytelling, supported by positive Likert ratings, consistent authoring patterns (Overview as backbone, Tracking/Comparison for key moments), and qualitative reports of reduced analysis effort and useful insights.","tokens_in":20386,"tokens_out":2633,"duration_ms":31392,"significance":"If the result holds, this is a useful step beyond prior work on embedding visualizations in sports videos: it treats authoring as coordinating data layers, views, transitions, and pacing over time, and it grounds the design in both an automated data pipeline and observed broadcast practices. The paper is unusually transparent: the study is pre-registered, questionnaires and authored videos are shared on OSF, pipeline source code and LLM prompts are released, and the authors explicitly acknowledge several limitations. These open-science practices strengthen the credibility of the qualitative evaluation. The main risk is that the truthfulness of the authored narratives depends on unvalidated pipeline accuracy; if tracking, transcription, or LLM extraction produce incorrect race facts, the resulting videos could mislead viewers, which would undermine the central effectiveness claim.","major_comments":[{"comment":"The pipeline is load-bearing for the central claim, but its accuracy is not validated against ground truth. Tracking, speed/gap derivation, WhisperX transcription, name normalization, and GPT-5.4 insight extraction are all used without quantitative evaluation. The only reported check (Sec. 3.3, Fig. 3) is a 98.15% coverage rate of coder-identified key moments by the proposed insight types; this measures taxonomy completeness, not whether the LLM correctly identified events in each race, and no inter-coder reliability is reported for the key-moment coding. Since the user study (Sec. 6) used pipeline output for the women's 100 m butterfly final, any tracking jitter or mis-transcription directly propagates into the authored videos. Please add validation against official results/split times and, at minimum, a manually labeled subset of positions and events, with precision/recall for insight","section":"Sec. 3, especially 3.2 and 3.3"},{"comment":"The study has 9 participants recruited through the authors' networks, no baseline or comparison condition, and the Likert analysis is purely descriptive. The paper's wording 'generally enables effective narrative authoring and storytelling' (Sec. 1) is stronger than what this design can support. I do not require a fully powered experiment for a technology probe, but the claim should be softened (e.g., 'was perceived to enable...') or complemented with a within-subject comparison to a simpler authoring tool. At minimum, report confidence intervals or effect sizes for the Likert items and be explicit that the observed 'consistent patterns' are qualitative.","section":"Sec. 6.1-6.4"},{"comment":"The Likert item presentation is hard to read: the counts are not shown per response option, and the reported percentages (e.g., 'Features Coverage 100%') appear to combine 4-7 ratings rather than the full distribution. Please provide the full response distribution per item (1-7) and a precise definition of the percentage. This is a reporting clarity issue, but it affects the interpretability of the central evidence.","section":"Sec. 6.4 / Fig. 5"}],"minor_comments":[{"comment":"The LLM is named as 'GPT-5.4'; check whether this is the intended model identifier and whether the exact version/temperature settings are captured in the released prompts. Model-specific behavior may matter for reproducibility.","section":"Sec. 3.3"},{"comment":"The public access URL '43.163.231.237/' is an IP address that may not be persistent. It would be safer to point readers to the OSF repository for a stable demo link.","section":"Sec. 5"},{"comment":"The age-range sentence is garbled ('18–54 3 5 10 00 1') and needs to be formatted as a proper distribution table.","section":"Sec. 6.2"},{"comment":"In the text, '[41] Chen et al.' and '[39] Chen et al.' do not match the reference entries, which list 'C. Zhu-Tian' as the first author. Please reconcile author names in citations and the bibliography.","section":"References"},{"comment":"The observational analysis covers only three broadcasts (one per sport) and is described as illustrative. This is fine, but the manual annotation reliability is not reported; a short paragraph acknowledging this and noting that no inferential claims are made would improve precision.","section":"Sec. 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is honest and well-positioned relative to prior work, and the open-science practices are a strong point. My main concern is the mismatch between the strength of the central claim and the validation depth: the pipeline accuracy is unvalidated, and the user study is small and non-comparative. This is fixable within the manuscript's scope by adding validation experiments or carefully narrowing the claims. I do not see a circularity problem beyond the taxonomy being developed and tested on the same kind of commentary; the coverage result is about taxonomy completeness, not correctness. The paper fits the journal's scope and, after revision, could be a solid contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know about this paper: it is a genuine integration advance, not a repackaging. Prior work on visualization in motion mostly embedded isolated overlays; this paper tackles full-race storytelling by coordinating data layers, views, transitions, and pacing over time. The automatic multimodal pipeline (video tracking, commentary transcription, LLM insight extraction, metadata retrieval) is a real engineering contribution, and the broadcast analysis gives design constraints rather than mere inspiration. The technology probe is evaluated as a probe, with all materials on OSF, source code released, and the study pre-registered. That is reproducible, honest work.\n\nWhat it does well: the framing as a technology probe is appropriate and the claims are scoped. The Likert data and qualitative coding show consistent authoring patterns, and the discussion of functionality/usability trade-offs is thoughtful. The observational analysis of broadcasts is modest but useful, and the pipeline's design is sensible.\n\nSoft spots, in proportion: the stress-test concern is valid. The pipeline's accuracy is not validated against ground truth. The 98.15% coverage rate measures taxonomy completeness on coder-identified key moments, not whether the LLM correctly extracted lead changes or record watches from commentary, and tracking/transcription accuracy is unreported. If a participant displays a gap that is off by several tenths, the authored story misleads. That matters, but it is not fatal for a technology probe — the central effectiveness claim is about authoring support, not data correctness. Still, it leaves a gap between the pipeline output and the narrative visuals. The user study is small (9 participants, mostly from the authors' networks) and has no baseline comparison; for a probe that is acceptable, but the paper would be stronger with at least a sanity check on the pipeline or a comparison against manual data preparation. The insight taxonomy being developed and validated on the same kind of commentary is a minor circularity, not a load-bearing flaw. The heavy reliance on the authors' own prior visualization designs is fine — they built on SwimFlow and say so clearly.\n\nWho this is for: people working on visualization in motion, sports video authoring, or data-video tools. It gives them a concrete system, a reusable pipeline, and an honest evaluation. It deserves peer review and will likely get useful revision comments rather than a desk reject. My recommendation: send it out, with the expectation that the authors address pipeline validation before acceptance.","headline":"Solid tech-probe contribution on narrative authoring for swimming videos; the main soft spot is the unvalidated data pipeline, but the paper is honest and deserves a serious referee.","tokens_in":20843,"tokens_out":1379,"would_cite":true,"duration_ms":18003,"reading_group":"yes","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 authoring narrative visualizations in motion is feasible through timeline-based coordination of views, transitions, data layers, and pacing, and that a nine-participant study supports this.","keywords":["narrative visualization","visualization in motion","sports video","storytelling","technology probe","swimming","data-driven video","authoring tool"],"falsifier":"A reader could process a known race through the published pipeline and compare the computed frame-level rankings, split times, and gap measurements against the official timing results for that race. If the pipeline's live data deviates systematically from official splits, or if the LLM-extracted insights fail to match a human-coded set of key moments, the central claim that the tool supports truthful narrative authoring would be falsified.","tokens_in":19980,"feed_emoji":"🏊","tokens_out":8742,"duration_ms":75343,"temperature":0.7,"pith_summary":"The paper asks how content creators can author narrative visualizations in motion — charts and data overlays embedded in video that move with the athletes they describe — and uses swimming races as the case study. The central claim is that this kind of storytelling is best supported not by simply placing charts on frames, but by giving creators the ability to coordinate views, transitions, data layers, and pacing over time. The paper builds an automatic pipeline that turns a race video, commentary audio, and a competition name into structured, visualization-ready data, and a technology probe called SwimComposer that exposes these elements as editable timeline segments. A study with nine experienced content creators and designers found that the probe enabled effective authoring with consistent patterns: overview as the narrative backbone, tracking and comparison for key moments, and an event-driven layering of athlete metadata, live data, and results. A sympathetic reader would care because this reframes a previously manual, skill-heavy task into a coordination problem that tools can support.","feed_headline":"Nine creators author full race narratives with SwimComposer","feed_subtitle":"A new tool turns raw swim footage and commentary into structured, timeline-edited data stories for general audiences.","key_machinery":"The central object is SwimComposer, a technology probe — a functional but intentionally minimal tool for studying how people author. It exposes four sets of controls: a layer library that organizes race data into live data, insights, athlete information, and records; a narrative viewing panel with three view modes (overview, tracking, comparison); timelines that arrange layer segments, view segments, and playback speed; and configuration panels for transitions and layer design. Carrying the argument is also the automatic multimodal data preparation pipeline that feeds the probe: video processing with lane partitioning and prompt-based detection and tracking to extract live metrics, audio pro","core_discovery":"The paper's central claim is that authoring narrative visualizations in motion can be supported by reframing the task as the coordination of views, transitions, data layers, and pacing over time. Drawing on an observational analysis of professional broadcasts in swimming, basketball, and soccer, the authors found a recurring alternation between overview and focus-on views, with transitions used to signal emphasis. They implemented this insight in SwimComposer, a technology probe that lets creators arrange data layers, view changes, transitions, and playback speed on timelines. In a pre-registered study with nine participants experienced in content creation or graphic design, all produced com","pith_inferences":["Because the view-alternation and event-driven emphasis patterns appeared in professional broadcasts across swimming, basketball, and soccer, the coordination model likely generalizes beyond swimming; a concrete test would be to implement the same timeline-based probe for a sport with less linear motion, such as soccer, and observe whether the same authoring patterns emerge.","The paper validates the insight category design against commentary coverage (98.15% of key moments across 15 races) but does not measure per-frame tracking accuracy or transcription reliability; a natural next step is to compare the pipeline's computed positions, speeds, and gaps against official timing data for the same races, which would test whether authored narratives rest on truthful data.","The study's nine participants and single task window make the strongest evidence the internal consistency of patterns rather than statistical power; a longer deployment with professional broadcast editors on real deadlines would reveal whether the authoring approach survives production pressures."],"forward_implications":["If SwimComposer's central claim holds, authoring narrative visualizations in motion becomes a coordination problem of arranging views, transitions, data layers, and pacing on a timeline, rather than a programming or manual animation task.","The consistent authoring patterns observed across participants suggest that race-based narratives share a temporal logic — overview as backbone, detailed views for key moments, event-driven data layering — that tool designers can build into future systems.","The automatic pipeline means new races can be turned into data stories from just a video, commentary audio, and a competition name, making the approach reproducible and transferable across events without per-race manual annotation.","Participants reported that the tool produced clearer results than they could achieve in general-purpose video editing tools, indicating that a specialized probe can fill a gap in current editing workflows for sports storytelling."],"fun_headline_variants":["SwimComposer lets creators edit data narratives into swim videos","Tool turns raw swim footage into timeline-edited data narratives","From swim video to data-driven narrative: SwimComposer","Nine creators author race narratives with SwimComposer's timelines","SwimComposer: authoring motion visualizations for swimming races"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The paper's argument rests on the automatic pipeline producing structured data — swimmer positions, speeds, gaps, and commentary-derived insights — that is accurate enough to support truthful narrative videos; the authors validate the insight category coverage (98.15% on 15 races) but do not measure per-frame tracking or transcription accuracy against ground truth.","fun_headline_variants_meta":{"raw":{"variants":["SwimComposer lets creators edit data narratives into swim videos","Tool turns raw swim footage into timeline-edited data narratives","From swim video to data-driven narrative: SwimComposer","Nine creators author race narratives with SwimComposer's timelines","SwimComposer: authoring motion visualizations for swimming races"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00099,"raw_usage":{"total_tokens":4033,"prompt_tokens":747,"completion_tokens":3286,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":491,"completion_tokens_details":{"reasoning_tokens":3202}},"tokens_in":491,"tokens_out":3286,"duration_ms":23963,"temperature":1.0,"reasoning_tokens":3202,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T00:39:42.611937+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could process a known race through the published pipeline and compare the computed frame-level rankings, split times, and gap measurements against the official timing results for that race. If the pipeline's live data deviates systematically from official splits, or if the LLM-extracted insights fail to match a human-coded set of key moments, the central claim that the tool supports truthful narrative authoring would be falsified.","supporting_citations":[],"review_version":1}