{"id":"3d6a94b6-9136-4fc7-8b46-2e5bb547a09e","arxiv_id":"2504.21393","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper describes JASMINE, a JavaScript web application that combines a hierarchical autoencoded overview with multiple linked detail views for exploring large multivariate astronomical datasets.","lead":"JASMINE is a browser-based prototype that lets astronomers inspect large datasets by opening multiple views of one object, such as galaxy images and rotating 3D particle clouds. It is a demonstration of a multimodal data-exploration concept rather than a new scientific result, and no code is currently released.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The overview navigation rests on Spherinator's latent morphology space being interpretable and preserved in the HiPSter sphere; the paper asserts this without measurement, so the central 'picking a data point' workflow is unestablished.","rationale":"The reader's weakest assumption identifies the Spherinator latent-space interpretability as the load-bearing premise. My reading agrees: the paper's own description in Section 3.1 makes the entire 'arranged by a key property' overview depend on this learned representation, and the paper supplies no new evidence or evaluation for it. I additionally note a second, closely related unverified step: the 3D latent space must be faithfully projected onto the 2D HiPS sphere used in the overview, and the zoom hierarchy must preserve neighbor relationships. Neither step is shown quantitatively. Because this is the same core concern raised by the reader, I do not propose a different verdict. The CONDITIONAL verdict remains appropriate: the idea is plausible and builds on published components (Aladin Lite, Illustris, Spherinator/HiPSter), but the central usability claim cannot be accepted as verified without a concrete test of the latent-space arrangement or a runnable demonstration. A quantitative neighborhood-preservation test would settle whether the concern lands, and the result would either support or undercut the central claim directly.","tokens_in":3871,"tokens_out":3769,"duration_ms":45827,"concrete_test":"Run the released Spherinator/HiPSter implementation on the same 1000 Illustris TNG mock images, then compute the 3D latent-space positions and the 2D spherical HiPS projection positions at several zoom depths. For a random sample of 100 galaxies, compute the Jaccard overlap between the 10 nearest neighbors in latent space and the 10 nearest neighbors in the projected overview. If the median overlap is not significantly above the overlap expected from random placement, the overview arrangement does not preserve morphological similarity and the central navigation claim fails. If no public implementation is available, request a live demo and perform the same neighborhood-overlap test on the displayed layout.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central navigation feature of JASMINE is the overview window from which users pick a data point. Section 3.1 asserts that Spherinator's latent representation is 'interpretable' and that 3D positions reflect morphological properties, and the paper then uses this arrangement as the structured view that 'enables picking a data point.' This is load-bearing because if the latent layout is not semantically meaningful, then the overview is an arbitrary scatter of galaxies and the claimed overview-first, details-on-demand workflow collapses. The paper provides no new evidence for this assumption, and it also skips a second necessary condition: even if the 3D latent positions are meaningful, the displayed spherical HiPS projection and its zoom hierarchy must preserve local neighborhood structure so that galaxies near each other in the overview are actually morphologically similar. Section 3.2 describes zooming as increasing the number of projected galaxies but gives no test of whether neighbor relations in the projected view correspond to neighbor relations in latent space. Without a quantitative check of either step, the central claim that JASMINE offers a useful structured route from global overview to individual data points is unsupported. The absence of a code repository or live demo further prevents independent verification, but the most load-bearing issue is the unverified latent-space arrangement that underpins the core workflow.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces JASMINE (JAvaScript Multimodal INformation Explorer), a client-side web application intended to support overview-first, details-on-demand exploration of large astronomical datasets. The proposed workflow uses an autoencoded latent representation of galaxy morphology, projected onto a hierarchical HiPS sphere via the authors' Spherinator/HiPSter framework, as the structured overview from which users select individual objects. Detail modals then display per-object data such as mock images and interactive 3D particle clouds from IllustrisTNG, with user-selectable data fields. The manuscript is a short prototype description: it states that the tool exists and describes its design and planned interaction model, but contains no screenshots, code, demo URL, or quantitative evaluation.","tokens_in":4197,"tokens_out":3963,"duration_ms":45899,"significance":"If the described prototype works as claimed and the latent-space overview is genuinely informative, JASMINE would be a useful contribution to interactive astronomical data exploration, linking global morphology-based navigation to rich per-object views. The paper has several strengths: it follows a well-established visualization paradigm, reuses existing open components (Aladin Lite, IllustrisTNG data, Spherinator/HiPSter), and is honest about its prototype status. However, the central claim that JASMINE 'allows users' to perform the described navigation rests entirely on textual assertion. No artifact is available for inspection, and the load-bearing premise that the autoencoded morphological layout is semantically meaningful in the displayed spherical projection is not supported by any evidence in this manuscript. The significance is therefore conditional on verification that is currently absent.","major_comments":[{"comment":"The central navigation feature of JASMINE is the structured overview from which users pick data points, and that structure depends on the claim in Section 3.1 that 'The learned latent representation of galaxy morphological space is interpretable' and that 'The 3D positions in latent space reflect the morphological properties of each galaxy.' This is load-bearing: if the latent layout is arbitrary, the overview is an unstructured scatter and the overview-first, details-on-demand workflow collapses. The manuscript provides no evidence for this claim, either new measurements or a pointer to a concrete evaluation in the prior Spherinator/HiPSter work. Section 3.2 adds a second unverified step: even if latent positions are meaningful, the HiPS projection and its zoom hierarchy must preserve local neighborhood structure so that nearby galaxies in the displayed sphere are morphologically similar, but the paper only says that zooming increases the number of projected galaxies. Please provide a quantitative or visual validation of both steps, or explicitly scope the prototype as based on an unverified representation.","section":"Section 3.1"},{"comment":"The paper's central claim is that JASMINE is a working multimodal data explorer, yet the manuscript contains no screenshot of the running application, no code repository URL, no live demo link, and no description of any functional test or user evaluation. For a software/instrumentation paper, the existence of the tool is the primary factual claim and must be verifiable. Please include at least one figure showing the overview window and detail modals, a persistent link to the prototype or source code, and a brief list of tested interactions (e.g., clicking a cell, opening multiple modals, rendering 3D particle clouds, toggling color fields).","section":"Abstract and Sections 3.2–3.3"},{"comment":"The multimodal data-cube format is a core concept of the paper, but its implementation is not described in enough detail to assess feasibility. The paper states that each data point is stored in a data cube with sides and subsides, and that 'The chosen side has to of the lowest hierarchy' (the sentence is incomplete), but it does not specify how the cube is represented (e.g., JSON schema, nested arrays), how the different data types (images, point clouds, metadata) are attached to sides, or how the on-the-fly HiPS tile generation accesses these structures. Without code or a more precise specification, the reader cannot determine whether the multimodal comparison workflow is actually implemented as described. Please clarify the data model and the mapping between cube sides and renderers.","section":"Sections 2.2 and 3.1"}],"minor_comments":[{"comment":"The sentence 'Zooming in increases the number of HealPiX cells and therefore the number of galaxies projected to the sphere surface' is ambiguous: for a fixed sample of 1000 galaxies, zooming changes the tiling resolution, not the number of distinct objects; please rephrase to describe how the same objects are shown at higher resolution or how additional objects become visible.","section":"Section 3.2"},{"comment":"The sentence 'each field as a different color scale' appears to be missing a verb; it should read 'each field has a different color scale.'","section":"Section 3.3"},{"comment":"The text says the simulations follow 'multiple thousands galaxies' but the subsample used is 1000 galaxies; please correct the phrasing to 'several thousand galaxies' or similar.","section":"Section 3.1"},{"comment":"The sentence referring to Faaique (2023) says the issue is 'still under review,' but the reference list gives a published article; please clarify the intended meaning and update the reference or phrasing.","section":"Introduction"},{"comment":"The statement that dimensionality reduction and clustering techniques create representations where 'The relations between the representation and the real data are preserved' is too general; this is not guaranteed for all such methods. Please qualify it as dependent on the technique and its training objective.","section":"Section 2.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads more like an ADASS conference demo note than a complete journal paper. The core idea is reasonable and the reuse of established components is a strength, but the lack of any artifact (screenshots, repo, demo) and the unverified latent-space navigation premise are substantial barriers. I would not reject outright—the issues are addressable by adding evidence and tightening claims—but the current version does not yet support the central functionality claim. I also note that the authors cite their own prior work for the interpretability of the latent space without summarizing or reproducing any validation, which is a key gap."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"JASMINE is a short software-prototype paper, and it should be read as one. What is new is the integration: taking an autoencoded spherical overview (Spherinator/HiPSter), Aladin Lite as the zoomable layer, per-object modals with mock images and interactive TNG particle clouds, and organizing fields in nested data cubes. The nested-cube idea is a genuinely useful organizing device. The paper is also honest about scope: it calls itself a prototype, notes that a window-management strategy is not implemented, and lists bidirectional communication and generalized interfaces as future work. For a six-page ADASS-style description, that is appropriate.\n\nThe soft spots are the usual ones for prototype papers. There are no screenshots, no demo URL, no code repository, and no usability or performance evaluation. The central claim that JASMINE works is therefore supported only by the text. That is a real limitation, but it is addressable and not fatal. The stress-test note worries that the overview arrangement is load-bearing and unverified because Spherinator's latent space is asserted to be interpretable. I think that concern is somewhat overdrawn: the paper does not claim to validate the morphology layout; it reuses a framework from the same group and uses it as the mechanism for structured browsing. For a prototype report, depending on a published component is acceptable. What would upgrade this paper is a live demo or a few screenshots showing the layout and modals, plus an explicit sentence about whether the latent-space arrangement was visually inspected for the 1000-galaxy sample.\n\nThe citation pattern looks appropriate: Spherinator/HiPSter, UltraPINK, Aladin Lite, Illustris TNG, and the standard visualization references are all relevant. Self-citation here is component reuse, not inflation. There are no free parameters or quantitative claims, so the paper is internally consistent; the math and data are not really in play, only software credibility.\n\nThis paper is for people building astronomy visualization or exploration tools, or for reviewers tracking the HiPSter ecosystem. It deserves a serious referee, not a desk reject, but the referee should ask for a demo or code and probably a brief usability note before acceptance. I would not cite it in my own work until the code is released.","headline":"A clear, honest prototype description; the nested data-cube plus linked-modals design is the real contribution, but the paper needs screenshots or a demo before the 'it works' claim can be verified.","tokens_in":4633,"tokens_out":2551,"would_cite":false,"duration_ms":27018,"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":"JASMINE is a web browser that lets astronomers move from an autoencoder-arranged overview of a galaxy set to individual objects shown as images or interactive 3D particle clouds.","keywords":["JASMINE","multimodal data exploration","visualization","autoencoding","latent space","point clouds","galaxy morphology","interactive web application"],"falsifier":"Take the same galaxy sample, sort all pairs by distance in the autoencoded latent space, and compare that ranking with independent human or classifier judgments of visual similarity; if nearby pairs are no more similar than random pairs, the overview's arrangement is arbitrary and the explorer's core navigation claim fails.","tokens_in":3638,"feed_emoji":"🔭","tokens_out":11059,"duration_ms":100066,"temperature":0.7,"pith_summary":"JASMINE is a prototype web application whose central claim is that a multimodal data explorer can carry an astronomer from a global overview of a very large dataset down to individual objects, with distinct views of the same object opened side by side. The overview is arranged automatically by a key property learned by an autoencoder, so similar objects appear near one another; selecting an object opens detail modals that can show an image or an interactive point cloud, with the user choosing which stored field to display. This gives astronomers a way to inspect massive, many-sided datasets without forcing every property into one overloaded visualization, and it makes the same overview-to-detail pattern reusable across survey and simulation data. The paper demonstrates the concept on a sample of simulated galaxies, with the arranged overview and the per-object views drawn from the simulation's own fields.","feed_headline":"JASMINE bridges a survey overview and one galaxy's 3D views","feed_subtitle":"Click a spot on a learned morphology map; inspect the same object as image, gas, or dark-matter particles.","key_machinery":"The load-bearing mechanism is a three-part pipeline. First, an autoencoder—a neural network that learns a compact latent representation of galaxy morphology—supplies the key property that orders the overview, and the paper asserts that this latent space is interpretable. Second, a hierarchical spherical projection tiles that latent space so the overview can be zoomed smoothly from the whole set down to a cell containing one or a few objects. Third, a nested data cube stores every field of a data point as an accessible side, and a modal viewer displays one selected side at a time. The modal windows communicate one-way from the overview to the details, which keeps the interaction model simple and predictable.","core_discovery":"The paper's claim is that the full arc of data inspection—overview, zoom, filter, details-on-demand—can be realized in a single browser-based tool by separating the arrangement of the dataset from the rendering of individual points. The arrangement comes from an autoencoded latent space: a neural-network projector learns compact morphological features from galaxy images, and those features are projected as an interactive hierarchical spherical tiling so that position on the sphere encodes morphology. Each data point is stored as a nested data cube, with one side per field and subsides for subfields, and the detail window shows whichever side the user selects. The prototype implements this for simulated galaxies, linking the overview to mock images and to color-coded particle clouds for gas, stars, or dark matter, together with access to the objects' simulation metadata. The paper states that the learned latent representation is interpretable and that 3D positions in it reflect morphology; that property is what turns the overview into a meaningful navigation surface.","pith_inferences":["Beyond the simulated-galaxy case, the modal-and-cube architecture is generic: any survey or simulation with a learned embedding could feed the same explorer, so the contribution is a viewer pattern rather than a specific map.","The true test of the arrangement is user behavior: if nearby points in the latent-space overview are not judged similar by astronomers, the navigation is decorative, and a controlled study comparing search times against a random grid would settle this.","Bidirectional links—selecting an object in one detail modal and highlighting its neighbors in the overview—would turn the current one-way browser into a clustering tool, though the paper only lists this as future work.","Because the paper supplies a prototype, the code itself can be stress-tested against browser scale: loading hundreds of thousands of points through the same hierarchical tiling would show whether the approach survives outside the 1,000-galaxy demo."],"forward_implications":["A user can start from a global morphological map, zoom to a cell, and open one or several detail windows for the same or different galaxies, comparing image and particle views side by side.","Adding a new measured property to a data point requires adding a side to the cube rather than redesigning the viewers, so the explorer can grow with the dataset.","The same overview-to-detail pattern applies to any large dataset for which a meaningful key property can be learned and per-object views can be rendered.","Because communication is one-way from overview to detail, the behavior is predictable; the paper explicitly identifies window management and bidirectional communication as the next improvements.","The overview itself becomes a scientific surface: navigating the latent space is navigating galaxy morphology, not just a list of files."],"supporting_citations":[{"why":"Supplies the 'overview first, zoom and filter, details-on-demand' mantra that structures the entire data-browser workflow.","marker":"Shneiderman 2003"},{"why":"Provides the taxonomy and strategies for multiple-view browsers, including the window-management trade-offs the prototype consciously defers.","marker":"Plaisant et al. (2003)"},{"why":"Supplies the argument for machine-learning representations as a basis for interactive overviews of large datasets.","marker":"Kollasch & Polsterer (2022)"},{"why":"Provides the autoencoder and hierarchical spherical tiling that produce the interpretable latent arrangement of galaxy images.","marker":"Polsterer et al. (2024)"},{"why":"Supplies the cosmological simulation snapshots, particle fields, and mock images used as the demonstration dataset.","marker":"Nelson et al. (2018)"},{"why":"Provides the interactive sky-viewer component that handles zooming, hierarchical sphere rendering, and coordinate overlays in the overview window.","marker":"Baumann et al. (2022)"}],"fun_headline_variants":["JASMINE lets you fly from galaxy maps into 3D particle clouds","One tool, many views: JASMINE maps galaxies to your browser","JASMINE's autoencoder turns galaxy images into an explorable sphere","Explore galaxies: JASMINE links images, particles, and metadata"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole navigation is only meaningful if the autoencoder's latent space really places morphologically similar galaxies near each other, and the paper asserts this interpretability without adding a new test of it.","fun_headline_variants_meta":{"raw":{"variants":["JASMINE lets you fly from galaxy maps into 3D particle clouds","One tool, many views: JASMINE maps galaxies to your browser","JASMINE's autoencoder turns galaxy images into an explorable sphere","Explore galaxies: JASMINE links images, particles, and metadata"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000742,"raw_usage":{"total_tokens":3274,"prompt_tokens":870,"completion_tokens":2404,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":486,"completion_tokens_details":{"reasoning_tokens":2321}},"tokens_in":486,"tokens_out":2404,"duration_ms":17837,"temperature":1.0,"reasoning_tokens":2321,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:03:49.065576+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the same galaxy sample, sort all pairs by distance in the autoencoded latent space, and compare that ranking with independent human or classifier judgments of visual similarity; if nearby pairs are no more similar than random pairs, the overview's arrangement is arbitrary and the explorer's core navigation claim fails.","supporting_citations":[{"cited_title":"2003, in The Craft of Information Visualization, edited by B","cited_arxiv_id":null,"evidence_quote":"Supplies the 'overview first, zoom and filter, details-on-demand' mantra that structures the entire data-browser workflow."},{"cited_title":"2003, in The Craft of Information Visualization, edited by B","cited_arxiv_id":null,"evidence_quote":"Provides the taxonomy and strategies for multiple-view browsers, including the window-management trade-offs the prototype consciously defers."},{"cited_title":"2022, in Astronomical Data Analysis Software and Systems XXX, edited by J","cited_arxiv_id":null,"evidence_quote":"Supplies the argument for machine-learning representations as a basis for interactive overviews of large datasets."},{"cited_title":"2022, in Astronomical Data Analysis Software and Systems XXX, edited by J","cited_arxiv_id":null,"evidence_quote":"Provides the interactive sky-viewer component that handles zooming, hierarchical sphere rendering, and coordinate overlays in the overview window."}],"review_version":1}