{"id":"b3719ab8-637f-44e4-800b-7d1aaa43c431","arxiv_id":"2607.23309","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Artists work with AI models as craft materials; this case study shows how a public toolkit and community can make neural audio latent spaces learnable through hands-on making.","lead":"This paper proposes 'material explainability': turning AI models into hands-on design materials through tools, documentation, and community. A case study with four artists shows how a Max/MSP toolkit for neural audio latent spaces enabled musical interface designs.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim 'supports hands-on artistic explorations' rests on four curated positive cases; no baseline or negative cases, so material explainability cannot be separated from generic toolkit/community effects.","rationale":"We agree with the reader's weakest assumption: the evidence for the central claim is a small set of curated, positive artist cases without a baseline or negative cases. The paper's own framing as a 'heuristic' tempers but does not resolve the causal attribution problem. The toolkit and repository are genuinely public resources and the artist projects are real (credit given), but none of that establishes that the material-explainability approach, rather than the toolkit's novelty or the community's support, caused the observed outcomes. This is a serious but addressable limitation appropriate for a CONDITIONAL verdict. The proposed between-subjects comparison would directly test whether the full package outperforms the raw toolkit alone, settling whether the specific explainability artifacts matter. No internal inconsistency or fatal flaw was found, so the reader's CONDITIONAL verdict should remain.","tokens_in":10295,"tokens_out":3948,"duration_ms":41799,"concrete_test":"Recruit ~16 artists with roughly matched backgrounds and randomly assign them to (A) the full Latent Terrain package with documentation, tutorials, and community showcase, or (B) the raw nn_tilde/RAVE toolkit with only generic help files, both for 4 weeks; measure number of completed projects, time to first working patch, self-reported understanding, and 3-month sustained use. If group B matches group A on these metrics, the material-explainability artifacts are not the active ingredient.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—that material explainability 'supports hands-on artistic explorations of neural audio models in NIME design' (Sec. 5)—is supported only by four invited artists' projects and their positive self-reports (Sec. 3.3). These artists 'shared a mutual interest in exploring neural audio synthesis' and were invited/curated by the authors; no negative cases, attrition, or comparison conditions are reported. The Latent Terrain intervention bundles multiple components: a Max/MSP toolkit, documentation, video tutorials, pretrained autoencoders, and a community showcase. From these data alone, one cannot tell whether the observed exploratory practices are due to the material-explainability activities/artifacts as defined, or to the novelty of neural audio synthesis, the usability of the toolkit, the effects of authorial attention/community curation, or the artists' prior expertise (several are ML/composition researchers). Section 4's three recommendations are built directly on this same evidence, so the prescriptive contribution inherits the selection/causal threat. The paper itself calls the project 'a heuristic' (Sec. 3), yet the conclusion's 'supports' is stated without this hedge. This is not a fatal flaw, but it is load-bearing: if a different adequate toolkit with comparable attention produced the same outcomes, the specific notion of material explainability adds little beyond generic tool-building and community support.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes 'material explainability' as a range of activities and artifacts that make AI models accessible, inclusive design materials for artists. It presents Latent Terrain, a Max/MSP toolkit for exploring latent spaces of neural audio autoencoders, together with online documentation, video tutorials, pre-trained models, and a website for artist project showcases. Four invited artists contributed projects; excerpts from conversations with them are used to reflect on how the artists learned the models through trial and error. From this case study, the paper derives three recommendations for XAI design in NIME: make visible how AI models can be tailored into artistic assemblages; enable learning through crafting; and foster community exchange of techniques. It concludes that material explainability supports hands-on artistic explorations of neural audio models.","tokens_in":10581,"tokens_out":5707,"duration_ms":50858,"significance":"If the conclusion can be backed by stronger evidence, the paper would be a valuable operationalization of the XAIxArts manifesto's call, providing a concrete open repository and detailed documentation in a domain where material-oriented XAI has been mostly conceptual. Strengths include the public Latent Terrain toolkit, the range of supported autoencoders, and the plural artist projects that give texture to the notion of material understanding. The central weakness is that the evidence is limited to four self-selected, curated positive cases, so the causal or enabling role of 'material explainability' specifically remains undemonstrated. As a design case study and a set of design recommendations, it is useful; as support for the stated conclusion, it currently overreaches.","major_comments":[{"comment":"The central claim in §5 that 'the material explainability approach supports hands-on artistic explorations' is stronger than the reported data. §3.3 describes four artists who 'shared a mutual interest' and were invited/showcased on the authors' site; the quotes are from 'our conversations' with those artists. No recruitment criteria, interview protocol, analysis method, negative cases, or comparison condition are reported. The intervention bundles a new toolkit, documentation, tutorials, pretrained autoencoders, authorial attention, and community curation, so the observed explorations cannot be attributed to material explainability as opposed to generic usability, novelty, or artist expertise. The paper itself calls the project 'a heuristic' in §3; I recommend either adding methodological detail and explicit discussion of alternative explanations, or limiting the conclusion to 'illustra","section":"§3.3, §5"},{"comment":"The three recommendations are presented as grounded in the four projects, but the link between the evidence and each recommendation is loose. For example, the first recommendation claims that making visible how AI models can be tailored enabled the iterative integration of materials; yet none of the artist quotes in §3.3 identifies a specific feature of the Latent Terrain toolkit, documentation, or tutorial that created this visibility. Similarly, the second recommendation's 'tacit understanding' is asserted rather than demonstrated by an analysis of artist behavior or verbal reports. Please map each recommendation to observable evidence (e.g., which artifact was used, what changed in the artists' practice over time) or explicitly present the recommendations as design lessons needing further validation.","section":"§4"},{"comment":"Material explainability is defined broadly as 'a range of activities or artifacts' (§1) and is noted in §2.3 to have been 'raised' in the authors' own XAIxArts Manifesto [7]. The breadth of the definition makes it difficult to falsify: almost any tool, documentation, or community activity could count. As such, the conclusion that the approach 'supports' artistic exploration risks being circular unless the paper specifies (a) what distinguishes material explainability from ordinary toolkit development and community building, and (b) what evidence would count against the approach. Please clarify the relationship to [7] and to earlier 'understanding in action' work [13,30,37] so that the contribution is clearly delimited.","section":"§2.3, §5"}],"minor_comments":[{"comment":"'ocio-material environments' appears to be a typo for 'socio-material environments'.","section":"§4"},{"comment":"'around the edges explanations' is quoted but not defined; please briefly explain what actions or insights this type of explanation affords.","section":"§2.2"},{"comment":"'XAIxArts' is used without expansion; spell out 'Explainable AI for the Arts' at first use.","section":"Abstract, §2"},{"comment":"Clarify whether the shared repository contains source code/patches or only descriptions and media; the claim about a 'repository of resources' is stronger if the artifacts themselves are available.","section":"§3.3, Fig. 3"},{"comment":"The styling 'RA VE' is inconsistent; use 'RAVE' or the accepted typography for the model name.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a book chapter rather than a full empirical paper. Its main weakness is the evidence base for the causal claim; however, the conceptual framing and the public toolkit make it suitable if the authors revise to align claims with evidence. I recommend major revision, primarily to address the evaluative gaps outlined above."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a solid practice chapter, not a rigorous empirical study. What's actually new: it takes 'material explainability' from a slogan in the XAIxArts Manifesto and gives it a concrete form — a Max/MSP toolkit, public documentation and tutorials, open autoencoders, and four external artist projects. That operationalization is the real contribution, and it is useful. The writing is clear, the framework is coherent, and the toolkit is real and public. That last point matters for a practice-oriented field; you can go use the thing.\n\nWhat the paper does well: it positions explainability as access to a crafting and hacking space rather than as model transparency, and the three recommendations (tailor AI into artists' material assemblages, enable learning through crafting, build communities for technical know-how) are sensible and grounded in the development experience. The artist reflections are varied and believable. The authors also explicitly call the project 'a heuristic' in Section 3, which is an honest framing.\n\nThe soft spots are load-bearing but not fatal. The central claim — that material explainability 'supports hands-on artistic explorations' — is supported only by four invited artists who shared a mutual interest in neural audio synthesis, with no negative cases, no attrition, no comparison baseline, and no way to separate the material-explainability package from the novelty of neural audio, the usability of the toolkit, authorial attention, or the artists' prior expertise. The conclusion drops the heuristic hedge and states the claim without qualification. Since all three recommendations are built on the same four cases, they inherit the same selection and causal threats. That's worth naming clearly, but for a chapter proposing a framework and illustrating it, it does not sink the work.\n\nThere's also a conceptual circularity worth acknowledging: the term comes from the authors' own manifesto, the toolkit is the first author's doctoral output, and the artists were curated by the authors. That doesn't mean the work is invalid — self-extension is normal in this literature — but the validation loop should be stated more plainly. There are no equations or fitted parameters, so this is not mathematical circularity; it's a matter of interpretive caution.\n\nThe citation pattern is fine. Self-citations point to prior work that this chapter is explicitly extending, and the more general HCI and NIME references are appropriate.\n\nWho is this for? People working on XAIxArts, NIME, or tool-building for creative AI. It would be a reasonable chapter in a Springer collection on explainable AI for the arts. The reader's CONDITIONAL verdict and the stress-test note both land; the paper needs a limitations subsection, a softened conclusion, and ideally a note about how future work could evaluate the approach against a baseline. That's a revision, not a rejection.\n\nSend it to peer review. A careful referee can push on the evidence without throwing out the toolkit.","headline":"A useful operationalization of material explainability with a real public toolkit, but the central claim rests on four curated positive cases; deserves refereeing with a clear limitations section.","tokens_in":11063,"tokens_out":1857,"would_cite":true,"duration_ms":21336,"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":"This paper argues that explainable AI for artists should make neural audio models into hands-on design materials, not just transparent black boxes.","keywords":["material explainability","XAI for the arts","AI as design material","neural audio models","latent space","NIME design","musical interface design","community building"],"falsifier":"A matched-group study in which musicians with similar backgrounds build a musical interface with a neural audio model either through the full Latent Terrain package (toolkit, documentation, tutorials, community) or through a bare API and a technical paper; if the bare-API group develops comparable musical interfaces and reports comparable understanding, the central claim loses its support.","tokens_in":10186,"feed_emoji":"🎛️","tokens_out":4826,"duration_ms":46977,"temperature":0.7,"pith_summary":"The paper argues that explainable AI for artists should not be primarily about opening the black box, but about turning AI models into materials that artists can probe, configure, and craft with over time. It proposes 'material explainability': a range of activities and artifacts—software tools, documentation, tutorials, and shared community spaces—that make AI models accessible and inclusive design materials. The case study follows Latent Terrain, a toolkit that brings neural audio latent spaces into a visual programming environment, and documents four artist-built musical interfaces that each used the model differently. Artist reflections emphasize trial-and-error learning, embodied listening, and the need for better tooling. If the claim holds, explainability for creative practice should be judged by whether it enables sustained hands-on engagement, not by how completely it explains a model's inner workings.","feed_headline":"Make AI a material, not a mystery","feed_subtitle":"A toolkit and community turned neural audio latent spaces into hands-on design materials for four musical interface projects.","key_machinery":"The load-bearing artifact is Latent Terrain, a toolkit of Max/MSP objects that maps the latent space of a neural audio autoencoder into a 2D control space and places it inside a visual programming environment. The latent space is the compact, learned representation an autoencoder uses to compress and reconstruct audio—normally treated as opaque and hard to navigate. Surrounding the toolkit are video tutorials, documented APIs, example patches, an open repository of autoencoders, and a project showcase. Conceptually, the engine is the idea of material agency: a model's properties emerge through use, so explainability means giving artists resources to probe, configure, and hack the model rathe","core_discovery":"On the paper's own terms, the central discovery is that a deliberately assembled ecology of software, documentation, tutorials, and community exchange—what the authors call material explainability—can let artists develop a working, tacit understanding of a neural audio model's latent space. The paper documents Latent Terrain, a package of Max/MSP objects that maps an autoencoder's latent space to a 2D control space, alongside an online repository of docs and tutorials and a community showcase. Four artists contributed projects ranging from a virtual gallery to a live performance system, a meditation sonification, and an ambient soundtrack. Their reflections emphasize learning through trial-a","pith_inferences":["The material-explainability pattern likely extends beyond audio: the same toolkit-plus-documentation-plus-community recipe could be tested with diffusion models, LLM fine-tuning, or visual generative models in creative settings.","If the claim is right, evaluation of XAI for artists should shift from measures of comprehension to measures of sustained practice, reuse, and the range of artifacts produced over time.","A stronger test would separate the ingredients: do artists succeed because of the toolkit itself, the documentation, or the curated community? The paper's case study does not isolate these factors."],"forward_implications":["XAI for creative practice should produce tools and resources that make models modifiable and integrable, not just explanations of how they work.","Artists can acquire operational understanding of opaque neural models through iterative crafting, trial-and-error, and embodied listening.","Documentation and tutorials written for creative practitioners in their own environments lower the barrier to hands-on engagement with AI.","Community spaces where artists share domain-specific techniques sustain long-term engagement with AI materials.","The latent space of a neural audio model can serve as a design material with its own grain, resistances, and propensities."],"fun_headline_variants":["Neural audio models as hands-on design materials","Material explainability: AI as a craftable medium for music","Artists turn AI latent space into tactile design tools","Community toolkit makes neural audio models explorable"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The paper's case rest on the four invited artists being representative, and on their positive reflections being caused by the material-explainability package rather than by the novelty of the tools or the curated community environment.","fun_headline_variants_meta":{"raw":{"variants":["Neural audio models as hands-on design materials","Material explainability: AI as a craftable medium for music","Artists turn AI latent space into tactile design tools","Community toolkit makes neural audio models explorable"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000239,"raw_usage":{"total_tokens":1317,"prompt_tokens":674,"completion_tokens":643,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":418,"completion_tokens_details":{"reasoning_tokens":581}},"tokens_in":418,"tokens_out":643,"duration_ms":6268,"temperature":1.0,"reasoning_tokens":581,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T23:46:20.157102+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A matched-group study in which musicians with similar backgrounds build a musical interface with a neural audio model either through the full Latent Terrain package (toolkit, documentation, tutorials, community) or through a bare API and a technical paper; if the bare-API group develops comparable musical interfaces and reports comparable understanding, the central claim loses its support.","supporting_citations":[],"review_version":1}