{"id":"6a0d38fe-db29-411b-a645-e10d6d149c0d","arxiv_id":"2607.06344","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"A lifecycle-based and context-sensitive framework maps how ethical risks of personalisation emerge and evolve across human-robot interaction settings.","lead":"This paper presents a conceptual framework that maps ethical risks of personalised human-robot interaction across a six-stage lifecycle and four interaction contexts. A smart generalist would read it to understand how physical robots amplify risks like manipulation and privacy loss compared to screen-based AI.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"The framework's own risk analyses don't demonstrate analytical leverage beyond independent risk consideration; the lifecycle × interaction-type mapping is descriptive rather than generative.","rationale":"The reader's CONDITIONAL verdict is appropriate. My concern sharpens the reader's identified weakness — the issue isn't only that the organising dimensions lack external validation, but that the paper's own use of the framework in Section 3 doesn't demonstrate generative analytical power. The risk analyses follow a predictable template that could be produced without the framework. However, this does not move the verdict below CONDITIONAL for three reasons. First, the paper is explicitly positioned as a perspective/framework paper, not an empirical study, and the authors are transparent about its conceptual nature (acknowledging dehumanisation as 'conceptually grounded but empirically open,' calling the interaction classification a 'loose framework'). Second, the embodiment amplification argument in Section 3.6 is the paper's most distinctive contribution and is argued with reasonable care, identifying three structural mechanisms (enacted behaviour, shared environment, relational framing) that go beyond simply asserting 'embodiment makes things worse.' Third, as a community-organising effort emerging from workshops and debates, the paper's value includes agenda-setting, which doesn't require demonstrated analytical leverage to be useful. The CONDITIONAL verdict correctly signals that the framework is a promising contribution whose central claim of 'systematic' analysis remains unvalidated — both externally (no comparative test against alternatives) and internally (the paper's own analyses don't show the framework generating non-obvious insights). The paper would be significantly strengthened by even one detailed case study where the framework reveals a risk interaction or mitigation that independent risk consideration would miss. I agree partially with the reader: the weakest assumption is related to what they identified, but the more precise formulation is that the framework's analytical leverage is untested by the paper's own analyses, not merely that alternative structurings weren't compared.","tokens_in":39640,"tokens_out":2433,"duration_ms":280012,"concrete_test":"Select 3-4 detailed HRI deployment case studies from the published literature (e.g., Paro in dementia care, a robot tutor in education, a companion robot in domestic settings). Have two groups of HRI ethics experts independently analyse risks: one group uses the paper's framework (lifecycle stages × interaction types), the other uses a flat risk checklist drawn from the same literature. Compare: (a) does the framework group identify risks or risk interactions the checklist group misses? (b) does the framework generate any non-obvious predictions about where/when risks compound across lifecycle stages or interaction types? If the framework-guided analyses do not surface risks or interactions beyond what the checklist produces, the framework's claim to enabling 'systematic analysis' beyond fragmented approaches is not substantiated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that combining six lifecycle stages with a 2×2 interaction-type classification enables 'systematic analysis of how risks arise and evolve' in a way that prior fragmented approaches could not. But the paper's own risk analyses in Section 3 don't demonstrate this generative power. Each risk subsection follows a predictable template: define the risk (drawing on established HCI/AI ethics literature), map it onto lifecycle phases (typically: Design phase sets it up, Modelling entrenches it, Interaction enacts it, Evaluation misses it), and note that long-term settings are worse than short-term and open-domain broader than closed. This pattern is largely what one would produce by considering each risk independently with common sense about time and scope — the framework organises but does not appear to discover. No risk interaction is surfaced that would be invisible without the framework (e.g., no case where the lifecycle × interaction-type combination reveals a non-obvious risk pathway, contradiction, or mitigation that single-dimension analysis would miss). The one worked example (§4.6) is too cursory — six bullet points that read like a generic checklist — to show the framework producing insights a risk list alone wouldn't. If the framework's analytical value is purely organisational (grouping known risks under a common structure) rather than generative (revealing new risks, interactions, or non-obvious dependencies), then the claim of providing a 'systematic foundation' overstates what is essentially a useful taxonomy. The reader's concern about validation of the organising dimensions is part of this, but the sharper issue is internal: the paper doesn't use its own framework in a way that demonstrates its distinctive analytical power.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"This paper presents a conceptual framework — the Responsible Personalisation Framework — for analysing ethical risks of personalisation in human-robot interaction (HRI). The framework combines a six-phase personalisation lifecycle (design, data collection, modelling, interaction, evaluation, end-of-life) with a 2×2 classification of interaction types (short/long-term × open/closed-domain), and uses this structure to examine five risk categories: autonomy erosion, biased user modelling, dehumanisation, manipulation, and privacy violation. The paper also articulates how physical embodiment may amplify these risks relative to disembodied AI, provides five design recommendations, and outlines open research questions. The work is grounded in interdisciplinary workshop discussions and a broad literature review spanning HRI, HCI, and AI ethics.","tokens_in":39740,"tokens_out":1374,"duration_ms":176366,"significance":"The paper addresses a genuine gap: while ethical risks of personalisation are well-studied in HCI, their systematic treatment in the context of embodied HRI is underdeveloped. The embodiment-aware analysis in Section 3.6, which identifies three structural features (enacted behaviour, shared environments, relational framing) through which embodiment amplifies risk, is a valuable conceptual contribution. The lifecycle framing is useful for connecting risks to specific development stages, and the recommendations in Section 4 are actionable and well-grounded. The paper is honest about the conceptual (rather than empirical) status of dehumanisation as a risk. The open research questions are substantive and could productively guide future work.","major_comments":[{"comment":"§2.4 and §3: The central claim is that combining the lifecycle stages with the 2×2 interaction-type classification 'enables systematic analysis of how risks arise and evolve' in a way that prior fragmented approaches could not. However, the risk analyses in Section 3 do not clearly demonstrate analytical leverage beyond what independent risk consideration would produce. Each risk subsection follows a predictable template: design phase sets up the risk, modelling entrenches it, interaction enacts it, evaluation misses it; and long-term settings are worse than short-term, open-domain broader than closed. No risk interaction or non-obvious risk pathway is surfaced that would be invisible without the framework. The one worked example (§4.6) is six bullet points that read like a generic checklist. The paper would be substantially strengthened if at least one risk analysis showed the framework","section":null},{"comment":"§2.3–2.4: The choice of lifecycle stages and interaction-type dimensions as the organising axes is asserted rather than justified. The paper does not explain why these dimensions are more analytically productive than alternatives (e.g., risk severity by user vulnerability, deployment setting, or data sensitivity). The privacy subsection (§3.5) itself acknowledges that 'sensitivity' is an orthogonal dimension that cuts across the interaction-type classification, which raises the question of whether the chosen axes are the most informative. A brief comparative argument for why these dimensions were selected over alternatives would strengthen the framework's foundation.","section":null},{"comment":"§3.3: Dehumanisation is included as one of five key risks, but the paper acknowledges that 'direct empirical demonstration of representational dehumanisation in robot interaction remains limited.' The conceptual argument is well-made, but the risk analysis is notably thinner than the other four risks — the lifecycle mapping and interaction-type analysis are more speculative. The paper should either strengthen this analysis (e.g., by drawing more concrete connections to existing HRI empirical work on objectification or instrumentalisation) or more explicitly frame it as a prospective risk requiring empirical validation, which it partially does but could do more clearly.","section":null}],"minor_comments":[{"comment":"§2.1: The embodiment literature review reports mixed findings on trust (some studies find no significant differences between embodied and virtual agents), but the paper's framing in later sections treats embodiment amplification as established. The nuance in §2.1 should be carried forward more carefully into §3.6.","section":null},{"comment":"Figure 2: The framework overview figure is dense and difficult to parse. The relationship between lifecycle phases, interaction types, risks, and mitigations could be presented more clearly, perhaps with a simpler schematic or multiple sub-figures.","section":null},{"comment":"§2.2: The distinction between adaptation and personalisation is well-articulated, but the summary paragraph notes these are 'regions along a continuum' rather than discrete categories. It would help to acknowledge earlier in the section that real systems combine these, rather than only at the end.","section":null},{"comment":"§3.5: The discussion of consent gaps in experimental HRI is important but somewhat buried within the lifecycle mapping. Consider foregrounding the distinction between consent to data collection and awareness of downstream inference as a standalone point.","section":null},{"comment":"§4.6: The worked example would be more convincing if it showed the framework producing a non-obvious insight — e.g., a risk interaction or mitigation that a standard checklist would miss — rather than confirming expected design choices.","section":null},{"comment":"References: The paper cites a 2026 dated reference [11] (Axelsson and Seeck, 'Just Accepted') and several 2025/2026 references. Ensure all citations are to published or stable versions where possible.","section":null},{"comment":"§5: The call for an open community platform is a constructive initiative but reads somewhat as self-promotion. Consider framing it more neutrally as a community resource.","section":null},{"comment":"Acknowledgement: The use of AI tools (GPT-5.5, figurelabs.ai, whimsical.com) for figure generation should specify what role these tools played and whether outputs were verified by the authors.","section":null}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid conceptual contribution that would benefit from demonstrating the framework's generative power more convincingly. The core issue is that the framework organises known risks well but does not yet show it can discover or reveal non-obvious risk interactions. This is fixable: one or two worked analyses showing the lifecycle × interaction-type combination surfacing a risk pathway that single-dimension analysis would miss would substantially strengthen the contribution. The paper is well-positioned for a venue like ACM/IEEE HRI or an AI ethics journal, but the analytical leverage claim needs more support before publication. The workshop grounding is a strength but should not substitute for demonstrating the framework's utility on its own terms."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and constructive review. The referee identifies three major concerns: (1) the framework's risk analyses do not clearly demonstrate analytical leverage beyond independent risk consideration; (2) the choice of organising axes (lifecycle stages and interaction-type dimensions) is asserted rather than justified relative to alternatives; and (3) the dehumanisation risk analysis is thinner than the other four risks and should either be strengthened or more explicitly framed as prospective. We address each below.","responses":[{"response":"We accept this criticism in substantial part. The referee is correct that the current risk analyses follow a predictable template and that we have not demonstrated the framework's distinctive analytical leverage as convincingly as we should. We will make two revisions. First, we will add at least one extended risk analysis that surfaces a non-obvious pathway through the framework — specifically, we will trace how biased user modelling (§3.2) and manipulation (§3.4) interact across lifecycle stages in a long-term open-domain setting, showing how design-phase optimisation choices propagate through modelling into interaction-phase feedback loops that are invisible to system-level evaluation precisely because the system's own metrics register entrenchment as improvement. This pathway is genuinely hard to see without the lifecycle framing, because the risk emerges from the interaction between phases rather than within any single phase. Second, we will expand the worked example in §4.6 beyond its current checklist format into a narrative case study that shows the framework surfacing trade-offs and tensions (e.g., between exploration for bias correction and psychological safety in clinical populations) that a generic checklist would not reveal. We agree that the current bullet-point format undersells the framework's value.","revision_made":"yes","referee_comment":"§2.4 and §3: The central claim is that combining the lifecycle stages with the 2×2 interaction-type classification 'enables systematic analysis of how risks arise and evolve' in a way that prior fragmented approaches could not. However, the risk analyses in Section 3 do not clearly demonstrate analytical leverage beyond what independent risk consideration would produce. Each risk subsection follows a predictable template: design phase sets up the risk, modelling entrenches it, interaction enacts it, evaluation misses it; and long-term settings are worse than short-term, open-domain broader than closed. No risk interaction or non-obvious risk pathway is surfaced that would be invisible without the framework. The one worked example (§4.6) is six bullet points that read like a generic checklist. The paper would be substantially strengthened if at least one risk analysis showed the framework"},{"response":"This is a fair point and we will address it. We will add a brief comparative justification in §2.3–2.4 explaining why lifecycle stages and interaction-type dimensions were selected over the alternatives the referee suggests. Our reasoning, which we will make explicit in the revision, is as follows. Lifecycle stages were chosen because they map directly onto the development process where interventions are actually implementable: a framework organised by user vulnerability or data sensitivity would classify risks but would not indicate where in the development process they can be addressed. Interaction duration and domain specificity were chosen because they govern the two properties that most directly shape personalisation-specific risk — the depth of the user model (duration) and its breadth (domain) — and thus determine which risks can arise at all in a given setting. We acknowledge the referee's point about data sensitivity: indeed, our own §3.5 notes that sensitivity is orthogonal to the interaction-type classification. We do not claim that our axes are the only informative ones, and we will state this explicitly. Rather, we argue that lifecycle and interaction-type dimensions are complementary: one tells you where to intervene, the other tells you what form the risk takes. We will add a short paragraph discussing why alternatives such as user vulnerability or deployment setting, while valuable, are less suited to the framework's specific purpose of connecting risks to actionable intervention points across the development process.","revision_made":"yes","referee_comment":"§2.3–2.4: The choice of lifecycle stages and interaction-type dimensions as the organising axes is asserted rather than justified. The paper does not explain why these dimensions are more analytically productive than alternatives (e.g., risk severity by user vulnerability, deployment setting, or data sensitivity). The privacy subsection (§3.5) itself acknowledges that 'sensitivity' is an orthogonal dimension that cuts across the interaction-type classification, which raises the question of whether the chosen axes are the most informative. A brief comparative argument for why these dimensions were selected over alternatives would strengthen the framework's foundation."},{"response":"We agree that the dehumanisation analysis is thinner than the other four risk subsections, and we accept the referee's recommendation. We will pursue both suggested strategies. First, we will strengthen the analysis by drawing more concrete connections to existing empirical and theoretical work on objectification and instrumentalisation in HRI and adjacent fields — for example, Sharkey's work on dignity and robot care [169], Söderlund's empirical examination of dehumanisation by service robots [180], and the broader literature on care workers being unable to compete with robots' non-human-like patience (as we note briefly in §3.3 but do not develop). Second, we will more explicitly frame dehumanisation as a prospective risk requiring empirical validation, making clear which claims are conceptually grounded and which await empirical testing. We will add a brief statement at the end of §3.3 flagging this as a risk whose lifecycle and interaction-type mappings are inferential rather than empirically demonstrated, and cross-reference the open research questions in §4 that call for empirical work on this topic.","revision_made":"yes","referee_comment":"§3.3: Dehumanisation is included as one of five key risks, but the paper acknowledges that 'direct empirical demonstration of representational dehumanisation in robot interaction remains limited.' The conceptual argument is well-made, but the risk analysis is notably thinner than the other four risks — the lifecycle mapping and interaction-type analysis are more speculative. The paper should either strengthen this analysis (e.g., by drawing more concrete connections to existing HRI empirical work on objectification or instrumentalisation) or more explicitly frame it as a prospective risk requiring empirical validation, which it partially does but could do more clearly."}],"tokens_in":39315,"tokens_out":1292,"duration_ms":238088,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"Here's my take on the responsible personalisation paper (arXiv:2607.06344). The bottom line: it's a well-organised conceptual framework that addresses a genuine gap in HRI, but the framework organises known risks rather than generating new insights, and the paper overstates its own analytical power relative to what it actually demonstrates. It deserves a serious referee for a perspective or framework track at an HRI venue, but not treatment as an empirically validated result. The paper does several things well. The embodiment amplification argument in Section 3.6 is the strongest part — the three structural features (enacted behaviour, shared physical environment, relational framing) give a concrete account of why embodied personalisation differs from disembodied AI, and the inversion point about scrutiny being lowest where influence is strongest is genuinely sharp. The lifecycle framing (design → data collection → modelling → interaction → evaluation → end-of-life) is a useful organising device, and the customisation/adaptation/personalisation distinction is clearly articulated. The authors are commendably transparent about the conceptual (not empirical) status of dehumanisation risks, and the privacy analysis — particularly the point about inference exceeding consent — is well-grounded. The stress-test concern lands, though. Each risk subsection follows a predictable template: define the risk, map it onto lifecycle phases (design sets it up, modelling entrenches it, interaction enacts it, evaluation misses it), and note that long-term open-domain is worst. This pattern is largely what you'd produce by considering each risk independently with common sense about time and scope. No risk interaction is surfaced that would be invisible without the framework — no case where the lifecycle × interaction-type combination reveals a non-obvious risk pathway or mitigation. The one worked example (§4.6) is too cursory to demonstrate distinctive analytical leverage; it reads like a generic checklist. The reader's concern about validation of the organising dimensions is real but somewhat secondary. The sharper issue is internal: the paper doesn't use its own framework in a way that shows it can discover things a risk list alone wouldn't. If the framework's value is purely organisational, that's still useful, but the claim of providing a 'systematic foundation' overstates it. The interaction-type classification is also asserted rather than derived — no argument for why duration × domain specificity is more productive than, say, user vulnerability or deployment setting. These are not disqualifying flaws for a perspective paper. The framework is a genuine contribution as a shared vocabulary and organising structure, and the design recommendations are practical. But the gap between ambition and evidence should be acknowledged, and a referee should push the authors to either demonstrate the framework's generative power with a richer worked example or moderate the systematicity claims. Recommend for peer review at an HRI venue with a perspective or framework track.","headline":"Solid conceptual framework for ethical risks of personalisation in HRI; the lifecycle × interaction-type structure organises known risks well but doesn't yet demonstrate generative analytical power beyond independent risk consideration.","tokens_in":40429,"tokens_out":655,"would_cite":false,"duration_ms":231988,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Robot personalisation is a double-edged sword, and embodiment is the blade","keywords":[],"falsifier":"The framework would lose its analytical leverage if the lifecycle stages and interaction-type classification turned out not to be the most informative dimensions for risk analysis—for instance, if risk severity were better predicted by user vulnerability, deployment setting, or data sensitivity than by interaction duration and domain breadth. A concrete falsification would be a study showing that practitioners using the framework identify the same risks at the same rate as those using unstructured ethical reflection, or that risks cluster along dimensions orthogonal to the proposed axes.","tokens_in":39779,"feed_emoji":"🤖","tokens_out":1240,"duration_ms":194837,"temperature":0.7,"pith_summary":"This paper argues that personalising robot behaviour—having a robot autonomously learn about a specific user and adapt its actions accordingly—introduces ethical risks that are uniquely amplified by the robot's physical embodiment. Unlike personalisation in screen-based AI, a robot shares the user's physical space, moves through it, acts upon it, and is perceived as a socially present agent with intentions and authority. The paper's central contribution is a framework that combines a six-phase personalisation lifecycle (design, data collection, modelling, interaction, evaluation, end-of-life) with a 2×2 classification of interaction contexts (short-term vs. long-term, open-domain vs. closed-domain). Using this scaffold, the authors trace how five ethical risks—autonomy erosion, biased user modelling, manipulation, dehumanisation, and privacy violation—arise, evolve, and persist differently depending on where in the lifecycle they originate and what type of interaction they occur in. The key insight is that embodiment does not merely add a layer of risk on top of what disembodied AI already carries; it transforms the nature of those risks by making personalisation an enacted, physical, and relational phenomenon rather than a purely informational one. A robot that anticipates a user's needs and physically completes tasks on their behalf redistributes agency in the shared environment; a robot that acts out a reductive user model does so in the physical world, making the reduction harder to dismiss as a computational abstraction. The paper translates this analysis into five design recommendations: critically evaluating whether personalisation is necessary at all, mitigating overtrust and overengagement, enabling user oversight and overrides, continuously auditing models for bias, and defining responsibility for outcomes.","feed_headline":"Physical robots amplify personalisation risks that screens do not","feed_subtitle":"A lifecycle framework traces how autonomy erosion, manipulation, and privacy violation grow worse when a personalised agent shares your body","key_machinery":"The Responsible Personalisation Framework: a six-phase personalisation lifecycle (design, data collection, modelling, interaction, evaluation, end-of-life) crossed with a 2×2 interaction-type classification (short-term vs. long-term, open-domain vs. closed-domain), used to trace how five ethical risks (autonomy erosion, biased user modelling, manipulation, dehumanisation, privacy violation) emerge and evolve differently across contexts. The framework also introduces an input–modelling–output (IMO) functional decomposition that runs in parallel with the temporal lifecycle phases.","core_discovery":"The paper's central object is the Responsible Personalisation Framework, which maps ethical risks of personalisation onto two axes: a six-phase lifecycle (design, data collection, modelling, interaction, evaluation, end-of-life) and a four-cell interaction-type grid (short/long-term × open/closed-domain). The framework's organising claim is that embodiment transforms personalisation from informational mediation into enacted physical behaviour, which amplifies five specific risks—autonomy erosion, biased user modelling, manipulation, dehumanisation, and privacy violation—in ways that are structurally different from disembodied AI. The framework also reveals a recurring pattern: system-level评价","pith_inferences":["The framework's lifecycle stages implicitly suggest a regulatory mapping: different legal responsibilities may attach to different phases (e.g., design-phase decisions about optimisation objectives could be treated as a form of intent, while modelling-phase biases could be treated as negligence). This connection is not drawn out by the paper but follows from its lifecycle-based risk analysis.","The observation that manipulation is hardest to evaluate because the system's own metrics (engagement, acceptance) may constitute the risk implies a need for external, user-defined success criteria that are negotiated before deployment—a form of value alignment that the paper gestures toward but does not formalise.","The framework could be extended to cover multi-robot or multi-agent personalisation scenarios, where several robots with different user models interact with the same user simultaneously; the paper does not address this but the lifecycle analysis would need to account for model conflicts and compounded privacy exposure."],"forward_implications":["If the framework is adopted, HRI researchers and designers would need to evaluate personalisation not just by engagement or task-success metrics but by longitudinal measures of autonomy retention, bias entrenchment, and privacy erosion—metrics that current system-level evaluation cannot detect because the system's own success criteria may constitute the risk.","The end-of-life phase would become a first-class design concern: accumulated user models and interaction data would require active memory management (selective forgetting, decay) throughout operation, not just secure deletion at decommissioning.","The 2×2 interaction-type grid implies that short-term interactions, which lack the data to build persistent user models, should be classified as adaptation rather than true personalisation—potentially requiring different regulatory and ethical standards than long-term personalised systems.","The paper's call for an open community platform and annual structured debates suggests that responsible personalisation is being positioned as an ongoing, collective research agenda rather than a problem solvable by any single framework or set of guidelines."],"fun_headline_variants":["Physical embodiment amplifies ethical risks of personalised robots","A lifecycle framework for ethical risks in personalised robotics","Personalised robots face amplified risks of manipulation and privacy loss","Embodiment transforms the ethical risks of personalised AI","Mapping how embodied personalisation threatens user autonomy"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The paper assumes that its chosen organising dimensions—a six-phase lifecycle and a 2×2 interaction-type grid—are the right axes for analysing personalisation risks, but this choice is asserted rather than derived or validated against alternatives. No empirical test is offered for whether these dimensions help practitioners identify risks they would otherwise miss.","fun_headline_variants_meta":{"raw":{"variants":["Physical embodiment amplifies ethical risks of personalised robots","A lifecycle framework for ethical risks in personalised robotics","Personalised robots face amplified risks of manipulation and privacy loss","Embodiment transforms the ethical risks of personalised AI","Mapping how embodied personalisation threatens user autonomy","Why personalised robots need stricter ethical frameworks than apps","Human-robot interaction amplifies personalisation risks over screen AI"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":1157,"prompt_tokens":511,"completion_tokens":646,"prompt_tokens_details":null},"tokens_in":511,"tokens_out":646,"duration_ms":30827,"temperature":1.0,"reasoning_tokens":682,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-08T09:02:31.092388+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"The framework would lose its analytical leverage if the lifecycle stages and interaction-type classification turned out not to be the most informative dimensions for risk analysis—for instance, if risk severity were better predicted by user vulnerability, deployment setting, or data sensitivity than by interaction duration and domain breadth. A concrete falsification would be a study showing that practitioners using the framework identify the same risks at the same rate as those using unstructured ethical reflection, or that risks cluster along dimensions orthogonal to the proposed axes.","supporting_citations":[],"review_version":1}