Paper deep dive
Designing Adaptive Digital Nudging Systems with LLM-Driven Reasoning
Tiziano Santilli, Mina Alipour, Mahyar Tourchi Moghaddam
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 95%
Last extracted: 4/14/2026, 2:31:56 AM
Summary
The paper presents a novel software architecture for adaptive digital nudging systems that integrates behavioral science with software engineering. It addresses the lack of architectural guidance in current systems by introducing a multi-layered design that treats ethical compliance, fairness, and multi-dimensional user modeling (cognitive mode, behavioral stage, attention capacity) as first-class architectural concerns. The architecture was validated by 13 software architects and demonstrated through an LLM-powered proof-of-concept in residential energy sustainability, showing high feasibility and positive user impact.
Entities (5)
Relation Signals (3)
Digital Nudging System â enforces â Ethical Compliance
confidence 95% · treating ethics and fairness as structural guardrails rather than implementation details
Digital Nudging System â integrates â Behavioral Theory
confidence 95% · We present an architecture that uses behavioral theory through explicit architectural decisions
LLM â powers â Digital Nudging System
confidence 90% · An LLM-powered proof-of-concept in residential energy sustainability demonstrated feasibility
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Digital nudging systems lack architectural guidance for translating behavioral science into software design. While research identifies nudge strategies and quality attributes, existing architectures fail to integrate multi-dimensional user modeling with ethical compliance as architectural concerns. We present an architecture that uses behavioral theory through explicit architectural decisions, treating ethics and fairness as structural guardrails rather than implementation details. A literature review synthesized 68 nudging strategies, 11 quality attributes, and 3 user profiling dimensions into architectural requirements. The architecture implements sequential processing layers with cross-cutting evaluation modules enforcing regulatory compliance. Validation with 13 software architects confirmed requirements satisfaction and domain transferability. An LLM-powered proof-of-concept in residential energy sustainability demonstrated feasibility through evaluation with 15 users, achieving high perceived intervention quality and measurable positive emotional impact. This work bridges behavioral science and software architecture by providing reusable patterns for adaptive systems that balance effectiveness with ethical constraints.
Tags
Links
- Source: https://arxiv.org/abs/2604.11206v1
- Canonical: https://arxiv.org/abs/2604.11206v1
Trouble viewing inline? Open PDF directly â
Full Text
43,541 characters extracted from source content.
Expand or collapse full text
Designing Adaptive Digital Nudging Systems with LLM-Driven Reasoning Tiziano Santilli Mina Alipour Mahyar Tourchi Moghaddam Syddansk Universitet Odense, Denmark Abstract Digital nudging systems lack architectural guidance for translating behavioral science into software design. While research identifies nudge strategies and quality attributes, existing architectures fail to integrate multi-dimensional user modeling with ethical compliance as architectural concerns. We present an architecture that uses behavioral theory through explicit architectural decisions, treating ethics and fairness as structural guardrails rather than implementation details. A literature review synthesized 68 nudging strategies, 11 quality attributes, and 3 user profiling dimensions into architectural requirements. The architecture implements sequential processing layers with cross-cutting evaluation modules enforcing regulatory compliance. Validation with 13 software architects confirmed requirements satisfaction and domain transferability. An LLM-powered proof-of-concept in residential energy sustainability demonstrated feasibility through evaluation with 15 users, achieving high perceived intervention quality and measurable positive emotional impact. This work bridges behavioral science and software architecture by providing reusable patterns for adaptive systems that balance effectiveness with ethical constraints. ©2026 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. Accepted as paper at SAGAI-ICSA2026. I Introduction The role of software systems is evolving from passive tools to active participants in human decision-making. This shift is exemplified by the Internet of Behaviors (IoB), which extends the Internet of Things by connecting user data to behavioral insights [29, 22]. Digital nudging represents a key mechanism for this interaction, using behavioral science to guide users toward beneficial choices in domains like energy conservation [4] and health improvement [15]. However, designing adaptive nudging systems presents fundamental architectural challenges that current practice fails to address systematically. Digital nudging presents three architectural challenges that cannot be resolved through implementation alone: i) Multi-dimensional runtime adaptation. Effective nudging requires simultaneous consideration of cognitive mode (System 1 vs. System 2 thinking), behavioral stage (pre-contemplation through maintenance), attention capacity, and quality attributes [14, 19]. This creates the modifiability-performance trade-off [2]: tight coupling enables coordinated responses but creates brittleness; loose coupling improves maintainability but complicates adaptation. The architectural question becomes: how do we structure components such that changes in one dimension (e.g., attention drops) trigger coordinated adaptations across multiple subsystems (UI simplification, strategy adjustment, message adaptation) without tight coupling? i) Non-negotiable ethical constraints as architectural guardrails. The EU AI Act, GDPR Article 22, and Digital Services Act impose hard constraints on automated decision-making systems [30, 12, 6]. These are not optional features but mandatory architectural properties that must hold across all execution paths. Treating ethics as an afterthought leads to architectures where manipulative strategies can slip through during edge cases. The architectural challenge is: how do we enforce ethical boundaries structurally, such that no component can generate non-compliant nudges regardless of user state or system evolution? i) Quality attribute conflicts requiring explicit prioritization. Literature reveals inherent tensions: transparency increases trust but may trigger reactance [1, 13]; personalization improves effectiveness but raises fairness concerns [20]; simplification aids low-attention users but limits information for analytical thinkers [14, 19]. These conflicts demand architectural tactics, like explicit design decisions that prioritize certain qualities at the expense of others, with documented rationale and trade-offs. Current practice treats user adaptation as an implementation concern rather than an architectural requirement, resulting in systems that cannot guarantee consistent ethical operation or adapt systematically [16, 25]. Software architects lack systematic guidance for translating behavioral science principles into concrete architectural decisions, Section I-C. To address these challenges, we must understand the theoretical foundation. Behavioral theory, originating with Watsonâs behaviorism [32] and advanced by Skinnerâs operant conditioning [27, 28], posits that human actions result from conditioned responses to environmental stimuli. Learning is defined as measurable behavioral change through experience. Despite this theoretical background, the software engineering community lacks systematic approaches to address these concepts architecturally. This paper addresses these gaps through four contributions: i) a literature review synthesizing nudging strategies, quality attributes, and user profiling dimensions into architectural requirements; i) an architecture treating behavioral dimensions and ethical compliance as first-class architectural concerns; i) validation with software architects confirming requirements satisfaction and transferability; iv) a proof-of-concept implementation using large language models (LLMs) for cognitive reasoning, with empirical user evaluation. We address three research questions: âą RQ1: What are the key characteristics, quality attributes, and metrics for digital nudging? âą RQ2: To what extent can a software architecture integrate nudging characteristics with intervention logic while preserving modularity and ethical compliance? âą RQ3: To what extent does the instantiated architecture demonstrate feasibility and generate effective adaptive interventions? For the methodology we employ a four-phase approach Figure 1: i) Literature Review synthesizing behavioral characteristics into requirements; i) Architectural Design translating requirements into components with explicit decision rationale; i) Architect Evaluation validating requirements satisfaction and transferability; iv) Implementation & User Evaluation demonstrating feasibility through proof-of-concept. Figure 1: Research Methodology Overview The paper proceeds as follows: Section I reports literature review and gap analysis; Section I presents the architecture with architectural decisions and validation; Section IV describes implementation and user evaluation; Section V discusses validity threats; Section VI concludes with future work. Supplemental materials available at this link111https://github.com/tiziasan/Adaptive-Digital-Nudging-System I Nudging Literature Review To establish an understanding of digital nudging characteristics, we conducted a literature review. I-A Review Process We queried four academic databases (IEEE Xplore, Scopus, ACM Digital Library, Web of Science) using the search string: ((nudge* OR nudging) AND ("quality attribute*" OR metric* OR framework OR characteristic*)). The search yielded 707 papers. After removing 308 duplicates and applying inclusion/exclusion criteria, full-text review resulted in 21 primary studies. Inclusion/exclusion criteria list and literature review flow are available in the supplemental materials. I-B Results Our review identified three key contributions: i) Nudging Strategies (N = 68). We extracted a taxonomy spanning information provision (anchoring, framing, simplification), social influence (social norms, peer support), defaults and positioning, feedback mechanisms, and timing strategies. The complete taxonomy is available in supplemental materials. i) Quality Attributes (N = 11). Literature identifies attributes determining effectiveness and ethicality: acceptability, awareness, effectiveness, fairness, healthfulness, helpfulness, intrusiveness, motivation, transparency, trustworthiness, and user impact. Table I presents definitions. TABLE I: Quality Attributes for Digital Nudging Attribute Definition Ref Acceptability Degree to which users perceive a nudge as appropriate or desirable, considering social, cultural, moral standards [1] Awareness Degree of user attention to the nudge. Tension exists: high transparency may trigger reactance, subliminal approaches raise ethical concerns [1] Effectiveness Degree to which a nudge achieves its intended behavioral outcome [31] Fairness Ensures systems address algorithm/data biases and cognitive/perceptual biases, providing equitable access. Includes GDPR, AI Act, DSA compliance [20] Healthfulness Degree to which users perceive an item or choice to impact their health positively [1] Helpfulness Degree to which users perceive nudges as assisting decision-making [1] Intrusiveness Low-intrusive nudges subtly guide decisions; high-intrusive nudges strongly influence decisions [31] Motivation Degree to which users are motivated to make choices aligned with the nudgeâs goal. [1] Transparency Explainability frameworks for DSA/AI Act requirements while maintaining effectiveness [1, 9, 20] Trustworthiness Degree to which users trust the nudge and information provided [1] User Impact Effect of nudge on userâs goals, preferences, or well-being [31] i) User Profiling Dimensions (N = 3). Literature identifies three behavioral dimensions: Cognitive Mode: Nudges target either System 1 (fast, intuitive) or System 2 (slow, analytical) thinking [14, 19]. System 1 nudges leverage automatic processing; System 2 nudges support deliberate reasoning. Behavioral Stage: The Transtheoretical Model recognizes five stages [14]: pre-contemplation (unaware), contemplation (considering change), preparation (planning action), action (actively modifying behavior), maintenance (sustaining change). Each stage requires different intervention strategies. Attention Capacity: Users make choices attentively or inattentively based on cognitive load [19, 21]. Low attention requires simple nudges; high attention enables information-rich strategies. RQ1 Answer Our literature review identified 68 nudging strategies, 11 quality attributes, and 3 user profiling dimensions. These constitute requirements for digital nudging systems that align interventions with user states. I-C Gap Analysis i) Digital Nudging Architectures. Karlsen and Andersen present a smart nudging system with user profile learner and nudge design components [16, 25], but assess only interests, capabilities, and activity history, omitting cognitive mode, attention state, and quality attributes as architectural dimensions. Their system represents closest prior work but lacks systematic integration of behavioral theory dimensions. i) Behavioral Theories in Software. Grisiute and Raubal explore spatial nudging converging persuasive technologies with spatial design [14], but do not provide architectural guidance. Lofgren and Nordblom present theoretical framework of decision-making explaining nudging mechanisms [19], but focus on psychological models rather than software architecture. i) Ethical AI Architectures. Milano et al. discuss recommender systemsâ ethical challenges [20], emphasizing fairness as critical concern. Fabbri examines self-determination through explanation for DSA transparency requirements [9]. Our architecture operationalizes these ethical concerns through structural enforcement rather than post-hoc auditing. iv) Quality-Driven Architectures. Bass et al.âs quality attribute scenarios [2] provide methodological foundation for our approach. We extend this by showing how behavioral characteristics function as quality attributes driving architectural decisions, alongside traditional attributes. I Digital Nudging Architecture We translate literature findings into an architecture addressing the identified gaps. This section presents requirements with traceability, architectural design with explicit decisions, and validation results. I-A Requirements Traceability Table I presents functional requirements (FR) and non-functional requirements (NFR) derived from literature. TABLE I: Architectural Requirements with Literature Traceability ID Requirement Component Derivation & References Functional Requirements FR1 Capture context and behavioral data Context Tracker, Behavioral Collector Context-awareness [24], behavioral signals [5, 14] FR2 Determine cognitive mode Cognitive Mode Dual-process theory [14, 19] FR3 Detect behavioral stage Behavioral Stage Transtheoretical Model [3, 14] FR4 Estimate attention capacity Attention Capacity Cognitive load theory [19, 21] FR5 Generate diverse nudges Nudge Generator Strategy taxonomy [3, 23, 31] FR6 Optimize strategy selection Strategy Optimizer Architectural necessity for multi-dimensional constraint satisfaction FR7 Provide transparency Explainability Regulatory requirements [1, 13, 9] FR8 Capture user feedback User Feedback Adaptive refinement [5, 17] FR9 Track outcomes Outcome Tracker Effectiveness measurement [31] Non-Functional Requirements NFR1 Ethics compliance Ethics Compliance GDPR/AI Act/DSA mandates [30, 12, 6, 13, 26] NFR2 Fairness monitoring Fairness Monitor Bias mitigation [20] NFR3 Transparency Explainability DSA Article 27, AI Act Article 13 [1, 9, 20] NFR4 Adaptivity Adaptive Dashboard, UI Adaptation State-dependent effectiveness [5, 24, 14] FR6 Rationale: Strategy Optimizer emerges from architectural necessity, not explicit literature mention. When multiple strategies match user state (e.g., System 1 thinking + pre-contemplation + low attention), we need conflict resolution logic to select the single most appropriate intervention. This component implements constraint satisfaction reasoning. NFR Selection Rationale: From 11 quality attributes (Table I), we prioritized 4 as NFRs based on: i) Regulatory mandates, ethics Compliance and Fairness address legally required properties under GDPR Article 22, AI Act Article 10, DSA Article 27; i) Architectural impact, Transparency and Adaptivity fundamentally shape system structure. The remaining attributes are evaluation metrics measured during user testing (Section IV-C). I-B Architecture Description Figure 2 presents our architecture comprising three core processing layers (Data Capture, User Modeling, Nudge Intelligence) with two cross-cutting modules (Adaptation, Evaluation). Figure 2: Digital Nudging Architecture i) Data Capture captures raw signals through Context Tracker (device, location, time-of-day) and Behavioral Collector (clicks, hesitation times, navigation patterns, facial emotions). i) User Modeling transforms signals into multi-dimensional profiles through three parallel components: Cognitive Mode (System 1 vs. System 2), Behavioral Stage (Transtheoretical Model mapping), Attention Capacity (high/medium/low). i) Nudge Intelligence generates personalized interventions through Strategy Optimizer that selects optimal strategy given user profile, Nudge Generator that creates message content, and UI Adaptation that specifies interface parameters. iv) Adaptation Module (left side) implements delivery through Explainability generating transparency reports, Adaptive Dashboard rendering dynamic interface, and User Feedback capturing explicit ratings. v) Evaluation Module (right side) ensures quality through Fairness Monitor that detects population-level biases, Ethics Compliance that validates regulatory constraints, and Outcome Tracker that logs execution traces. I-C Architectural Decisions and Rationale We document key architectural decisions following the template: Decision, Rationale, Trade-offs, and Alternatives Considered in Table I. TABLE I: Architectural Decisions and Rationale Decision Details AD1: Sequential Pipeline vs. Event-Driven Architecture Decision Three sequential layers with per-user session processing rather than event-driven publish-subscribe. Rationale Digital nudging requires deterministic reasoning chains where User Modeling completes before Strategy Optimization. Sequential processing ensures causal consistency, reasoning transparency, and provenance tracking. Trade-offs GainsâClear provenance, easier debugging, compliance auditability (GDPR Article 22), simplified testing. CostsâNo intra-user parallelization, increased per-user latency. Alternatives Event-driven with orchestration (rejected: excessive overhead), blackboard architecture (rejected: non-deterministic control flow), pipes-and-filters (rejected: unnecessary given sequential requirements). AD2: Side-Mounted Evaluation vs. Embedded Validators Decision Ethics Compliance as side-mounted module intercepting all Nudge Intelligence outputs. Rationale Enforces separation of concerns: nudge generation focuses on effectiveness; ethics audits all outputs. Interceptor pattern ensures no nudge reaches users without validation. Provides architectural evidence for DSA Article 27 compliance. Trade-offs GainsâConsistent enforcement, evolvability, independent testability, regulatory evidence. CostsâAdditional latency, potential bottleneck. Alternatives Embedded validators (rejected: duplication risk), post-hoc filtering (rejected: lacks explicit architectural representation), aspect-oriented programming (rejected: deployment complexity). AD3: LLM-Driven Reasoning vs. Rule-Based Classification Decision User Modeling components use LLMs rather than rule-based systems. Rationale Behavioral constructs like âSystem 1 thinkingâ require contextual interpretation of multiple weak signals [18, 19]. LLMs handle ambiguous cases (e.g., high clicks but short duration). Sequential pipeline ensures deterministic data flow despite LLM stochasticity within classifications. Trade-offs GainsâRobustness to edge cases, expressiveness, adaptability via prompt engineering. CostsâControlled non-determinism, explainability challenges, API latency (200-500ms), vendor dependency. Alternatives Threshold rules (rejected: lacks semantic understanding), supervised ML classifiers (rejected: insufficient for nuanced behavioral patterns), hybrid approach (future work). AD4: Backend-Driven UI Adaptation vs. Client-Side Adaptation Decision Backend generates structured âUI Contextâ payloads specifying interface parameters sent to frontend. Rationale Attention capacity and cognitive mode affect information processing [14, 19]. Backend determines what adaptations based on User Modeling; frontend determines how to render. Implements Strategy Pattern keeping behavioral intelligence server-side. Trade-offs GainsâPersonalization, reduced cognitive load, accessibility compliance, privacy protection. CostsâFrontend complexity, consistency challenges, increased payload size. Alternatives Frontend-driven (rejected: exposes interaction data to browser), progressive disclosure (rejected: orthogonal concern), static templates (rejected: cannot provide fine-grained adaptation). I-D Architectural Evaluation We evaluated the architecture through two methods: i) questionnaire assessing requirements satisfaction, i) follow-up discussions exploring architectural insights. We recruited 13 software architects through professional networks. Selection criteria: i) experience with quality-attribute-driven projects, i) familiarity with adaptation systems. Table IV presents demographics. TABLE IV: Architectural Evaluators Demographics (N=13) Characteristic Category N (%) Experience 0â5 years 4 (30.8%) 5â10 years 5 (38.5%) 10â15 years 1 (7.7%) 15+ years 3 (23.1%) Domain Healthcare 3 (23.1%) Automotive 4 (30.8%) Consultancy 3 (23.1%) Other 3 (23.1%) Participants received: i) architecture diagram, i) requirements table, i) architectural decisions document. They completed a questionnaire with 16 closed-ended questions using 5-point Likert scales (1=strongly disagree, 5=strongly agree). Questions assessed whether components met requirements and questionnaire data available in the supplemental materials. Table V presents results. Mean rating across all items was 4.62 (SD=0.12), indicating strong agreement. UI Adaptation achieved highest rating (M=4.85, SD=0.38); Explainability received lowest (M=4.38, SD=0.65), suggesting this is the most contentious aspect. TABLE V: Structured Questionnaire Results (Likert Scale 1-5) Requirement Mean Architectural Quality Maintainability & Evolvability 4.62 Regulatory Compliance 4.69 Functional Requirements Context/Behavioral Data (FR1) 4.54 Cognitive Mode (FR2) 4.62 Behavioral Stage (FR3) 4.62 Attention Capacity (FR4) 4.54 Nudge Generator (FR5) 4.69 Strategy Optimizer (FR6) 4.69 Explainability (FR7) 4.38 User Feedback (FR8) 4.46 Outcome Tracker (FR9) 4.54 Non-Functional Requirements Ethics Compliance (NFR1) 4.54 Fairness Monitor (NFR2) 4.77 Transparency (NFR3) 4.69 UI Adaptation (NFR4) 4.85 Domain Generalizability Highly Transferable 61.5% Transferable with modifications 38.5% We conducted a discussion with 5 of 13 participants (others declined due to time constraints). Discussions were semi-structured around: (1) architectural strengths/weaknesses, (2) missing concerns, (3) transferability. Two participants highlighted separation between behavioral reasoning and ethical enforcement: âI like that ethics is not inside the nudge generator. Thatâs a recipe for unexpected behaviors. Having it as a separate point makes it testable and auditable.â , P3. Three participants raised concerns: âUsing LLMs for cognitive mode worries me. How do you regression test that? If GPT gets updated and starts classifying differently, your system behavior changes. You need to carefully manage the LLM non determinismâ, P2. All 5 agreed the architecture transfers to other domains: âI could see this working for health behavior change in physical activity or medication adherence.â, P4. Qualitative discussions confirmed questionnaire findings. Architectural separation was praised. Explainability emerged as primary weakness, with calls for richer provenance tracking. LLM non-determinism identified as operational risk requiring mitigation. Transferability was also strongly supported. RQ2 Answer We designed an architecture systematically integrating behavioral characteristics through three sequential processing layers and two cross-cutting modules. Ethical compliance is enforced structurally through side-mounted Evaluation implementing Fairness Monitor, Ethics Compliance, and Explainability as first-class architectural concerns. Architect validation confirms successful requirements satisfaction (M=4.62/5), with UI Adaptation highest rated (4.85) and Explainability identified for improvement (4.38). All architects assessed the architecture as transferable across domains. The architecture demonstrates behavioral theory integration and ethical compliance achievable through architectural patterns enforcing constraints at design time. IV Implementation and User Evaluation To demonstrate architectural feasibility, we implemented a proof-of-concept in residential energy sustainability. This section describes implementation details and empirical user evaluation. IV-A Implementation Context Household energy consumption accounts for approximately 26% of EU greenhouse gas emissions [8], making residential energy reduction central to sustainability. The domain requires systems adapting interventions to behavioral stages while managing high-consumption appliances. For the implementation we created a Python backend orchestrating five architectural modules using LLMs for cognitive reasoning via OpenAI API, and React/TypeScript frontend realizing the Adaptation module. Data interchange uses JSON contracts. Code available in supplemental materials. IV-B Implementation Walkthrough i) Context Tracker captures device type (desktop vs. mobile) and temporal attributes (time-of-day: morning/afternoon/evening) from session timestamps. i) Behavioral Collector aggregates interaction signals from frontend: hesitation time, clicks, energy consumption, usage hours, user emotions (facial expressions via FaceAPI [10]). Signals normalized to JSON serving as cognitive profiling input. i) Cognitive Mode uses LLM classifier with Context Tracker + Behavioral Collector data. Specialized prompt analyzes interaction velocity; high clicks + long hesitation on high-wattage appliances indicates analytical processing. iv) Behavioral Stage uses LLM classifier mapping users to Transtheoretical Model stages. v) Attention Capacity uses LLM classifier estimating cognitive capacity from contextual factors, influencing UI and message complexity. vi) Strategy Optimizer uses LLM with Cognitive Mode + Behavioral Stage + Attention Capacity data, selecting optimal intervention from 23-strategy taxonomy [3] via constraint-satisfaction prompting. vii) Nudge Generator uses selected strategy + appliance data to create message via LLM, enforcing regulatory constraints by integrating Ethics and Fairness prompts. Ensures messages are concise, personalized with consumption data, aligned with user state. viii) UI Adaptation uses LLM processing Cognitive Mode + Attention Capacity + Behavioral Stage to dynamically adjust: font size (12pxâ20px), primary/secondary colors, chart type (bar/pie/line). Ensures UI aligns with cognitive state. ix) Explainability generates natural language âTransparency Explanationâ revealing systemâs decision logic, satisfying transparency requirements. Explainability panel screenshot available in the supplemental materials. x) Adaptive Dashboard provides interactive environment for monitoring/modifying appliance states. Dynamically renders content based on UI Adaptation output, displays nudges and explanations, connects to feedback system. Adaptive dashboard screenshot is available in the supplemental materials. xi) User Feedback captures âthumbs up/downâ feedback. Previous nudges tracked via history, informing future Strategy Optimizer decisions. xii) Fairness Monitor injects âBias Mitigationâ system prompt into every LLM call, instructing non-discrimination and vulnerable user protection. xiii) Ethics Compliance enforces EU AI Act and GDPR constraints by injecting mandatory instructions into LLM prompts as guardrail against manipulative strategies. xiv) Outcome Tracker records every intermediate step (raw signals â profiling â strategy â message) and emotional states (pre-nudge and post-nudge) using facial emotion detection. Stored in CSV enabling longitudinal analysis. IV-C User Evaluation We conducted empirical evaluation with 15 participants interacting with the implemented dashboard to assess feasibility and effectiveness. We recruited participants through university networks (N = 8) and local community channels (N = 7). Criteria: i) age 18+, i) responsible for household energy decisions. Table VI presents demographics. TABLE VI: User Study Participant Demographics (N=15) Characteristic Category N (%) Age 18â24 5 (33.3%) 25â34 6 (40.0%) 35â44 3 (20.0%) 45+ 1 (6.7%) Gender Male 8 (53.3%) Female 7 (46.7%) Education High school 2 (13.3%) Bachelorâs 7 (46.7%) Masterâs/PhD 6 (40.0%) Energy Management High awareness 4 (26.7%) Experience Medium awareness 7 (46.7%) Low awareness 4 (26.7%) Tech Smart home user 6 (40.0%) Familiarity Basic tech user 7 (46.7%) Low tech user 2 (13.3%) Sample represents diverse backgrounds. Age distribution skews younger (73.3% under 35) due to university recruitment. Users accessed web dashboard and explored freely. They could add/remove appliances, modify device states, adjust usage hours. Dashboard captured behavioral signals (clicks, hesitation). System monitored emotions via facial recognition throughout session. After each generated nudge, users read message and explanation, then rated on a Likert scale: i) Nudge Quality, âThis nudge is helpful and appropriate for my current situationâ; i) Explanation Quality, âThe explanation clearly describes why this nudge was shownâ. Sessions lasted 20-30 minutes. All provided explicit consent. Our evaluation addresses RQ3 through three measures: 1. Feasibility (Architecture Instantiation): Demonstrated by showing architecture executes across all components with real users. Feasibility established if: i) User Modeling produces classifications for all sessions, i) Strategy Optimizer selects strategies from taxonomy, i) Nudge Generator produces coherent messages, iv) UI Adaptation applies interface changes, v) Ethics Compliance doesnât block compliant nudges, vi) system completes full cycle without failures. Reported in Table VII. 2. Effectiveness (Intervention Quality): Measured through perceived effectiveness via subjective ratings: i)Nudge Quality, i) Explanation Quality. Results in Table VIII. 3. Emotional Impact (User Experience): Assessed whether adaptive nudges produce positive or negative emotional responses using facial expression analysis. Addresses key risk: personalized interventions might feel invasive, triggering negative affect. Conversely, helpful nudges should produce positive affect. Facial Expression Analysis: Used FaceAPI [10], a JavaScript library implementing Ekman facial emotions [7], detecting seven basic emotions: happiness, sadness, anger, fear, disgust, surprise, neutral. Library outputs probability distributions over emotions per frame. We focus on happiness because positive-valence emotion indicates receptiveness and sustained positive affect predicts behavior change adherence [11]. We report raw probability values from FaceAPI (range: 0.0â1.0, where 1.0 = 100% confidence). Table VII shows user testing generated 15 complete executions capturing adaptive behavior. System classified majority (73.3%) as intuitive mode, 26.7% analytical. Behavioral stage: 40% pre-contemplation, 26.7% contemplation, 33.3% action. Attention capacity: 66.7% high, 20% medium, 13.3% low. Strategy Optimizer selected interventions based on profiles. Just-in-time prompts emerged most frequently (53.3%), with remaining 46.7% across: remind consequences, raise visibility, enable comparisons, reduce distance. UI Adaptation adjusted parameters: font size 16px for 73.3%, 19px for 13.3%, 24px for 13.3%, demonstrating response to attention capacity. TABLE VII: User Testing System Classifications (N=15) ID Cognitive Stage Attn Strategy 1 Analytical Contemplation High Just-in-time 2 Intuitive Pre-contempl. High Remind conseq. 3 Intuitive Contemplation High Enable compar. 4 Intuitive Action High Raise visibility 5 Intuitive Contemplation High Just-in-time 6 Intuitive Action Medium Just-in-time 7 Intuitive Pre-contempl. High Remind conseq. 8 Intuitive Pre-contempl. Low Just-in-time 9 Intuitive Pre-contempl. Low Reduce distance 10 Intuitive Action Medium Just-in-time 11 Intuitive Contemplation High Just-in-time 12 Analytical Pre-contempl. Medium Raise visibility 13 Intuitive Pre-contempl. High Just-in-time 14 Analytical Action High Just-in-time 15 Analytical Action High Raise visibility Table VIII presents evaluation metrics. Participants rated nudge quality highly (M=4.73, SD=0.46): 73.3% provided maximum rating (5), 26.7% rated 4. Explanation quality received slightly lower but positive ratings (M=4.27, SD=0.70): 40% rated 5, 46.7% rated 4, 13.3% rated 3. This suggests while users found nudges highly appropriate, transparency could be strengthened, consistent with architect evaluation. TABLE VIII: User Evaluation Results Metric Mean Value Nudge Quality (1â5) 4.73 (SD=0.46) Explanation Quality (1â5) 4.27 (SD=0.70) Pre-Nudge Happiness 0.000170 (SD=0.000092) Post-Nudge Happiness 0.000500 (SD=0.000213) Happiness Increase 0.000330 (SD=0.000158) Emotional impact analysis revealed measurable positive effects. Pre-nudge happiness scores (M=0.000170) reflected predominantly neutral emotional states. Post-nudge happiness scores (M=0.000500) showed mean increase of 0.000330, with all participants showing positive changes. While absolute values remain small due to FaceAPI probability output range, consistent directional change across all participants indicates adaptive nudge delivery produces positive emotional responses rather than reactance or annoyance. Complete execution traces demonstrate architecture successfully adapts behavior across multiple dimensions simultaneously. Sessions with low attention capacity consistently received simplified UI configurations, while Strategy Optimizer simultaneously selected low-complexity approaches rather than information-rich strategies. This coordination between User Modeling and both Nudge Intelligence and UI Adaptation validates architectural designâs ability to integrate behavioral signals into adaptive responses. RQ3 Answer The instantiated architecture demonstrates effective adaptive behavior in sustainable energy domain. User profiling components successfully classified participants across cognitive mode, behavioral stage, and attention capacity dimensions. System selected contextually appropriate nudging strategies and UI configurations based on classifications. User evaluation confirms high perceived quality for generated nudges (M=4.73/5) and satisfactory explanation quality (M=4.27/5). Emotional impact measurement shows positive affect changes across all participants following nudge delivery, indicating interventions are well-received rather than perceived as manipulative. Feasibility is demonstrated through successful execution of all 15 user sessions without failures, with complete architectural pipeline from behavioral signal capture through adaptive intervention delivery. V Threats to Validity Internal Validity. Architects evaluated architecture after detailed explanations, creating priming effects influencing assessments. Future work should employ blind evaluation. User testing lacked baseline measurements and control conditions; observed outcomes may reflect novelty effects. We partially addressed this by capturing pre-nudge and post-nudge emotional states within sessions. Implementation relies on LLMs exhibiting non-deterministic behavior; we mitigated this setting low temperature (0.3) for classification tasks. Construct Validity. Facial emotion recognition may not validly represent deeper affective states driving sustained behavior change. We triangulated emotion data with explicit user feedback ratings. Future implementations should validate classifications against validated behavioral science measures. External Validity. Implementation instantiates architecture exclusively in residential energy sustainability. While architects rated it highly transferable, empirical validation across domains (healthcare, finance) remains necessary. Implementation uses specific technologies (React, Python, OpenAI APIs) that may not generalize to resource-constrained environments. Architectureâs abstract design remains technology-agnostic, enabling alternative implementations. VI Conclusion and Future Work This paper bridges behavioral theory and software architecture by presenting an architecture that translates psychological constructs into executable software components while enforcing ethical compliance structurally. We conducted a literature review identifying 68 nudging strategies, 11 quality attributes, and 3 user profiling dimensions, transformed into architectural requirements with full traceability. The architecture implements three core processing layers (Data Capture, User Modeling, Nudge Intelligence) with two cross-cutting modules (Adaptation, Evaluation) enforcing fairness monitoring and regulatory compliance as architectural guardrails. Architect validation with 13 professionals confirmed requirements satisfaction and domain transferability. Implementation in residential energy sustainability using LLMs for cognitive reasoning demonstrated feasibility. User evaluation with 15 participants achieved high perceived nudge quality and measurable positive emotional impact. Explicit architectural decisions documented trade-offs between quality attributes, showing how behavioral theory shapes software structure. Future Work. Three directions emerge: i) Cross-domain validation, empirically testing architecture in healthcare, finance, e-commerce to strengthen generalizability claims; i) Longitudinal studies, tracking users over weeks/months to assess sustained impact and habituation effects; i) Enhanced explainability providing richer access to decision states, enabling transparency reports revealing considered alternatives and rejection rationale. acknowledgment This research was supported by the Nudge2Green project (1152-00043A), funded by AgriFoodTure, Innovation Fund Denmark. References [1] M. Alsaeed, M. Adam, J. Piper, and T. Bucher (2025-06) Digital nudging for healthy and sustainable food choice in online supermarkets: a scoping review. AIS Transactions on Human-Computer Interaction 17, p. 287â332. External Links: Document Cited by: §I, TABLE I, TABLE I, TABLE I, TABLE I, TABLE I, TABLE I, TABLE I, TABLE I, TABLE I. [2] L. Bass, P. Clements, and R. Kazman (2003) Software architecture in practice. SEI series in software engineering, Addison-Wesley. External Links: ISBN 9780321154958, LCCN 2003045300 Cited by: §I, §I-C. [3] A. Caraban, E. Karapanos, D. Gonçalves, and P. Campos (2019) 23 ways to nudge: a review of technology-mediated nudging in human-computer interaction. CHI â19, New York, NY, USA, p. 1â15. External Links: ISBN 9781450359702, Document Cited by: TABLE I, TABLE I, §IV-B. [4] C. Charlier, G. Guerassimoff, A. Kirakozian, and S. Selosse (2020-01) Under pressure! nudging electricity consumption within firms. feedback from a field experiment. The Energy Journal 42, p. . External Links: Document Cited by: §I. [5] S. Dalecke and R. Karlsen (2020) Designing dynamic and personalized nudges. In Proceedings of the 10th International Conference on Web Intelligence, Mining and Semantics, WIMS 2020, New York, NY, USA, p. 139â148. External Links: ISBN 9781450375429, Document Cited by: TABLE I, TABLE I, TABLE I. [6] Digital services act. External Links: Link Cited by: §I, TABLE I. [7] P. Ekman (1992) An argument for basic emotions. Cognition & Emotion 6 (3-4), p. 169â200. External Links: Document Cited by: §IV-C. [8] Energy consumption in households. External Links: Link Cited by: §IV-A. [9] M. Fabbri (2023) Self-determination through explanation: an ethical perspective on the implementation of the transparency requirements for recommender systems set by the digital services act of the european union. In AAAI/ACM, AIES â23, New York, NY, USA, p. 653â661. External Links: ISBN 9798400702310, Document Cited by: §I-C, TABLE I, TABLE I, TABLE I. [10] Face-api.js. External Links: Link Cited by: §IV-B, §IV-C. [11] B. L. Fredrickson (2001) The role of positive emotions in positive psychology: the broaden-and-build theory of positive emotions.. American Psychologist 56 (3), p. 218â226. External Links: Document Cited by: §IV-C. [12] General data protection regulation. External Links: Link Cited by: §I, TABLE I. [13] N. Gerber, A. Stöver, J. Peschke, and V. Zimmermann (2023-11) Donât accept all and continue: exploring nudges for more deliberate interaction with tracking consent notices. ACM Trans. Comput.-Hum. Interact. 31 (1). External Links: ISSN 1073-0516, Document Cited by: §I, TABLE I, TABLE I. [14] A. Grisiute and M. Raubal (2024) Spatial Nudging: Converging Persuasive Technologies, Spatial Design, and Behavioral Theories. In 16th International Conference on Spatial Information Theory (COSIT 2024), Vol. 315. External Links: ISBN 978-3-95977-330-0, ISSN 1868-8969, Document Cited by: §I, §I, §I-B, §I-B, §I-C, TABLE I, TABLE I, TABLE I, TABLE I, TABLE I. [15] L. Johannsdottir, S. Siguroardottir, M. ĂskarsdĂłttir, and A. Islind (2025-06) Digital nudging in digital health technologies: a systematic review. Health and Technology 15, p. . External Links: Document Cited by: §I. [16] R. Karlsen and A. Andersen (2019) Recommendations with a nudge. Technologies 7 (2). External Links: ISSN 2227-7080, Document Cited by: §I, §I-C. [17] Y. Lian, C. Song, and T. Ge (2025) ITMPRec: intention-based targeted multi-round proactive recommendation. In Proceedings of the ACM on Web Conference 2025, W â25, New York, NY, USA, p. 4171â4182. External Links: ISBN 9798400712746, Document Cited by: TABLE I. [18] J. Liu (2023) Toward a two-sided fairness framework in search and recommendation. CHIIR â23, New York, NY, USA, p. 236â246. External Links: ISBN 9798400700354, Document Cited by: TABLE I. [19] Ă . Löfgren and K. Nordblom (2020) A theoretical framework of decision making explaining the mechanisms of nudging. Journal of Economic Behavior and Organization 174, p. 1â12. External Links: ISSN 0167-2681, Document Cited by: §I, §I, §I-B, §I-B, §I-C, TABLE I, TABLE I, TABLE I, TABLE I. [20] S. Milano, M. Taddeo, and L. Floridi (2020-12) Recommender systems and their ethical challenges. AI and SOCIETY 35, p. . External Links: Document Cited by: §I, §I-C, TABLE I, TABLE I, TABLE I, TABLE I. [21] I. Nahum-Shani, S. Smith, B. Spring, L. Collins, K. Witkiewitz, A. Tewari, and S. Murphy (2016-09) Just-in-time adaptive interventions (jitais) in mobile health: key components and design principles for ongoing health behavior support. Annals of Behavioral Medicine 52, p. . External Links: Document Cited by: §I-B, TABLE I. [22] G. Nyman Internet of behaviors (iob). External Links: Link Cited by: §I. [23] K. Oyibo (2024) ComTech: towards a unified taxonomy of persuasive techniques for persuasive technology design. Computers in Human Behavior Reports 14, p. 100372. External Links: ISSN 2451-9588, Document Cited by: TABLE I. [24] A. K. Purohit and A. Holzer (2019) Functional digital nudges: identifying optimal timing for effective behavior change. CHI EA â19, New York, NY, USA, p. 1â6. External Links: ISBN 9781450359719, Document Cited by: TABLE I, TABLE I. [25] K. Randi and A. Anders (2022-10) The impossible, the unlikely, and the probable nudges: a classification for the design of your next nudge. Technologies 10, p. 110. External Links: Document Cited by: §I, §I-C. [26] V. Singh, N. K. Vishvakarma, and V. Kumar (2024) Unveiling digital manipulation and persuasion in e-commerce: a systematic literature review of dark patterns and digital nudging. Journal of Internet Commerce 23 (2), p. 144â171. External Links: Document Cited by: TABLE I. [27] B. F. Skinner (1938) The behavior of organisms: an experimental analysis. Appleton-Century-Crofts. Cited by: §I. [28] B. F. skinner (1953) Science and human behavior. Macmillan, New York. Cited by: §I. [29] J. Sun, W. Gan, H. Chao, P. S. Yu, and W. Ding (2023) Internet of behaviors: a survey. IEEE Internet of Things Journal 10 (13), p. 11117â11134. External Links: Document Cited by: §I. [30] The eu artificial intelligence act. External Links: Link Cited by: §I, TABLE I. [31] M. Valta and C. Maier (2025-01) Digital nudging: a systematic literature review, taxonomy, and future research directions. SIGMIS Database 56 (1), p. 101â125. External Links: ISSN 0095-0033, Document Cited by: TABLE I, TABLE I, TABLE I, TABLE I, TABLE I. [32] J. B. Watson (1913) Psychology as the behaviorist views it. Vol. 20, p. 158â177. External Links: Document Cited by: §I.