PaperlayerView exact revision on arXiv

Source-linked AI summary

MAIC-UI: Making Interactive Courseware with Generative UI

Shangqing Tu, Yanjia Li, Keyu Chen, Sichen Zhang, Jifan Yu, Daniel Zhang-Li, Lei Hou, Juanzi Li, Yu Zhang, Huiqin Liu

arXiv:2604.25806v1cs.CLcs.AIcs.HC

TL;DR

Creating interactive STEM courseware requires programming expertise, while existing generative-AI tools often produce static content, struggle with long materials, and lack pedagogical accuracy mechanisms. MAIC-UI provides zero-code authoring with structured analysis, verified generation, and rapid incremental editing; in a three-month deployment, the pilot class achieved 9.21-point STEM gains compared to -2.32 points in control classes.

  • Problem

    Creating interactive STEM courseware traditionally requires substantial programming expertise, while existing generative-AI tools often produce static content, struggle with long materials, and lack pedagogical accuracy mechanisms.

  • Method

    MAIC-UI combines structured multimodal knowledge analysis, a two-stage generate-verify-optimize pipeline, and Click-to-Locate editing with Unified Diff-based incremental generation.

  • Results

    9.21-point STEM gains in the pilot class compared to -2.32 points in control classes were observed during a three-month classroom deployment with 53 high school students.

  • Takeaways & Limitations

    MAIC-UI’s classroom benefits included learning agency and reduced outcome disparities, with larger gains among lower-performing pilot-class students and lower gain variance.

  • Takeaways & Limitations

    MAIC-UI focuses on single-page simulations, and its studies used proxy instructors and a single Chinese public high school, limiting generalizability.

Abstract

from arXiv · show

Creating interactive STEM courseware traditionally requires HTML/CSS/JavaScript expertise, leaving barriers for educators. While generative AI can produce HTML codes, existing tools generate static presentations rather than interactive simulations, struggle with long documents, and lack pedagogical accuracy mechanisms. Furthermore, full regeneration for modifications requires 200--600 seconds, disrupting creative flow. We present MAIC-UI, a zero-code authoring system that enables educators to create and rapidly edit interactive courseware from textbooks, PPTs, and PDFs. MAIC-UI employs: (1) structured knowledge analysis with multi-modal understanding to ensure pedagogical rigor; (2) a two-stage generate-verify-optimize pipeline separating content alignment from visual refinement; and (3) Click-to-Locate editing with Unified Diff-based incremental generation achieving sub-10-second iteration cycles. A controlled lab study with 40 participants shows MAIC-UI reduces editing iterations (4.9 vs. 7.0) and significantly improves learnability and controllability compared to direct Text-to-HTML generation. A three-month classroom deployment with 53 high school students demonstrates that MAIC-UI fosters learning agency and reduces outcome disparities -- the pilot class achieved 9.21-point gains in STEM subjects compared to -2.32 points in control classes. Our code is available at https://github.com/THU-MAIC/MAIC-UI.

1 INTRODUCTION

MAIC-UI is a zero-code authoring system that helps educators create interactive STEM courseware from textbooks, PPTs, and PDFs while addressing accuracy, editing, engagement, and visualization challenges. Its structured analysis, verified two-stage generation, and incremental editing support faster authoring and improved educational outcomes.

  • System motivation and contribution: MAIC-UI enables educators to create interactive courseware from textbooks, PPTs, and PDFs without programming expertise.The system was developed in response to educators’ concerns about knowledge accuracy, editing precision, passive learning, and procedural knowledge visualization.
  • Design goals: Its four design goals target pedagogical accuracy and visual professionalism, precise zero-code editing, active personalized exploration, and procedural knowledge visualization.These goals directly address the challenges identified in the formative study with six teaching-experienced participants.
  • Technical approach: MAIC-UI combines multimodal structured knowledge analysis for documents up to 50 pages with a two-stage generate-verify-optimize pipeline.Stage 1 creates content-aligned interactive simulations, while Stage 2 improves visual quality through layout verification and optimization.
  • Evaluation: Its evaluation included a controlled user study with 40 participants and a three-month classroom deployment involving 53 high school students.The studies assessed authoring efficiency, editing responsiveness, student engagement, and learning outcomes.
  • Results: MAIC-UI reduced authoring time from days to minutes, accelerated edit responses relative to full-regeneration baselines, and improved student engagement and learning outcomes.Teachers reported feeling empowered to create interactive content without programming, while students benefited from personalized, self-paced learning.

2 RELATED WORK

Prior educational LLM and co-creation systems improve content generation, interaction, and inspiration, but often lack scientific rigor, local editability, and mechanisms for learning-by-doing. MAIC-UI addresses these gaps through a two-stage workflow, incremental code updates, and click-to-locate editing.

  • Educational LLM Systems: Educational LLM tools support immersive narratives, pedagogical-question generation, dynamic project-based learning, and multimodal creation.These systems expand educational content production and interaction beyond static guidance.
  • Educational LLM Systems: Existing educational LLM tools often lack scientific rigor and visual professionalism, with hallucination, learning-objective misalignment, and limited parametric interaction [37] [47] [58].Many systems remain presentation-oriented rather than supporting personalized learning-by-doing.
  • Educational LLM Systems: MAIC-UI uses a two-stage workflow and interactive component generation with real-time feedback to improve pedagogical reliability, visual quality, and learner-controlled exploration.The workflow separates pedagogical reliability from visual quality while supporting manipulation-based exploration.
  • Human-AI Co-creation: Co-creation frameworks balance AI inspiration with human authority [38], but limited local editability and slow feedback make fine-grained revisions costly.Single-change-affects-all behavior and slow iteration interrupt educators’ creative flow.
  • End-User Programming: MAIC-UI combines Unified Diff-based incremental updates with click-to-locate editing, enabling precise natural-language revisions without DOM or CSS expertise.The approach reduces file-modification latency from minutes to seconds and avoids unstable full regeneration.

3 FORMATIVE STUDY

The formative study examined educators’ experiences creating and integrating AI-generated interactive courseware, revealing concerns about accuracy, editing control, and student engagement that informed four design goals for MAIC-UI.

  • 3 FORMATIVE STUDY: The study asked how educators navigate AI-assisted courseware creation, articulate instructional requirements, and perceive classroom integration of AI-generated interactive materials.These questions addressed educators’ creation processes, cognitive challenges, and classroom concerns.
  • 3.1 Process: Six senior university students with teaching experience completed one-hour hands-on sessions with MAIC-UI’s initial version followed by semi-structured interviews.Interviews explored creation experience, learning costs, creative amplification, classroom integration, and procedural knowledge visualization.
  • 3.2.1 F1: Knowledge Accuracy Concerns in AI-Generated Content.: Participants reported concerns about incorrect knowledge and mismatches between specified input content and generated output, motivating mechanisms for accuracy and content alignment.Participants specifically questioned how the system represents knowledge and noted that generated knowledge could be incorrect or omit specified content.
  • 3.2.2 F2: Limitations of Current Editing Mechanisms.: Participants needed localized, granular editing without coding, because failed edits and three to four iterations made precise modification difficult.They requested “what you see is what you get” control that targets specific elements without repeated regeneration cycles.
  • 3.2.3 F3: Passive Courseware Fails to Engage Active Exploration.: Participants viewed traditional fixed-content courseware as passive and boring, whereas interactive tools could offer actionable choices that increase participation through hands-on, self-paced exploration.This finding motivates interactive experiences that shift students from passive reception toward active exploration.
  • 3.2.3 F3: Passive Courseware Fails to Engage Active Exploration.: Participants identified a gap between textbook concepts and real-world application, motivating dynamic visualizations that make abstract procedural steps concrete and visible.Such visualizations were intended to bridge practical-learning gaps and reduce cognitive load.
  • 3.3 Design Goals: The findings yielded four design goals: pedagogical scientificity and visual professionalism, hands-on interaction with feedback, localized editing, and procedural-knowledge visualization.The goals correspond to structured knowledge analysis, a generate-verify-optimize pipeline, customizable interaction, and low-cost dynamic visualizations.

4 THE MAIC-UI SYSTEM

MAIC-UI is a web-based, no-code authoring system that transforms PDFs, PPTs, language instructions, or structured pedagogical input into interactive courseware. Its structured analysis, two-stage generation, and incremental Click-to-Locate editing connect pedagogical accuracy with rapid teacher-controlled refinement.

  • System Overview: MAIC-UI integrates structured knowledge analysis, two-stage validated generation, and Click-to-Locate editing to produce pedagogically sound, visually professional interactive courseware.These interconnected stages address knowledge accuracy, visual quality, and code-free editing challenges identified in the formative study.
  • Input and Knowledge Analysis: Teachers can upload PDFs or enter structured content directly, with both modes populating a shared knowledge representation for subsequent generation.PDF processing supports documents up to 50 pages and extracts concepts, objectives, prerequisites, and procedural knowledge through multimodal analysis.
  • Two-Stage Generation: The two-stage pipeline first creates content-aligned simulations, then verifies structure and refines themes, typography, layout, and animation.Stage 1 couples step-by-step process panels with interactive simulations, while Stage 2 applies visual refinement and HTML validation.
  • Click-to-Locate Editing: Unified Diff editing reduces generated output by approximately 90% and completes incremental edits in under 10 seconds, replacing 200-600-second full regeneration.Fuzzy context matching applies changed lines with minimal surrounding context, enabling fine-grained natural-language adjustments without programming.
  • Click-to-Locate Editing: Under-10-second iteration helps teachers maintain creative flow and iterate on courseware without technical friction.The system’s rapid editing capability supports pedagogical creativity by avoiding the several-minute waits associated with full regeneration.

5 EVALUATION

MAIC-UI was evaluated through a controlled lab study and an authentic classroom deployment, comparing its full pipeline with direct Text-to-HTML generation. It improved editing efficiency, usability, STEM learning gains, and the evenness of outcomes, with especially strong benefits for lower-performing students.

  • Lab User Study: MAIC-UI participants typically completed refinement in about 3–7 iterations, compared with 5–9 for the baseline, with Click-to-Locate reducing repeated global regeneration.The baseline directly sent prompts and revision instructions to the model without intermediate processing, structured decomposition, or validation.
  • Lab User Study: MAIC-UI was rated significantly higher than the baseline on two of four learnability and usability items, while the remaining two showed no significant difference.Responses used five-point Likert scores, two-sided Mann–Whitney U tests, n=20 per condition, and p<0.05 as the significance threshold.
  • In-Class Deployment: 9.21 points: Class C1’s average STEM gain exceeded the −2.32-point gain of other classes (p<0.001), while humanities gains were 6.43 points and less pronounced.The STEM comparison involved C1 (n=53) versus other classes (n=493), with Mann–Whitney U=16691.5 and r=0.14; the authors relate the stronger STEM advantage to dynamic presentation and interactive manipulation.
  • In-Class Deployment: 562: C1 had the lowest score-gain variance, compared with 598–1054 for other classes, indicating more evenly distributed improvement.The authors interpret this descriptive pattern as evidence that outcome gaps may have been mitigated, consistent with animated demonstrations externalizing procedural knowledge.
  • In-Class Deployment: 15.46 vs. 12.42: bottom-quartile C1 students gained more in STEM, showed more positive gains (78.6% vs. 63.5%), and had lower gain variance (251.02 vs. 671.34).These results indicate benefits extending beyond already stronger or more self-motivated learners.
  • In-Class Deployment: C1 students and teachers reported that interactive visualization and manipulation increased participation and supported intuitive understanding of processes and parameter changes.Students became more willing to ask questions, while teachers observed previously quiet students opening up in the online learning environment.

6 DISCUSSION

MAIC-UI balances pedagogical control with visual automation and transforms authoring through sub-10-second incremental iteration. Its classroom benefits extended beyond high-performing students, though applicability and generalizability remain limited.

  • Balancing Automation and Pedagogical Control: MAIC-UI separates content validation from visual polish, ensuring pedagogical substance is verified before aesthetics are applied.The design supports transparent communication of AI strengths and weaknesses so teachers can retain pedagogical control.
  • The Importance of Rapid Iteration Cycles: Unified Diff-based incremental generation enables sub-10-second iteration cycles that turn authoring from a batch process into an interactive conversation.Delays exceeding approximately 10 seconds create cognitive discontinuities that disrupt creative flow.
  • Addressing Equity in STEM Learning: Lower-performing pilot-class students showed larger score gains than counterparts in control classes, with substantially lower variance in gains.These findings suggest interactive courseware may help address achievement gaps by making abstract concepts accessible through multiple entry points.
  • Limitations: MAIC-UI focuses on single-page simulations, limiting applicability to extended narratives.The lab study used proxy instructors rather than practicing K-12 teachers, and classroom deployment occurred in a single Chinese public high school, requiring further generalizability research.

7 CONCLUSION … A.2.3 Addressing Equity in STEM Learning.

MAIC-UI enables zero-code creation and rapid editing of interactive STEM courseware while addressing pedagogical rigor, teacher control, creative flow, procedural learning, agency, and equity. Its classroom deployment suggests benefits for both learning engagement and more equitable STEM outcomes.

  • 7 CONCLUSION: MAIC-UI combines structured knowledge analysis, generate-verify-optimize generation, and Click-to-Locate Unified Diff editing to support rigorous, rapid, zero-code interactive courseware authoring.The system enables educators to create courseware without programming and supports sub-10-second iteration cycles.
  • A.1 Implications for Teacher-Facing GenUI Systems: Separating content-aligned simulation creation from visual polish addresses teachers’ need for AI assistance without compromising pedagogical accuracy.The formative study identified workload reduction and pedagogical accuracy as competing requirements for educational GenUI systems.
  • A.1.1 Balancing Automation and Pedagogical Control: Teachers who understood the system’s structured content analysis used its strengths more effectively, supporting transparent communication of AI capabilities and limitations.Future systems could visualize extraction and analysis processes to help teachers compensate for weaknesses.
  • A.1.2 The Importance of Rapid Iteration Cycles: Sub-10-second incremental updates preserve creative flow, whereas 200–600-second modifications can cause teachers to lose their train of thought and abandon refinement.Tighter feedback loops transform authoring from a batch process into an interactive conversation, with implications for other iterative GenUI domains.
  • A.1.3 Zero-Code Editing as End-User Programming: Click-to-Locate lets teachers select rendered elements and describe changes naturally, bridging visual interfaces and underlying code without requiring DOM or CSS knowledge.This design supports end-user programming by letting non-experts specify what they want without writing code while preserving precise intentions.
  • A.2 Implications for Procedural Knowledge Learning in STEM Education; A.2.1 Externalizing Procedural Thinking: Interactive simulations externalize procedural knowledge, making dynamic processes visible and manipulable while reducing working-memory demands for abstract STEM concepts.The deployment suggests especially large benefits for lower-performing students, who showed larger gains and a more even distribution of outcomes.
  • A.2.2 Fostering Learning Agency Through Interactive Exploration: Interactive courseware shifted classroom activity toward student-centered exploration, with students asking more questions and teachers exploring knowledge points with them rather than only presenting.The shift reflects active knowledge construction, but depends on interactions aligned with the underlying procedural structure; poorly designed interactions can distract from learning.
  • A.2.3 Addressing Equity in STEM Learning; A DETAILED DISCUSSION: Lower-performing pilot-class students achieved larger score gains than counterparts in control classes, while gain variance was substantially lower, indicating more evenly distributed benefits.Externalized procedural knowledge and multiple engagement entry points may help address STEM achievement gaps, motivating adaptive scaffolding and personalized interaction pathways.

A.3 Limitations · A.3.1 System Limitations.

MAIC-UI’s limitations concern its single-page scope, possible errors in multimodal content extraction, and the use of proxy instructors in the lab study. The classroom deployment with actual high school teachers and students helps address the latter limitation.

  • A.3 Limitations: MAIC-UI currently generates single-page interactive simulations rather than comprehensive multi-page courseware.This scope supports focused procedural knowledge visualization but limits applicability to topics requiring extended narrative or sequential lesson structures.
  • A.3 Limitations: The single-page focus limits MAIC-UI’s applicability to extended narrative or sequential lesson structures.Future work should examine scaling Click-to-Locate editing to multi-page experiences while preserving rapid iteration cycles.
  • A.3.1 System Limitations.: Vision-language models may occasionally misinterpret specialized notation or complex diagrams during content extraction.This limitation is relevant across diverse STEM disciplines.
  • A.3.1 System Limitations.: Structured analysis prompting reduces extraction errors, but domain-specific fine-tuning or OCR–visual-understanding hybrids may be needed for optimal performance.These approaches are proposed to improve robustness across diverse STEM disciplines.
  • A.3.1 System Limitations.: The lab user study used proxy instructors—graduate students with teaching experience—instead of practicing K-12 teachers.This controlled comparison may not reflect target users’ technical comfort, pedagogical training, or authentic classroom pressures.
  • A.3.1 System Limitations.: The three-month classroom deployment with actual high school teachers and students helps address the lab study’s proxy-instructor limitation.The passage identifies this deployment as a complementary evaluation context, without specifying the extent of the remaining gap.

A.3.2 Evaluation Limitations. · A.3.3 Technical Constraints. · B FORMATIVE STUDY METHODOLOGY

The evaluation is limited by short-term use in a single school and by technical dependencies on stable code structure and internet access. The formative study recruited six experienced university-affiliated teachers-in-training, who completed a hands-on prototype session followed by semi-structured interviews.

  • A.3.2 Evaluation Limitations.: Findings may not generalize beyond one Chinese public high school or its physics-chemistry-biology student population, and longer-term classroom integration remains unexamined.Effects may differ across countries, subjects, student demographics, teaching practices, and student expectations.
  • A.3.3 Technical Constraints.: Unified Diff patching requires stable code structure; substantial structural changes after extensive editing may require full regeneration.Fuzzy context matching handles minor drift but not all major structural changes.
  • A.3.3 Technical Constraints.: Internet connectivity is required for AI model access, limiting use in settings with poor connectivity or strict data-privacy requirements.The authors suggest future local-model or hybrid deployment, but the passage is truncated before specifying the hybrid approach.
  • B.1 Participant Recruitment: Six participants were purposively recruited from local university networks, all with formal teaching experience and no prior exposure to MAIC-UI or similar systems.Participants were senior undergraduates or graduate students from top-tier Chinese universities.
  • B.1 Participant Recruitment: Recruitment used departmental mailing lists, teaching-center announcements, and snowball sampling, followed by screening for teaching experience, disciplinary background, and educational-technology familiarity.The sample included students from computer science, physics, mathematics, chemistry, biology, and engineering.
  • B.1 Participant Recruitment: Participants provided written informed consent and received 200 RMB compensation, approximately $28 USD.This passage documents the study’s consent and compensation procedures.
  • B.2 Interview Protocol: Each participant completed a 1-hour hands-on session with the initial prototype and an approximately 45-minute semi-structured interview.The interview addressed creation experience, perceived learning costs, and creative amplification, among other areas.

B.3 Data Analysis · C EXTENDED PARTICIPANT QUOTES

The data analysis used iterative thematic coding with strong inter-rater reliability, while the extended-quotes section preserves formative-study excerpts shortened or omitted from the main text.

  • B.3 Data Analysis: Two researchers independently coded the interview transcripts using deductive and inductive thematic analysis.
  • B.3 Data Analysis: The preliminary coding scheme was based on the research questions and refined through iterative coding.
  • B.3 Data Analysis: Cohen’s kappa reached 0.82 on 20% of the transcripts, indicating the reported inter-rater reliability.
  • B.3 Data Analysis: Coding discrepancies were resolved through discussion until consensus was reached.
  • C EXTENDED PARTICIPANT QUOTES: The extended-quotes section presents formative-study excerpts abbreviated or omitted from the main text because of space constraints.
  • C EXTENDED PARTICIPANT QUOTES: These excerpts are presented as extended verbatim quotes from the formative study.

C.1 Knowledge Accuracy Concerns … C.4 Theory-Practice Gap

Educators identified knowledge accuracy, content alignment, and fine-grained editing as barriers to reliable interactive courseware. They also emphasized that interactive, visual learning can promote active exploration, intuition, and connections between abstract theory and real-world application.

  • C.1 Knowledge Accuracy Concerns: Teachers reported that generated explanations can miss scientific nuance or contain errors, requiring careful verification and occasional rewriting.One example gave a correct statement of Newton’s laws but omitted nuances students commonly struggle with.
  • C.1 Knowledge Accuracy Concerns: Generated courseware sometimes replaced teacher-provided examples with generic explanations, weakening alignment with instructional materials and student needs.Educators stressed that the system should respect the supplied PDF content rather than infer what matters.
  • C.2 Editing Experience: Editing often required repeated back-and-forth because the system misunderstood requested changes, making several iterations potentially slower than manual work.Participants described needing three or four attempts before a result became usable, with each regeneration adding waiting time.
  • C.2 Editing Experience: The system also struggled with granular layout and visual-hierarchy edits, sometimes changing the wrong element and leaving teachers to accept the output.Detailed procedural restructuring and section-specific adjustments were particularly difficult to control.
  • C.3 Passive vs. Active Learning: Traditional presentations were described as fixed and passive, whereas interactive tools can let students manipulate variables, explore randomized cases, and take ownership of understanding.Teachers associated unique, actionable interactions with greater concentration and more active participation.
  • C.4 Theory-Practice Gap: Textbooks often present idealized situations that leave students unable to connect memorized principles with messy real-world scenarios or flexible applications.Without actively bridging the gap, students may retain knowledge without truly understanding or applying it.
  • C.4 Theory-Practice Gap: Visualization can bridge abstract symbols and complex problem descriptions by making principles and phenomena observable, manipulable, and more intuitive.Examples included seeing force vectors and motion, or experiencing how changing velocity feels beyond reciting v = d/t.

D FULL QUESTIONNAIRE ITEMS · E DETAILED IMPLEMENTATION

The appendix reports the complete lab-study questionnaires, interview prompts, and technical implementation scope for MAIC-UI. It specifies Likert-rated usability and perceived-quality items, participant demographics, interview topics, and implementation details.

  • D FULL QUESTIONNAIRE ITEMS: The appendix provides the complete questionnaire items used in the lab user study.The items cover usability, editing experience, perceived quality, demographics, and post-study interviews.
  • D.1 Post-Task Questionnaire (RQ2: Usability and Editing Experience): Participants rated learnability, time cost, editing controllability, and usage preference on a 5-point Likert scale.The scale ranged from 1 = Strongly Disagree to 5 = Strongly Agree.
  • D.1 Post-Task Questionnaire (RQ2: Usability and Editing Experience): MAIC-UI-only perceived-quality items assessed layout intuitiveness, attention attraction, concept accuracy, content coverage, language clarity, and concept intuitiveness.These items evaluated presentation quality and alignment with source teaching content.
  • D.2 Demographic Questionnaire: The demographic questionnaire asked about academic status, disciplinary background, teaching experience, programming experience, AI-tool experience, and prior courseware creation.Response options included categorical levels such as undergraduate, beginner, intermediate, advanced, and yes/no experience.
  • D.3 Post-Study Interview Guide (Lab Study): The lab-study interview guide asked about helpful and frustrating aspects, editing decisions, expectations, final quality, desired features, and teaching use.Its prompts covered both usability and perceived quality during and after courseware creation.
  • E DETAILED IMPLEMENTATION: The technical appendix states that it provides implementation details of the MAIC-UI system.The supplied passage identifies implementation details as the scope of Section E without enumerating specific mechanisms.

E.1 System Architecture … E.8 Error Handling and Fallbacks

The implementation combines a client-server stack, structured multimodal analysis, staged generation, and targeted incremental editing. Performance optimizations and layered fallbacks support rapid, resilient courseware authoring.

  • E.1 System Architecture: The client uses React 18/TypeScript 5.0 with Zustand, React Query, Tailwind CSS, Monaco Editor, and a sandboxed iframe, while the server uses Python 3.11/FastAPI, SQLite/SQLAlchemy, Celery/Redis, and local or CDN-backed storage.
  • E.2 AI Model Configuration: The configured models are GLM-4.6V with temperature=0.2 and max_tokens=4096, falling back to GLM-4.5V, and GLM-4.7 with temperature=0.3 and max_tokens=8192, falling back to GLM-4.6.
  • E.3 Structured Analysis Prompt: The structured prompt extracts topics, key concepts, objectives, prerequisites, procedures with adjustable parameters, subject area, grade level, and content suited to interactive visualization as valid JSON.
  • E.4 Two-Stage Generation Pipeline: The two-stage pipeline first generates scientifically accurate, responsive HTML/JavaScript simulations with process panels, controls, coupling, and tooltips, then applies themed visual polish while preserving functionality.
  • E.5 Click-to-Locate Implementation: Click-to-Locate tracks selected DOM elements using XPath, CSS selectors, truncated HTML, and bounding boxes, then streams an edit request containing the selector, instruction, and document context.
  • E.6 Unified Diff Processing: Unified Diff processing computes line-level HTML changes and applies removals, additions, and context lines with fuzzy matching to accommodate context drift.
  • E.8 Error Handling and Fallbacks: Graceful degradation returns metadata or manual input after analysis failure, uses reduced-validation or CSS fallbacks for stage failures, retries edits with expanded context, and regenerates fully after 3 failures.If both stages fail, the system returns an emergency template with a user-friendly error message.
Loading 2604.25806v1…