pith. sign in

arxiv: 2605.26807 · v1 · pith:5YPPRXAYnew · submitted 2026-05-26 · 💻 cs.SE · cs.AI

HTMLCure: Turning Browser Experience into State Guided Repair for Interactive HTML

Jiajun Wu , Jian Yang , Tuney Zheng , Wei Zhang , Haowen Wang , Yihang Lou , Xianglong Liu This is my paper

Pith reviewed 2026-06-29 15:53 UTC · model grok-4.3

classification 💻 cs.SE cs.AI
keywords HTML generationinteractive web pagesbrowser evaluationstate-guided repairLLM fine-tuningVLM feedbackweb interaction testing
0
0 comments X

The pith

Browser-executed interaction states guide repairs that turn LLM-generated HTML into reliably functional pages.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

LLMs often output HTML that renders correctly at first glance but breaks when users scroll, hover, click, or resize the page. HTMLCure runs each page through a browser across multiple viewports and states, captures deterministic evidence from those runs, and feeds curated keyframes to a repair process. The repair engine selects state-specific families of fixes, re-executes candidates, and keeps only the pages that pass the interaction checks. From an initial 97K-prompt corpus this yields 40K quality-cleared pages for supervised fine-tuning. The resulting 27B model matches strong reference systems on HTMLBench-400 and raises average score by 15.3 points on MiniAppBench validation.

Core claim

HTMLCure evaluates generated HTML after the browser has executed it across interaction states, records deterministic browser evidence, supplies the VLM with trajectory keyframes, and drives a closed-loop repair engine that chooses state-specific repair families; the quality-cleared pages produced this way expand the usable seed set and, when used for SFT, raise interactive-task performance to the level of much larger reference models.

What carries the argument

State-guided repair engine that diagnoses the current page from browser trajectory evidence and selects a state-specific repair family before re-testing.

If this is right

  • The fraction of directly usable pages rises from the original seed to roughly two-thirds of the 97K corpus.
  • Evaluation of HTML output must include executed trajectories rather than static screenshots alone.
  • Repair can be iterated in a closed loop until the page passes the chosen state checks.
  • The same state signal can be reused both for diagnosis and for selecting which repair family to apply.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The method may reduce reliance on aggressive filtering that discards repairable pages.
  • Similar trajectory-based feedback could be applied to other stateful code generation settings such as game logic or UI component libraries.
  • If the repair families are made public they could serve as a reusable library for post-generation HTML polishing.

Load-bearing premise

Browser trajectories and state-specific repairs produce training pages whose quality gains transfer to new interactive tasks rather than only matching the particular test suites used.

What would settle it

Train two 27B models on identical recipes except one uses the HTMLCure-refined 40K set and the other uses the raw seed set, then measure pass rates on a fresh interactive HTML benchmark whose test cases were never seen during repair or evaluation.

Figures

Figures reproduced from arXiv: 2605.26807 by Haowen Wang, Jiajun Wu, Jian Yang, Tuney Zheng, Wei Zhang, Xianglong Liu, Yihang Lou.

Figure 1
Figure 1. Figure 1: Problem setting and full pipeline objective. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: HTMLCURE pipeline. The evaluator gathers browser evidence and curated keyframes, the controller chooses a state specific repair family, and the export stage adds quality cleared pages to the refined SFT candidate pool before the final training export. where vt records visual evidence such as rendered frames, viewport snapshots, and frame differences, and bt records behavioral evidence such as probe outcome… view at source ↗
Figure 3
Figure 3. Figure 3: Corpus level repair effect. Panel A compares the score density before and after repair, Panel [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Policy diagnostic for state aware repair. The figure compares action utility, fixed policies, [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Iteration level repair value. The figure shows how retained checkpoints improve over [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Repair scope across semantic task families. The figure summarizes score gains, lift [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Strategy profile for repairable pages. Panel A compares action families over return, success, [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: State dependent stopping behavior. Panel A locates the first best checkpoint for each [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Real repair trace for Games & Simulations. The card contrasts the original broken gameplay [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Real repair trace for Apps & Tools. The card shows how an early off target application [PITH_FULL_IMAGE:figures/full_fig_p015_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Real repair trace for Data Visualization. The card contrasts a missing analytic structure [PITH_FULL_IMAGE:figures/full_fig_p015_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Real repair trace for Visual Art & Animation. The card shows how browser evidence [PITH_FULL_IMAGE:figures/full_fig_p016_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Real repair trace for 3D/WebGL Scenes. The card keeps the radial map task and shows [PITH_FULL_IMAGE:figures/full_fig_p016_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Real repair trace for Content & Marketing. The card shows how repair turns a coarse [PITH_FULL_IMAGE:figures/full_fig_p017_14.png] view at source ↗
read the original abstract

LLMs can now produce full HTML pages, but many of those pages are only superficially correct: they render once, then fail under scroll, hover, click, resize, or gameplay. Evaluation from screenshots can miss these failures, and filtering discards many pages that are still repairable. We introduce HTMLCure, a browser experience framework that evaluates HTML after the system has interacted with it. The evaluator executes the page across viewports and interaction states, records deterministic browser evidence, and gives the VLM curated keyframes from the executed trajectory rather than isolated screenshots. The same state signal drives a closed loop repair engine: HTMLCure diagnoses the current page, chooses a state specific repair family, runs each candidate again, and exports quality cleared pages for SFT. On a 97K prompt corpus, this expands the directly usable seed into a candidate pool of 63703 quality cleared pages, from which we construct the final refined SFT set of 40K pages. Under the same backbone and training recipe, HTMLCure-27B-Refined reaches 50.6 on HTMLBench-400 with 45.2% deterministic test case pass, placing it in the same performance band as strong reference rows such as Kimi-K2.6 and GPT-5.4. On the released MiniAppBench validation split, it reaches 81.2 average, improving raw 27B SFT by 15.3 points and approaching the level of strong reference systems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces HTMLCure, a browser-based framework for evaluating and repairing interactive HTML pages using executed trajectories and state-specific repair families. From a 97K prompt corpus, it generates a 40K refined SFT dataset, resulting in a 27B model that scores 50.6 on HTMLBench-400 (45.2% deterministic test case pass) and 81.2 on MiniAppBench validation, improving 15.3 points over raw SFT and matching strong baselines like Kimi-K2.6 and GPT-5.4.

Significance. If the benchmark gains reflect genuine improvements in interactive HTML generation rather than alignment to specific evaluation metrics, the approach could offer a practical method for curating higher-quality training data for LLMs in web development tasks by incorporating execution feedback.

major comments (2)
  1. [Abstract] Abstract: The central performance claims (50.6 on HTMLBench-400 with 45.2% pass rate; 81.2 on MiniAppBench) are presented without any description of how the 40K pages were selected from the 63703 candidates, the definition of deterministic test cases, presence of error bars, or controls for potential data leakage between the repair process and the benchmarks. This information is load-bearing for interpreting whether the gains support the claim of improved generalization.
  2. [Abstract] Abstract: The state-specific repair families are described as driven by browser evidence, but no details are given on how these families were chosen or whether their definition was independent of the viewport and interaction states used in the HTMLBench-400 and MiniAppBench evaluations. If the families were derived from the same interaction patterns, the reported improvements could represent metric-specific optimization rather than broader quality gains.
minor comments (1)
  1. The term 'quality cleared pages' could be clarified as 'quality-cleared pages' for readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on the abstract. We agree that additional methodological details are needed for proper interpretation of the results and will revise the abstract accordingly. Our responses to the major comments are below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central performance claims (50.6 on HTMLBench-400 with 45.2% pass rate; 81.2 on MiniAppBench) are presented without any description of how the 40K pages were selected from the 63703 candidates, the definition of deterministic test cases, presence of error bars, or controls for potential data leakage between the repair process and the benchmarks. This information is load-bearing for interpreting whether the gains support the claim of improved generalization.

    Authors: We agree the abstract should summarize these elements. In revision we will add: the 40K set was obtained by further filtering the 63703 quality-cleared pages for interaction-state coverage and prompt diversity; deterministic test cases refer to fixed sequences of browser events with verifiable pass/fail outcomes; error bars are omitted because training used a single seed (computational cost); and benchmark prompts were excluded from the repair and SFT construction pipeline. These points are already elaborated in Sections 3–4; the abstract will now reference them concisely. revision: yes

  2. Referee: [Abstract] Abstract: The state-specific repair families are described as driven by browser evidence, but no details are given on how these families were chosen or whether their definition was independent of the viewport and interaction states used in the HTMLBench-400 and MiniAppBench evaluations. If the families were derived from the same interaction patterns, the reported improvements could represent metric-specific optimization rather than broader quality gains.

    Authors: The families were derived from a separate exploratory pass over failure modes observed in an initial 10K-page corpus unrelated to the benchmark viewports or interaction sequences. Categories (e.g., layout instability on resize, missing event handlers) are intentionally general. We will insert a clarifying clause in the revised abstract and expand the methods section to state the disjoint construction explicitly. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation self-contained against external benchmarks

full rationale

The paper's chain starts from a 97K prompt corpus, applies browser trajectory evaluation and state-specific repair families to produce a 40K curated SFT set, then trains a 27B model and reports scores on separate held-out sets (HTMLBench-400 at 50.6, MiniAppBench validation at 81.2). These performance figures are post-training measurements on benchmarks not used to define the repair families or curation criteria. No equations, self-definitions, or load-bearing self-citations reduce the claimed gains to the inputs by construction. The central result remains an empirical comparison against a raw SFT baseline under identical training.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no explicit free parameters, axioms, or invented entities; the 40K-page selection rule, the definition of quality clearing, and the choice of repair families are not detailed enough to classify.

pith-pipeline@v0.9.1-grok · 5816 in / 1463 out tokens · 59138 ms · 2026-06-29T15:53:15.134085+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

58 extracted references · 20 canonical work pages · 9 internal anchors

  1. [1]

    VIBEPASS: Can vibe coders really pass the vibe check?, 2026

    Srijan Bansal, Jiao Fangkai, Yilun Zhou, Austin Xu, Shafiq Joty, and Semih Yavuz. VIBEPASS: Can vibe coders really pass the vibe check?, 2026. arXiv preprint arXiv:2603.15921

    work page arXiv 2026

  2. [2]

    Teaching Large Language Models to Self-Debug

    Xinyun Chen, Maxwell Lin, Nathanael Schärli, and Denny Zhou. Teaching large language models to self-debug, 2023. arXiv preprint arXiv:2304.05128

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  3. [3]

    Mind2Web: Towards a Generalist Agent for the Web

    Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Samuel Stevens, Boshi Wang, Huan Sun, and Yu Su. Mind2Web: Towards a generalist agent for the web. InAdvances in Neural Information Processing Systems 36, Datasets and Benchmarks Track, 2023. NeurIPS 2023 Spotlight; arXiv preprint arXiv:2306.06070

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  4. [4]

    Textbooks Are All You Need

    Suriya Gunasekar, Yi Zhang, Jyoti Aneja, Caio César Teodoro Mendes, Allie Del Giorno, Sivakanth Gopi, Mojan Javaheripi, Piero Kauffmann, Gustavo de Rosa, Olli Saarikivi, Adil Salim, Shital Shah, Harkirat Singh Behl, Xin Wang, Sébastien Bubeck, Ronen Eldan, Adam Tau- man Kalai, Yin Tat Lee, and Yuanzhi Li. Textbooks are all you need, 2023. arXiv preprint a...

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  5. [5]

    WebVoyager: Building an End-to-End Web Agent with Large Multimodal Models

    Hongliang He, Wenlin Yao, Kaixin Ma, Wenhao Yu, Yong Dai, Hongming Zhang, Zhenzhong Lan, and Dong Yu. WebV oyager: Building an end-to-end web agent with large multimodal mod- els. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, pages 6864–6890, Bangkok, Thailand, 2024. Association for Computational Linguistics. a...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  6. [6]

    Vision2Web: A hierarchical benchmark for visual website development with agent verification,

    Zehai He, Wenyi Hong, Zhen Yang, Ziyang Pan, Mingdao Liu, Xiaotao Gu, and Jie Tang. Vision2Web: A hierarchical benchmark for visual website development with agent verification,

  7. [7]

    arXiv preprint arXiv:2603.26648

    work page arXiv

  8. [8]

    CodeRL: Mastering code generation through pretrained models and deep reinforcement learning

    Hung Le, Yue Wang, Akhilesh Deepak Gotmare, Silvio Savarese, and Steven Chu Hong Hoi. CodeRL: Mastering code generation through pretrained models and deep reinforcement learning. InAdvances in Neural Information Processing Systems 35, 2022. NeurIPS 2022; arXiv preprint arXiv:2207.01780

    work page arXiv 2022

  9. [9]

    WebCoderBench: Benchmarking web application generation with comprehensive and interpretable evaluation metrics, 2026

    Chenxu Liu, Yingjie Fu, Wei Yang, Ying Zhang, and Tao Xie. WebCoderBench: Benchmarking web application generation with comprehensive and interpretable evaluation metrics, 2026. arXiv preprint arXiv:2601.02430

    work page arXiv 2026

  10. [10]

    WebGen-Bench: Evaluating LLMs on generating interactive and functional websites from scratch, 2025

    Zimu Lu, Yunqiao Yang, Houxing Ren, Haotian Hou, Han Xiao, Ke Wang, Weikang Shi, Aojun Zhou, Mingjie Zhan, and Hongsheng Li. WebGen-Bench: Evaluating LLMs on generating interactive and functional websites from scratch, 2025. arXiv preprint arXiv:2505.03733

    work page arXiv 2025

  11. [11]

    Self- refine: Iterative refinement with self-feedback

    Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bodhisattwa Prasad Majumder, Katherine Hermann, Sean Welleck, Amir Yazdanbakhsh, and Peter Clark. Self- refine: Iterative refinement with self-feedback. InAdvances in Neural Information Processing Sy...

    2023

  12. [12]

    Olausson, Jeevana Priya Inala, Chenglong Wang, Jianfeng Gao, and Armando Solar- Lezama

    Theo X. Olausson, Jeevana Priya Inala, Chenglong Wang, Jianfeng Gao, and Armando Solar- Lezama. Is self-repair a silver bullet for code generation? InThe Twelfth International Conference on Learning Representations, 2024. ICLR 2024; arXiv preprint arXiv:2306.09896

    work page arXiv 2024

  13. [13]

    Explorer: Scaling exploration-driven web trajectory synthesis for multimodal web agents

    Vardaan Pahuja, Yadong Lu, Corby Rosset, Boyu Gou, Arindam Mitra, Spencer Whitehead, Yu Su, and Ahmed Hassan Awadallah. Explorer: Scaling exploration-driven web trajectory synthesis for multimodal web agents. InFindings of the Association for Computational Linguistics: ACL 2025, pages 6300–6323, Vienna, Austria, 2025. Association for Computational Linguis...

    work page arXiv 2025

  14. [14]

    Qwen. Qwen3.5. https://huggingface.co/collections/Qwen/qwen35 , 2026. Official Qwen3.5 model collection

    2026

  15. [15]

    Image2Struct: Benchmarking structure extraction for vision-language models

    Josselin Somerville Roberts, Tony Lee, Chi Heem Wong, Michihiro Yasunaga, Yifan Mai, and Percy Liang. Image2Struct: Benchmarking structure extraction for vision-language models. In Advances in Neural Information Processing Systems 37, Datasets and Benchmarks Track, 2024. 10

    2024

  16. [16]

    Reflexion: Language Agents with Verbal Reinforcement Learning

    Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Re- flexion: Language agents with verbal reinforcement learning. InAdvances in Neural Information Processing Systems 36, 2023. NeurIPS 2023; arXiv preprint arXiv:2303.11366

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  17. [17]

    De- sign2Code: Benchmarking multimodal code generation for automated front-end engineering

    Chenglei Si, Yanzhe Zhang, Ryan Li, Zhengyuan Yang, Ruibo Liu, and Diyi Yang. De- sign2Code: Benchmarking multimodal code generation for automated front-end engineering. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (V olume 1: Long Papers), pages ...

    work page arXiv 2025

  18. [18]

    FullFront: Benchmarking MLLMs across the full front-end engineering workflow, 2025

    Haoyu Sun, Huichen Will Wang, Jiawei Gu, Linjie Li, and Yu Cheng. FullFront: Benchmarking MLLMs across the full front-end engineering workflow, 2025. arXiv preprint arXiv:2505.17399

    work page arXiv 2025

  19. [19]

    Vibe Code Bench: Evaluating AI Models on End-to-End Web Application Development

    Hung Tran, Langston Nashold, Rayan Krishnan, Antoine Bigeard, and Alex Gu. Vibe code bench: Evaluating AI models on end-to-end web application development, 2026. arXiv preprint arXiv:2603.04601

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  20. [20]

    Jingyu Xiao, Ming Wang, Man Ho Lam, Yuxuan Wan, Junliang Liu, Yintong Huo, and Michael R. Lyu. DesignBench: A comprehensive benchmark for MLLM-based front-end code generation, 2025. arXiv preprint arXiv:2506.06251

    work page arXiv 2025

  21. [21]

    From code foundation models to agents and applications: A comprehensive survey and practical guide to code intelligence, 2025

    Jian Yang, Xianglong Liu, Weifeng Lv, Ken Deng, Shawn Guo, Lin Jing, Yizhi Li, Shark Liu, Xianzhen Luo, Yuyu Luo, Changzai Pan, Ensheng Shi, Yingshui Tan, Renshuai Tao, Jiajun Wu, Xianjie Wu, Zhenhe Wu, Daoguang Zan, Chenchen Zhang, Wei Zhang, He Zhu, Terry Yue Zhuo, Kerui Cao, Xianfu Cheng, Jun Dong, Shengjie Fang, Zhiwei Fei, Xiangyuan Guan, Qipeng Guo,...

    2025

  22. [22]

    Ma, Yuyang Song, Siwei Wu, Yuwen Li, L

    Jian Yang, Wei Zhang, Shawn Guo, Zhengmao Ye, Lin Jing, Shark Liu, Yizhi Li, Jiajun Wu, Cening Liu, X. Ma, Yuyang Song, Siwei Wu, Yuwen Li, L. Liao, T. Zheng, Ziling Huang, Zelong Huang, Che Liu, Yan Xing, Renyuan Li, Qingsong Cai, Hanxu Yan, Siyue Wang, Shikai Li, Jason Klein Liu, An Huang, Yongsheng Kang, Jinxing Zhang, Chuan Hao, Haowen Wang, Weicheng ...

    2026

  23. [23]

    WebShop: Towards scalable real-world web interaction with grounded language agents, 2023

    Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. WebShop: Towards scalable real-world web interaction with grounded language agents. InAdvances in Neural Information Processing Systems 35, 2022. NeurIPS 2022; arXiv preprint arXiv:2207.01206

    work page arXiv 2022

  24. [24]

    ReAct: Synergizing Reasoning and Acting in Language Models

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R. Narasimhan, and Yuan Cao. ReAct: Synergizing reasoning and acting in language models. InThe Eleventh International Conference on Learning Representations, 2023. ICLR 2023; arXiv preprint arXiv:2210.03629

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  25. [25]

    MiniApp- Bench: Evaluating the shift from text to interactive HTML responses in LLM-powered assistants,

    Zuhao Zhang, Chengyue Yu, Yuante Li, Chenyi Zhuang, Linjian Mo, and Shuai Li. MiniApp- Bench: Evaluating the shift from text to interactive HTML responses in LLM-powered assistants,

  26. [26]

    arXiv preprint arXiv:2603.09652

    work page internal anchor Pith review Pith/arXiv arXiv

  27. [27]

    WebArena: A Realistic Web Environment for Building Autonomous Agents

    Shuyan Zhou, Frank F. Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, Uri Alon, and Graham Neubig. WebArena: A realistic web environment for building autonomous agents. InThe Twelfth International Conference on Learning Representations, 2024. ICLR 2024; arXiv preprint arXiv:2307.13854. 11 A Limita...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  28. [28]

    If a feature is not supported by the Analyst report or probe data, do not assume it exists

    Evidence only. If a feature is not supported by the Analyst report or probe data, do not assume it exists

  29. [29]

    Done items raise the score; broken items penalize more than missing items because they indicate a failed implementation

    Functionality is driven by the requirement checklist. Done items raise the score; broken items penalize more than missing items because they indicate a failed implementation

  30. [30]

    Unresponsive buttons, keyboard failure, gameplay with no state change, or timed-out actions cap interaction

    Interaction is constrained by objective probes. Unresponsive buttons, keyboard failure, gameplay with no state change, or timed-out actions cap interaction

  31. [31]

    Polished but reusable templates should not receive top visual scores without prompt-specific design signals

    Visual design is judged from the Analyst's visual evidence. Polished but reusable templates should not receive top visual scores without prompt-specific design signals

  32. [32]

    Code quality is independent of visual quality and depends on runtime cleanliness, maintainability, event wiring, and implementation organization

  33. [33]

    rendering

    Score conservatively and cite evidence in every reason. Return JSON with the five dimensions and total: { "rendering": {"score": 0, "reason": "..."}, "visual_design": {"score": 0, "reason": "..."}, "functionality": {"score": 0, "reason": "..."}, "interaction": {"score": 0, "reason": "..."}, "code_quality": {"score": 0, "reason": "..."}, "total_score": 0, ...

  34. [34]

    Load the page and observe initial rendering

  35. [35]

    Activate entry points such as Start, Submit, Play, Enter, or OK

  36. [36]

    Test every feature mentioned in the task: - click buttons and interactive elements - fill and submit forms - test navigation, menus, tabs, pagination, and modes - observe and interact with animated or canvas-based content - navigate through multiple states or views

  37. [37]

    Complete at least one full user workflow

  38. [38]

    {interaction_guide} ## Report

    Report unresponsive elements, visual glitches, broken layout, on-page errors, and missing content. {interaction_guide} ## Report

  39. [39]

    Rendering: initial page state and visible failures

  40. [40]

    Feature status: working, partial, broken, or missing

  41. [41]

    Bug list: what happened versus what should have happened

  42. [42]

    patches": [{

    Overall quality: Excellent, Good, Fair, Poor, or Broken. D.6.3 State Aware Repair Agents The repair stage receives structured browser evidence collected from the current page. The controller selects the strategy, while the prompt exposes the current score, evidence, prior attempts, and preservation constraints to the code generating model. State aware rep...

  43. [43]

    canvas is not focusable or lacks tabindex

  44. [44]

    keydown/keyup listeners are attached to the wrong element

  45. [45]

    preventDefault is missing for arrow keys

  46. [46]

    key state is not read inside the game loop If {game_layer}=loop:

  47. [47]

    requestAnimationFrame is never called

  48. [48]

    loop starts only after a user action

  49. [49]

    update or draw throws and stops the loop If {game_layer}=canvas:

  50. [50]

    canvas width or height is zero

  51. [51]

    getContext is missing or called before the DOM is ready

  52. [52]

    clearRect runs without redraw

  53. [53]

    overlay or CSS hides the canvas If {game_layer}=overlay:

  54. [54]

    game-over or modal screen is visible on load

  55. [55]

    start screen cannot be dismissed

  56. [56]

    initial score, lives, or state incorrectly trigger a terminal state If {game_layer}=gameplay:

  57. [57]

    collision, scoring, state machine, timer, level progression, or physics is wrong

  58. [58]

    improved

    input works but game state does not change correctly ## Current HTML ```html {html} ``` Return the requested patch or full HTML according to {output_mode}. Contrastive visual feedback prompt template.After a candidate is run again, HTMLCUREcan compare before/after keyframes and inject the result into the next repair prompt. This turns the loop from blind ...