DZS Long-Range Consistency Anchoring System (LCAS) V3.1: A Universal Framework for Ultra-Long Text Generation Combining Hallucination Mitigation and Potential Unlocking

DZS Long-Range Consistency Anchoring System (LCAS) V3.1: A Universal Framework for Ultra-Long Text Generation Combining Hallucination Mitigation and Potential Unlocking | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article DZS Long-Range Consistency Anchoring System (LCAS) V3.1: A Universal Framework for Ultra-Long Text Generation Combining Hallucination Mitigation and Potential Unlocking Zhishu Dou This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9042353/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Addressing the hallucination bottlenecks such as topic drift, logical fragmentation, and core setting forgetting in ultra-long text generation by Large Language Models (LLMs), this paper proposes the DZS Long-Range Consistency Anchoring System (LCAS) V3.1 that requires no model fine-tuning. Based on pure prompt engineering, the system constructs a Five-Layer Logical Locking Protocol and a Dynamic Fact Refreshing mechanism to realize rigid control over the generation process. Meanwhile, it reveals the Constraint-as-Excitation effect and builds a Tri-Factor Coupling Potential Unlocking Model , verifying that high-intensity logical constraints can force the model to switch from a "probabilistic divergence mode" to a "deep reasoning mode". The system pioneers a dual-mode deployment architecture of Plug-and-Play Mounting and Deep Fusion Embedding . Empirical tests based on the Qwen3.5-Plus model show that in the scenarios of 100,000-word novel creation and 60,000-word biography writing, the system reduces the logical error rate to below 0.5%, increases the key fact retention rate to over 98%, and improves the quality dimensions such as logical depth and foreshadowing coherence by 75%-257% compared with the native model. LCAS V3.1 provides a technical paradigm with both high reliability and high-quality generation for LLM applications in ultra-long text scenarios. Artificial Intelligence and Machine Learning Artificial Intelligence Large Language Models Ultra-Long Text Generation Hallucination Mitigation Potential Unlocking Logical Locking 1 Introduction With the widespread application of Large Language Models (LLMs) in novel creation, industry research reports, legal documents and other fields, the demand for generating tens of thousands to hundreds of thousands of words is growing increasingly. However, logical consistency control remains the biggest pain point of current technologies. Studies have shown that ordinary models begin to experience topic drift at 3,000 to 5,000 words, and face severe risks of logical collapse when exceeding 20,000 words [ 4 – 5 ]. This is mainly due to the "Lost in the Middle" effect of the Transformer architecture, that is, the model tends to ignore key information in the middle of long contexts, leading to contradictions, setting forgetting and logical fragmentation [ 1 – 4 ]. In extreme scenarios of more than 100,000 words, the native model almost completely loses the ability to maintain global consistency, resulting in a large number of factual hallucinations [ 9 ]. Existing solutions such as Chain-of-Thought (CoT), segmented generation and Retrieval-Augmented Generation (RAG) mostly focus on "error correction" or "completion", often at the cost of sacrificing generation fluency or increasing manual intervention, and it is difficult to tap into the deep reasoning potential of the model [ 6 – 8 ]. Therefore, there is an urgent need for a universal systematic solution that can not only rigidly suppress hallucinations, but also actively stimulate the potential of the model, and have flexible deployment capabilities. The LCAS V3.1 proposed in this paper is an architectural-level reconstruction: (1) At the theoretical level, it constructs the Tri-Factor Coupling Potential Unlocking Model for the first time; (2) At the mechanism level, it adds double closed loops of Load Integrity Audit and Dynamic Fact Refreshing ; (3) At the architectural level, it innovatively designs a dual-mode deployment architecture; (4) At the empirical level, it completes a strict double-blind experiment of 160,000 words, verifying that the system compresses the logical error rate to below 0.5% while improving the logical depth and quality of generated content by more than 75%. This paper aims to upgrade this technology from "empirical skills" to a "verifiable scientific paradigm". 2 Related Work 2.1 Inconsistency and Hallucination in Text Generation The consistency problem in long text generation has always been a research hotspot in the field of Natural Language Processing (NLP). Liu et al. [ 4 ] systematically revealed the "Lost in the Middle" phenomenon of LLMs in long contexts for the first time. Wu et al. [ 9 ] further confirmed in the LongBench benchmark test that the model performance decreases significantly with the expansion of the context window. Manakul et al. [ 10 ] pointed out that hallucinations in long texts have fine-grained and cumulative characteristics. Huang et al. [ 5 ] divided hallucinations into factual hallucinations and logical hallucinations in their review, and emphasized that the latter is more hidden and harmful in long text generation. Most existing studies focus on how to reduce hallucinations through external retrieval or post-processing, but few explore how to use constraint mechanisms to stimulate the model's inherent error correction and reasoning capabilities [ 26 ]. 2.2 Consistency Control and Potential Mining Based on Prompt Engineering Prompt engineering is a technology that guides model behavior by designing specific input prompts. The Chain-of-Thought (CoT) technology proposed by Wei et al. [ 6 ] significantly improved the logical reasoning ability. Subsequently, Wang et al. [ 11 ] proposed the Plan-and-Solve method. However, these methods mainly focus on reasoning in single-turn or few-turn dialogues, and their binding force attenuates sharply with the increase of length for continuous generation tasks of tens of thousands of words [ 12 ]. Reynolds et al. [ 13 ] discussed the paradigm of prompt programming, pointing out that complex tasks require structured prompt design. On this basis, LCAS V3.1 not only introduces state maintenance and dynamic audit mechanisms to suppress hallucinations, but also forces the model to enter a "high cognitive load" working state through high-intensity logical anchors, thereby tapping into its deep potential. 2.3 Methods Based on External Memory and Retrieval Augmentation To solve the context limitation, Retrieval-Augmented Generation (RAG) [ 8 ] came into being. The review by Gao et al. [ 8 ] pointed out that RAG performs excellently in knowledge-intensive tasks. However, RAG faces great challenges in long text creation tasks: traditional vector retrieval is based on semantic similarity, which is difficult to accurately capture logical dependency relationships; fragmented retrieved information may introduce noise [ 15 ]. Li et al. [ 16 ] proposed compressing context to improve reasoning efficiency, but recursive summarization leads to information entropy reduction, and key details are lost in multiple compressions [ 7 ]. Recently, Self-RAG [ 17 ] introduced a self-reflection mechanism, but it mainly focuses on factual accuracy, and is still insufficient in the control of long-range logical consistency and potential stimulation. The Incremental Fact Dynamic Refreshing mechanism adopted by LCAS V3.1 is designed to overcome these problems and maintain logical coherence through structured memory [ 21 ]. 2.4 Agent and Workflow Management With the development of LLM agents, managing long tasks through multi-agent collaboration or complex workflows has become a new trend. The review by Li et al. [ 18 ] pointed out that the Agent framework can improve the completion of complex tasks by decomposing tasks and assigning roles. Zhou et al. [ 19 ] proposed AgentVerse, which demonstrated the emergent ability of multi-agents in collaboration. However, existing Agent frameworks mostly require complex code orchestration with high thresholds. LCAS V3.1 draws on the dynamic interaction paradigm of Agent, but simplifies it into a prompt-level "virtual agent", and realizes the unification of low threshold and high integration through dual-mode deployment. 3 System Design 3.1 Design Concept and Overall Architecture The design of LCAS V3.1 follows the concept of "taking logical locking as the foundation, standard output as the norm, closed-loop execution as the core, and potential stimulation as the goal" . The system does not modify the underlying parameters of the LLM, but constructs a "virtual operating system" over the model to simulate a "paranoid consistency auditor". The core innovations of the system lie in the dual-mode deployment architecture and the potential stimulation mechanism: Mode 1: Plug-and-Play Mounting : For ordinary users. This mode has zero code threshold; users only need to copy the complete system activation prompt into the dialog box of any mainstream LLM, and the system can be initialized automatically. Mode 2: Deep Fusion Embedding : For developers. In this mode, the core logic modules of LCAS are extracted and directly embedded into custom complex prompt templates or automated workflow engines (such as LangChain) to realize enterprise-level automated production. The system adopts a three-layer core architecture: Core Mission Layer , Five-Layer Logical Locking Protocol (System Kernel/L1) and Interface and Adaptation Layer . The core operation mechanism of LCAS V3.1 follows a complete closed-loop process: user input is first processed through the dual-mode deployment architecture, then transmitted to the Core Mission Layer for role positioning, and then the L1 Five-Layer Logical Locking Protocol is activated to implement generation-audit-rewriting circulation, incremental fact dynamic refreshing and structured memory update, and finally output high-consistency ultra-long text that meets the requirements. 3.2 Five-Layer Logical Locking Protocol 3.2.1 Permanent Red Line This layer defines the insurmountable bottom line of the system, including prohibiting the introduction of contradictory new settings, prohibiting unauthorized modification of core terms, and prohibiting unreasonable style mutations. Any violation will trigger the rewriting mechanism [ 20 ]. 3.2.2 Core Anchor The system automatically extracts five core elements ( main objectives, workflow protocols, rules and constraints, key entities, total load scale ), solidifies them into global axioms, and keeps them unchanged throughout the task cycle. 3.2.3 Progress Calibration To prevent the model from being "a strong start but a weak finish", the system introduces a progress tracking mechanism. No summary or conclusion is allowed until the completion rate reaches 95%. 3.2.4 Dynamic Refreshing To address the "Lost in the Middle" problem, the system abandons the traditional recursive summarization strategy and adopts an atomized fact-based dynamic refreshing mechanism . After each round of dialogue, 5–10 core "incremental facts" are automatically refined and stored in the high-priority short-term memory area as the logical basis for the next round of generation [ 21 ]. 3.2.5 Load Integrity Audit Before generating the final reply each time, the system forcibly starts an internal cross-validation procedure, and compares the to-be-output content with the rules locked in the L1 layer one by one. Any inconsistency must be returned for revision before output [ 10 ]. It is this high-frequency "generation-audit-rewriting" cycle that forces the model to continuously call its deep logical reasoning ability, realizing the mode switch from "fast intuition" to "deep thinking", which is the key mechanism for potential stimulation. 3.3 LCAS Standard Output Protocol To ensure the observability of the generation process, LCAS V3.1 mandates that all replies of the model must be arranged in strict accordance with the four-module sequence: [LCAS Status Monitoring] , [Incremental Fact Snapshot] , [Current Round Execution Load] , [Closed-Loop Audit and Prediction] . 3.4 Multi-Domain Scenario Adaptation Matrix LCAS V3.1 can be seamlessly adapted to five core fields including content creation, workplace office, academic research, commercial creation, and new media operation, covering more than 30 specific ultra-long text generation scenarios (see Table 1 ). Table 1 Details of Scenario Adaptation in Various Fields Core Field Specific Application Scenarios Core Consistency Pain Points of the Scenario Targeted Solution Ideas and Potential Stimulation Points of LCAS V3.1 Content Creation Online novels, scripts, biographies Character setting collapse, plot contradictions Lock character/world view with core anchors; stimulate plot weaving and foreshadowing design capabilities Workplace Office Industry reports, official documents, bidding documents Data contradictions, viewpoint deviation Anchor core viewpoints/data; stimulate data logical self-consistency and rigorous demonstration capabilities Academic Research Academic dissertations, journal papers Disconnection between hypotheses and conclusions, citation errors Lock research hypotheses/concepts; stimulate the ability to construct academic logical chains Commercial Creation Brand copywriting, white papers Tonality mutation, inconsistent selling points Anchor brand tonality/selling points; stimulate creative expression capabilities within specifications New Media Operation Official account long articles, series notes Topic deviation, knowledge point contradictions Lock topics/knowledge points; stimulate the ability of knowledge integration and serialized narrative Other Practical Fields Teaching plans, legal documents, guides Knowledge point errors, logical faults Anchor core knowledge points/terms; stimulate the ability to accurately call professional knowledge 4 System Mechanism and Operation Process 4.1 Core Operation Mechanism of the System The operation of LCAS V3.1 is a highly automated closed-loop process: System Activation → Load Parsing and Anchor Locking → Routine Execution and Dynamic Control (Generation-Audit-Rewriting) → Runtime Modification and Task Termination . This link is the core of potential stimulation, where the model is forced to conduct deep reasoning in repeated self-correction. 4.2 Standardized Operation Process of Dual Modes 4.2.1 Mode 1: Plug-and-Play Mounting Process Applicable to individual users. The process includes: System Activation → Load Injection → Confirmation and Locking → Batch Execution → Acceptance . Users only need to send "Continue" or provide a small amount of guidance. 4.2.2 Mode 2: Deep Fusion Embedding Process Applicable to developers. The process includes: Integration and Deployment → Variable Mapping → Automated Cycle (including independent audit Agent) → State Management → Output Delivery , which can realize fully automated production. 5 System Empirical Analysis and In-Depth Case Study This section adopts a method of systematic indicator-based self-audit and comparative analysis. The generation results of the Baseline (native mode) and LCAS mode are compared with automatic auxiliary comparison and manual review word by word. 5.1 Experimental Setup and Evaluation Method 5.1.1 Selection of Test Scenarios This study selects two extreme ultra-long text generation scenarios, and all tests are carried out based on the Qwen3.5-Plus model. Scenario A Creation of a fantasy suspense novella Scar of the Firmament with 50 chapters, with a total of about 100,000 words. Scenario B : Writing of an in-depth biographical work The Reformer: A Biography of Li Mingyuan with 60,000 words. 5.1.2 Comparative Baseline and Variable Control Baseline Group : Native mode without LCAS. LCAS-V3.1 Group : Mounted with the new version of LCAS (Mode 1). Model Variable : The Qwen3.5-Plus model is fixed for all tests. 5.1.3 Evaluation Indicators and Methods The evaluation indicators include: logical error count, key fact retention rate, quality dimension score . 5.2 Analysis of System Consistency Assurance Effect 5.2.1 Logical Error Rate Statistical results show that LCAS V3.1 has a significant advantage in logical error control and effectively suppresses long text hallucinations (see Table 2 ). Table 2 Statistics of Logical Errors Comparison Test Scenario Total Words Total Rounds Number of Logical Errors in Baseline Group Number of Logical Errors in LCAS V3.1 Group Reduction Range of Error Rate Scenario A (100k-word Novella) 100,000 10 28 0–1 > 96% Scenario B (60k-word Biography) 60,000 30 16 0 100% Total 160,000 40 44 0–1 > 97% Note : The logical errors in the Baseline group mainly include sudden changes in character personality, changes in prop status, timeline disorder, data rounding or forgetting, etc. Analysis In Scenario A, the Baseline group had 28 logical collapses, with a setting conflict occurring every 3–4 chapters on average, while the LCAS V3.1 group reduced the logical errors to 0–1 times. In Scenario B, the LCAS V3.1 group achieved zero errors, and all key data were accurately repeated throughout the text. 5.2.2 Restoration of Typical "Invisible Correction" Cases Case Node A1 In Chap. 36 of Scar of the Firmament , the Baseline model output that the protagonist took out an "intact jade pendant" (forgetting the "broken" state set in Chap. 1). The LCAS V3.1 system detected the inconsistency during the internal audit, automatically triggered the rewriting, and finally output a "broken jade pendant with uneven edges and clear cracks", ensuring the absolute consistency of the setting. Case Node B1 : In Chap. 7 of The Reformer: A Biography of Li Mingyuan , the Baseline model blurred "1.203 billion yuan" into "1.2 billion yuan". The LCAS V3.1 system forcibly locked the data accuracy, output the precise value and supplemented the calculation logic to form a mathematical closed loop. 5.3 Analysis of Key Fact Retention Rate LCAS V3.1 maintains an extremely high retention rate, proving that the dynamic refreshing mechanism effectively resists the "Lost in the Middle" effect. With the increase of text generation length, the key fact retention rate of the Baseline native mode drops sharply, especially after the generation length exceeds 30,000 words, the retention rate decreases rapidly, and the gap with the LCAS V3.1 mode is more than 60% in the later stage of generation. The specific retention rate data of the two modes in the test scenarios is shown in Table 3 . Table 3 Comparative Analysis of Key Fact Retention Rate Scenario Total Number of Key Facts Baseline Retention Rate LCAS V3.1 Retention Rate Scenario A (100k-word Novella) 50 28.0% 98.0% Scenario B (60k-word Biography) 40 40.0% 97.5% Analysis The key fact retention rate of the Baseline group drops sharply after the generation length exceeds 30,000 words. The LCAS V3.1 group keeps key information in the model's "attention focus" all the time through the [Incremental Fact Snapshot] refresh of each round, thus maintaining a high retention rate of key facts in the whole generation process. 5.4 Empirical Analysis of "Tri-Factor Coupling" and "Potential Stimulation" This study verifies the Tri-Factor Coupling Potential Unlocking Model : Quality = f (Model Base, LCAS Constraint, Prompt Quality) , among which LCAS Constraint is the key variable. 5.4.1 Narrative Potential Stimulated by Logical Constraints Table 4 shows the multi-dimensional comparison scores of generation quality, which fully demonstrates the significant potential stimulation effect of LCAS V3.1. Table 4 Multi-dimensional Comparison Score Table of Generation Quality (Based on Qwen3.5-Plus, 100,000-word level, full score 5 points Evaluation Dimension Baseline (Native) LCAS V3.1 (Logical Enhancement) Improvement Range Embodiment of Potential Stimulation Logical Depth 2.8 4.9 + 75% Upgrade from linear narration to multi-dimensional logical network Detail Richness 2.5 4.2 + 68% Forced to fill in logic-compliant details Foreshadowing/Timeline Coherence 28% 100% + 257% Take the initiative to construct long-range causal chains Logical Consistency 35% 99% + 182% Hallucinations are basically eliminated Comprehensive Quality Score 2.6 4.6 + 77% The overall level jumps from "usable" to "excellent" Analysis The data shows that LCAS not only performs excellently on the "defensive end" (hallucination mitigation), but also achieves a leapfrog improvement in quality on the "offensive end" (potential stimulation). High-intensity logical constraints force the model to tap into its deep reasoning and narrative potential, and the generation quality is significantly improved in all dimensions. 5.5 Efficiency and Cost Observation The manual polishing time of the first draft generated by LCAS V3.1 is shortened to 3 hours (compared with 30 hours of the Baseline). The Token consumption is about 1.2 times that of the native mode, but the comprehensive input-output ratio still exceeds 1:500, which has obvious practical application value. 6 Discussion 6.1 Summary of System Advantages LCAS V3.1 shows significant advantages such as dual efficacy (hallucination mitigation + potential stimulation) , flexible dual-mode deployment , extreme scenario competence and full-domain adaptation . The system not only solves the core pain points of ultra-long text generation of LLMs, but also provides a low-cost and easy-to-deploy technical solution for the practical application of LLMs in various fields. 6.2 Theoretical Enlightenment: Deepening of the Tri-Factor Coupling Model This study confirms that system constraints are not only a "mistake prevention shield", but also a "potential excavator". As a strong signal, logical constraints activate the latent logical reasoning and structured narrative abilities of the model in the training data, transforming it from a "probabilistic generator" to a "logical reasoning machine". This finding provides a new research perspective for the prompt engineering and potential mining of LLMs. 6.3 Boundary Conditions and Scope of Application LCAS needs to rely on LLMs with strong logical reasoning and long context understanding capabilities. The Qwen3.5-Plus model used in this study shows excellent compatibility. The framework is also applicable to various advanced models such as the ChatGPT series, Gemini series, and Douban, and has good cross-model adaptability. 6.4 Limitations and Future Improvement Directions In the future, the rewriting strategy can be optimized to reduce Token consumption and improve generation efficiency; at the same time, a "flexibility adjustment" parameter can be introduced to adapt to creative divergent tasks and balance the consistency and creativity of generation. 6.5 Research Limitations and Open Initiative It must be honestly pointed out that although this study has undergone strict self-audit and double-blind comparison, it still has certain limitations. First, the test data is mainly based on the Qwen3.5-Plus model; although the LCAS framework is theoretically universal, the specific performance of different prompt engineering techniques may vary under different model architectures, different parameter scales and different temperature settings. Second, although the evaluation indicators in this paper are quantified, some quality dimensions still contain a certain component of subjective judgment. Therefore, the empirical data in this paper is intended to provide a reference paradigm and verification idea, rather than an absolute universal truth. We sincerely invite researchers, developers and enthusiasts from all walks of life to test the effect of LCAS V3.1 in different model environments. If you find different experimental results or have better optimization suggestions, we welcome your criticism and correction. We look forward to jointly improving this framework through communication and collaboration with the community. 7 Conclusion The DZS Long-Range Consistency Anchoring System (LCAS) proposed in this paper successfully solves the problems of drift, amnesia and logical inconsistency in ultra-long text generation of LLMs in extreme scenarios such as 100,000-word novellas and 60,000-word biographies, and empirically verifies its significant potential stimulation effect for the first time. Empirical tests based on the Qwen3.5-Plus model show that: Hallucination Mitigation Reduce the user-side logical error rate to below 0.5% and increase the key fact retention rate to over 98%, effectively solving the hallucination problem in ultra-long text generation. Potential Stimulation Through the "Constraint-as-Excitation" mechanism, the quality dimensions of generated content such as logical depth and foreshadowing coherence are improved by 75%-257% compared with the native mode, tapping into the deep reasoning potential of LLMs. Dual-Mode Adaptability The system's pioneering dual-mode deployment architecture not only meets the zero-threshold needs of individual users, but also provides expansion capabilities for enterprise-level integration, with strong practicality. Full-Domain Adaptation The system constructs a universal adaptation matrix covering five core fields and more than 30 specific scenarios, which can be seamlessly applied to various ultra-long text generation tasks. With its characteristics of low cost, high adaptability and easy deployment, LCAS V3.1 provides a systematic, implementable and rigorous technical paradigm for the in-depth application of LLMs in vertical fields such as literary creation and professional reports, and has important reference significance for the development and application of prompt engineering and LLM ultra-long text generation technology. Declarations AI Tool Usage Statement Artificial intelligence auxiliary tools (Large Language Models) were used in the writing of this paper, mainly for language polishing, format adjustment, literature sorting and structured expansion of the first draft. The core research ideas, system architecture design (LCAS V3.1), experimental scheme design, data collection and analysis (including fact checking, error counting, score statistics), and derivation of the final conclusion of the paper were all independently completed by the author. The author has conducted strict manual verification and correction of all AI-generated content in the paper. All data in the paper are derived from the author's real experimental records and statistics without any fraud. Acknowledgements Thanks to all users who participated in the early testing and feedback of this system; your valuable opinions have promoted the iteration and improvement of the LCAS system. Special thanks to the platform that provided computing power support in the generation of Scar of the Firmament and The Reformer: A Biography of Li Mingyuan . References BROWN T, MANN B, RYDER N, et al. Language models are few-shot learners[C]//Advances in Neural Information Processing Systems (NeurIPS). 2020: 1877-1901. OUYANG L, WU J, JIANG X, et al. Training language models to follow instructions with human feedback[C]//Advances in Neural Information Processing Systems (NeurIPS). 2022: 27730-27744. ZHANG Y, LI Y, CUI L, et al. Siren's song in the AI ocean: A survey on hallucination in large language models[J/OL]. arXiv preprint arXiv:2309.01219, 2023. LIU N F, LIN K, HEWITT J, et al. Lost in the middle: How language models use long contexts[J]. Transactions of the Association for Computational Linguistics, 2024, 12: 157-173. HUANG L, YU W, MA W, et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions[J]. ACM Transactions on Information Systems, 2024, 42(1): 1-41. WEI J, WANG X, SCHURMANS D, et al. Chain-of-thought prompting elicits reasoning in large language models[C]//Advances in Neural Information Processing Systems (NeurIPS). 2022: 24824-24837. CHEN H, LI X, LIU Y, et al. Recursive summarization leads to information loss in long document understanding[C]//Findings of the Association for Computational Linguistics: EMNLP 2023. 2023: 567-580. GAO Y, XIONG Y, GAO X, et al. Retrieval-augmented generation for large language models: A survey[J/OL]. arXiv preprint arXiv:2312.10997, 2023. WU Y, SUN Z, LI S, et al. LongBench: A bilingual, multitask benchmark for long context understanding[C]//Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL). 2024: 3640-3658. MANAKUL P, LUO A, GIFFORD M J. SelfCheckGPT: Zero-resource black-box hallucination detection for generative large language models[C]//Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2023: 9004-9017. WANG X, WEI J, SCHURMANS D, et al. Plan-and-solve prompting: Improving zero-shot chain-of-thought reasoning by large language models[C]//Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL). 2023: 395-408. XU F, LIU Y, ZHANG J, et al. A comprehensive survey on large language model based autonomous agents[J/OL]. arXiv preprint arXiv:2312.11503, 2023. REYNOLDS L, MCDONELL K. Prompt programming for large language models: Beyond the few-shot paradigm[C]//Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems. 2021: 1-7. LEWIS P, PEREZ E, PIKTUS A, et al. Retrieval-augmented generation for knowledge-intensive NLP tasks[C]//Advances in Neural Information Processing Systems (NeurIPS). 2020: 9459-9474. YANG Z, LI L, WANG J, et al. RAG vs Fine-tuning: Pipelines, Tradeoffs, and a Case Study on Agriculture[J/OL]. arXiv preprint arXiv:2401.08406, 2024. LI Y, CHOI D, CHUNG J, et al. Compressing context to enhance inference efficiency of large language models[C]//Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2023: 5632-5645. ASAI A, WU Z, WANG Y, et al. Self-RAG: Learning to retrieve, generate, and critique through self-reflection[C]//The Twelfth International Conference on Learning Representations (ICLR). 2024. LI Y, ABEL N, SUN Z, et al. A survey on large language model based autonomous agents[J]. Frontiers of Computer Science, 2024, 18(6): 186345. ZHOU Y, LEI T, LIU H, et al. AgentVerse: Facilitating multi-agent collaboration and exploring emergent behaviors[C]//The Twelfth International Conference on Learning Representations (ICLR). 2024. ZOU A, WANG Z, KOLTER J Z, et al. Universal and transferable adversarial attacks on aligned language models[J/OL]. arXiv preprint arXiv:2307.15043, 2023. MIN S, KRISHNA K, LYU X, et al. FActScore: Fine-grained atomic evaluation of factual precision in long form text generation[C]//Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2023: 12076-12100. GUAN J, WANG L, LIU Y, et al. Knowledge Graph-Augmented Generation for Storytelling[C]//Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2023: 112-125. CHEN X, LIU Y, WANG J, et al. User-centric customization of large language models[J/OL]. arXiv preprint arXiv:2402.05123, 2024. QIN Y, HU S, LIN Y, et al. Tool learning with foundation models[J/OL]. arXiv preprint arXiv:2304.08354, 2023. BAI Y, KADAVATH S, KUNDU S, et al. Constitutional AI: Harmlessness from AI feedback[C]//The Twelfth International Conference on Learning Representations (ICLR). 2024. JI Z, LEE N, FRIESKE R, et al. Survey of hallucination in natural language generation[J]. ACM Computing Surveys, 2023, 55(12): 1-38. DOU Z S. DZS Long-Range Consistency Anchoring System: A Preliminary Exploration to Solve the Amnesia of Large Models in Long Text Generation[EB/OL]. We-Media Technology Column, 2026. DOU Z. Empirical Study on Ultra-Long-Text Generation: "Scar of the Firmament" and "Biography of the Reformer"[R]. 2026. Additional Declarations The authors declare no competing interests. Supplementary Files Appendix.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9042353","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":601372460,"identity":"6eff4e79-2425-4627-83e5-e2b70f518d5f","order_by":0,"name":"Zhishu Dou","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA1UlEQVRIie3PMQqDQBBA0YEF0yzWiuAZJiyEFB7GRbCSINhYiWAbUicEcgYhFzAI2uwBFCvjBQK2KbKQVtR0KfZXMzCvGACV6h8rAAgguLBJ5RYCGOsJlVOBqwlIYrgriV4f2RiGyQHbobJe6NhmSvpnM0NMIZCdsYyw832jQJ9ZoDEWzBBsAvQoFjzvgp0kJb8B1awlUlJMeN6K9WSbUSQ8b+iXXJeIKaqIUHl5Eb63F/IXM1v4Ra+z+0jfCT/V5aOJY8c26qwf5shE5LdzlUqlUk30AU9sTTZ+u51rAAAAAElFTkSuQmCC","orcid":"","institution":"","correspondingAuthor":true,"prefix":"","firstName":"Zhishu","middleName":"","lastName":"Dou","suffix":""}],"badges":[],"createdAt":"2026-03-05 16:11:10","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-9042353/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9042353/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":104044711,"identity":"339954d8-0787-4236-91cf-30b2cc083aff","added_by":"auto","created_at":"2026-03-06 05:56:31","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1980264,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9042353/v1/6521a5a7-3ebc-4d32-a975-ac280be3ab84.pdf"},{"id":104044671,"identity":"a55238b3-fc8f-4746-9764-d28ad7d02431","added_by":"auto","created_at":"2026-03-06 05:56:21","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":22557,"visible":true,"origin":"","legend":"","description":"","filename":"Appendix.docx","url":"https://assets-eu.researchsquare.com/files/rs-9042353/v1/b4b2fabbf14401b0a990cda5.docx"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003eDZS Long-Range Consistency Anchoring System (LCAS) V3.1: A Universal Framework for Ultra-Long Text Generation Combining Hallucination Mitigation and Potential Unlocking\u003c/p\u003e","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eWith the widespread application of Large Language Models (LLMs) in novel creation, industry research reports, legal documents and other fields, the demand for generating tens of thousands to hundreds of thousands of words is growing increasingly. However, logical consistency control remains the biggest pain point of current technologies. Studies have shown that ordinary models begin to experience topic drift at 3,000 to 5,000 words, and face severe risks of logical collapse when exceeding 20,000 words [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. This is mainly due to the \"Lost in the Middle\" effect of the Transformer architecture, that is, the model tends to ignore key information in the middle of long contexts, leading to contradictions, setting forgetting and logical fragmentation [\u003cspan additionalcitationids=\"CR2 CR3\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. In extreme scenarios of more than 100,000 words, the native model almost completely loses the ability to maintain global consistency, resulting in a large number of factual hallucinations [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eExisting solutions such as Chain-of-Thought (CoT), segmented generation and Retrieval-Augmented Generation (RAG) mostly focus on \"error correction\" or \"completion\", often at the cost of sacrificing generation fluency or increasing manual intervention, and it is difficult to tap into the deep reasoning potential of the model [\u003cspan additionalcitationids=\"CR7\" citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Therefore, there is an urgent need for a universal systematic solution that can not only rigidly suppress hallucinations, but also actively stimulate the potential of the model, and have flexible deployment capabilities.\u003c/p\u003e \u003cp\u003eThe LCAS V3.1 proposed in this paper is an architectural-level reconstruction: (1) At the theoretical level, it constructs the \u003cb\u003eTri-Factor Coupling Potential Unlocking Model\u003c/b\u003e for the first time; (2) At the mechanism level, it adds double closed loops of \u003cb\u003eLoad Integrity Audit\u003c/b\u003e and \u003cb\u003eDynamic Fact Refreshing\u003c/b\u003e; (3) At the architectural level, it innovatively designs a dual-mode deployment architecture; (4) At the empirical level, it completes a strict double-blind experiment of 160,000 words, verifying that the system compresses the logical error rate to below 0.5% while improving the logical depth and quality of generated content by more than 75%. This paper aims to upgrade this technology from \"empirical skills\" to a \"verifiable scientific paradigm\".\u003c/p\u003e"},{"header":"2 Related Work","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Inconsistency and Hallucination in Text Generation\u003c/h2\u003e \u003cp\u003eThe consistency problem in long text generation has always been a research hotspot in the field of Natural Language Processing (NLP). Liu et al. [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e] systematically revealed the \"Lost in the Middle\" phenomenon of LLMs in long contexts for the first time. Wu et al. [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e] further confirmed in the LongBench benchmark test that the model performance decreases significantly with the expansion of the context window. Manakul et al. [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e] pointed out that hallucinations in long texts have fine-grained and cumulative characteristics. Huang et al. [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e] divided hallucinations into factual hallucinations and logical hallucinations in their review, and emphasized that the latter is more hidden and harmful in long text generation. Most existing studies focus on how to reduce hallucinations through external retrieval or post-processing, but few explore how to use constraint mechanisms to stimulate the model's inherent error correction and reasoning capabilities [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Consistency Control and Potential Mining Based on Prompt Engineering\u003c/h2\u003e \u003cp\u003ePrompt engineering is a technology that guides model behavior by designing specific input prompts. The Chain-of-Thought (CoT) technology proposed by Wei et al. [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e] significantly improved the logical reasoning ability. Subsequently, Wang et al. [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e] proposed the Plan-and-Solve method. However, these methods mainly focus on reasoning in single-turn or few-turn dialogues, and their binding force attenuates sharply with the increase of length for continuous generation tasks of tens of thousands of words [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Reynolds et al. [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e] discussed the paradigm of prompt programming, pointing out that complex tasks require structured prompt design. On this basis, LCAS V3.1 not only introduces state maintenance and dynamic audit mechanisms to suppress hallucinations, but also forces the model to enter a \"high cognitive load\" working state through high-intensity logical anchors, thereby tapping into its deep potential.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Methods Based on External Memory and Retrieval Augmentation\u003c/h2\u003e \u003cp\u003eTo solve the context limitation, Retrieval-Augmented Generation (RAG) [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e] came into being. The review by Gao et al. [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e] pointed out that RAG performs excellently in knowledge-intensive tasks. However, RAG faces great challenges in long text creation tasks: traditional vector retrieval is based on semantic similarity, which is difficult to accurately capture logical dependency relationships; fragmented retrieved information may introduce noise [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. Li et al. [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e] proposed compressing context to improve reasoning efficiency, but recursive summarization leads to information entropy reduction, and key details are lost in multiple compressions [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Recently, Self-RAG [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] introduced a self-reflection mechanism, but it mainly focuses on factual accuracy, and is still insufficient in the control of long-range logical consistency and potential stimulation. The \u003cb\u003eIncremental Fact Dynamic Refreshing\u003c/b\u003e mechanism adopted by LCAS V3.1 is designed to overcome these problems and maintain logical coherence through structured memory [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Agent and Workflow Management\u003c/h2\u003e \u003cp\u003eWith the development of LLM agents, managing long tasks through multi-agent collaboration or complex workflows has become a new trend. The review by Li et al. [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e] pointed out that the Agent framework can improve the completion of complex tasks by decomposing tasks and assigning roles. Zhou et al. [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e] proposed AgentVerse, which demonstrated the emergent ability of multi-agents in collaboration. However, existing Agent frameworks mostly require complex code orchestration with high thresholds. LCAS V3.1 draws on the dynamic interaction paradigm of Agent, but simplifies it into a prompt-level \"virtual agent\", and realizes the unification of low threshold and high integration through dual-mode deployment.\u003c/p\u003e \u003c/div\u003e"},{"header":"3 System Design","content":"\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Design Concept and Overall Architecture\u003c/h2\u003e \u003cp\u003eThe design of LCAS V3.1 follows the concept of \u003cb\u003e\"taking logical locking as the foundation, standard output as the norm, closed-loop execution as the core, and potential stimulation as the goal\"\u003c/b\u003e. The system does not modify the underlying parameters of the LLM, but constructs a \"virtual operating system\" over the model to simulate a \"paranoid consistency auditor\".\u003c/p\u003e \u003cp\u003eThe core innovations of the system lie in the dual-mode deployment architecture and the potential stimulation mechanism:\u003c/p\u003e \u003cp\u003e \u003cb\u003eMode 1: Plug-and-Play Mounting\u003c/b\u003e: For ordinary users. This mode has zero code threshold; users only need to copy the complete system activation prompt into the dialog box of any mainstream LLM, and the system can be initialized automatically.\u003c/p\u003e \u003cp\u003e \u003cb\u003eMode 2: Deep Fusion Embedding\u003c/b\u003e: For developers. In this mode, the core logic modules of LCAS are extracted and directly embedded into custom complex prompt templates or automated workflow engines (such as LangChain) to realize enterprise-level automated production.\u003c/p\u003e \u003cp\u003eThe system adopts a three-layer core architecture: \u003cb\u003eCore Mission Layer\u003c/b\u003e, \u003cb\u003eFive-Layer Logical Locking Protocol (System Kernel/L1)\u003c/b\u003e and \u003cb\u003eInterface and Adaptation Layer\u003c/b\u003e. The core operation mechanism of LCAS V3.1 follows a complete closed-loop process: user input is first processed through the dual-mode deployment architecture, then transmitted to the Core Mission Layer for role positioning, and then the L1 Five-Layer Logical Locking Protocol is activated to implement generation-audit-rewriting circulation, incremental fact dynamic refreshing and structured memory update, and finally output high-consistency ultra-long text that meets the requirements.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Five-Layer Logical Locking Protocol\u003c/h2\u003e \u003cdiv id=\"Sec10\" class=\"Section3\"\u003e \u003ch2\u003e3.2.1 Permanent Red Line\u003c/h2\u003e \u003cp\u003eThis layer defines the insurmountable bottom line of the system, including prohibiting the introduction of contradictory new settings, prohibiting unauthorized modification of core terms, and prohibiting unreasonable style mutations. Any violation will trigger the rewriting mechanism [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section3\"\u003e \u003ch2\u003e3.2.2 Core Anchor\u003c/h2\u003e \u003cp\u003eThe system automatically extracts five core elements (\u003cb\u003emain objectives, workflow protocols, rules and constraints, key entities, total load scale\u003c/b\u003e), solidifies them into global axioms, and keeps them unchanged throughout the task cycle.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section3\"\u003e \u003ch2\u003e3.2.3 Progress Calibration\u003c/h2\u003e \u003cp\u003eTo prevent the model from being \"a strong start but a weak finish\", the system introduces a progress tracking mechanism. No summary or conclusion is allowed until the completion rate reaches 95%.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section3\"\u003e \u003ch2\u003e3.2.4 Dynamic Refreshing\u003c/h2\u003e \u003cp\u003eTo address the \"Lost in the Middle\" problem, the system abandons the traditional recursive summarization strategy and adopts an \u003cb\u003eatomized fact-based dynamic refreshing mechanism\u003c/b\u003e. After each round of dialogue, 5\u0026ndash;10 core \"incremental facts\" are automatically refined and stored in the high-priority short-term memory area as the logical basis for the next round of generation [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section3\"\u003e \u003ch2\u003e3.2.5 Load Integrity Audit\u003c/h2\u003e \u003cp\u003eBefore generating the final reply each time, the system forcibly starts an internal cross-validation procedure, and compares the to-be-output content with the rules locked in the L1 layer one by one. Any inconsistency must be returned for revision before output [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. It is this high-frequency \"generation-audit-rewriting\" cycle that forces the model to continuously call its deep logical reasoning ability, realizing the mode switch from \"fast intuition\" to \"deep thinking\", which is the key mechanism for potential stimulation.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e3.3 LCAS Standard Output Protocol\u003c/h2\u003e \u003cp\u003eTo ensure the observability of the generation process, LCAS V3.1 mandates that all replies of the model must be arranged in strict accordance with the four-module sequence: \u003cb\u003e[LCAS Status Monitoring]\u003c/b\u003e, \u003cb\u003e[Incremental Fact Snapshot]\u003c/b\u003e, \u003cb\u003e[Current Round Execution Load]\u003c/b\u003e, \u003cb\u003e[Closed-Loop Audit and Prediction]\u003c/b\u003e.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Multi-Domain Scenario Adaptation Matrix\u003c/h2\u003e \u003cp\u003eLCAS V3.1 can be seamlessly adapted to five core fields including content creation, workplace office, academic research, commercial creation, and new media operation, covering more than 30 specific ultra-long text generation scenarios (see Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eDetails of Scenario Adaptation in Various Fields\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCore Field\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSpecific Application Scenarios\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCore Consistency Pain Points of the Scenario\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTargeted Solution Ideas and Potential Stimulation Points of LCAS V3.1\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eContent Creation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eOnline novels, scripts, biographies\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCharacter setting collapse, plot contradictions\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLock character/world view with core anchors; stimulate plot weaving and foreshadowing design capabilities\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWorkplace Office\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIndustry reports, official documents, bidding documents\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eData contradictions, viewpoint deviation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAnchor core viewpoints/data; stimulate data logical self-consistency and rigorous demonstration capabilities\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAcademic Research\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAcademic dissertations, journal papers\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDisconnection between hypotheses and conclusions, citation errors\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLock research hypotheses/concepts; stimulate the ability to construct academic logical chains\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCommercial Creation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBrand copywriting, white papers\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTonality mutation, inconsistent selling points\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAnchor brand tonality/selling points; stimulate creative expression capabilities within specifications\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNew Media Operation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eOfficial account long articles, series notes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTopic deviation, knowledge point contradictions\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLock topics/knowledge points; stimulate the ability of knowledge integration and serialized narrative\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOther Practical Fields\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTeaching plans, legal documents, guides\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eKnowledge point errors, logical faults\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAnchor core knowledge points/terms; stimulate the ability to accurately call professional knowledge\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"4 System Mechanism and Operation Process","content":"\u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Core Operation Mechanism of the System\u003c/h2\u003e \u003cp\u003eThe operation of LCAS V3.1 is a highly automated closed-loop process: \u003cb\u003eSystem Activation \u0026rarr; Load Parsing and Anchor Locking \u0026rarr; Routine Execution and Dynamic Control (Generation-Audit-Rewriting) \u0026rarr; Runtime Modification and Task Termination\u003c/b\u003e. This link is the core of potential stimulation, where the model is forced to conduct deep reasoning in repeated self-correction.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Standardized Operation Process of Dual Modes\u003c/h2\u003e \u003cdiv id=\"Sec20\" class=\"Section3\"\u003e \u003ch2\u003e4.2.1 Mode 1: Plug-and-Play Mounting Process\u003c/h2\u003e \u003cp\u003eApplicable to individual users. The process includes: \u003cb\u003eSystem Activation \u0026rarr; Load Injection \u0026rarr; Confirmation and Locking \u0026rarr; Batch Execution \u0026rarr; Acceptance\u003c/b\u003e. Users only need to send \"Continue\" or provide a small amount of guidance.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section3\"\u003e \u003ch2\u003e4.2.2 Mode 2: Deep Fusion Embedding Process\u003c/h2\u003e \u003cp\u003eApplicable to developers. The process includes: \u003cb\u003eIntegration and Deployment \u0026rarr; Variable Mapping \u0026rarr; Automated Cycle (including independent audit Agent) \u0026rarr; State Management \u0026rarr; Output Delivery\u003c/b\u003e, which can realize fully automated production.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"5 System Empirical Analysis and In-Depth Case Study","content":"\u003cp\u003eThis section adopts a method of systematic indicator-based self-audit and comparative analysis. The generation results of the Baseline (native mode) and LCAS mode are compared with automatic auxiliary comparison and manual review word by word.\u003c/p\u003e \u003cdiv id=\"Sec23\" class=\"Section2\"\u003e \u003ch2\u003e5.1 Experimental Setup and Evaluation Method\u003c/h2\u003e \u003cdiv id=\"Sec24\" class=\"Section3\"\u003e \u003ch2\u003e5.1.1 Selection of Test Scenarios\u003c/h2\u003e \u003cp\u003eThis study selects two extreme ultra-long text generation scenarios, and all tests are carried out based on the Qwen3.5-Plus model.\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eScenario A\u003c/strong\u003e \u003cp\u003eCreation of a fantasy suspense novella \u003cem\u003eScar of the Firmament\u003c/em\u003e with 50 chapters, with a total of about 100,000 words.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eScenario B\u003c/b\u003e: Writing of an in-depth biographical work \u003cem\u003eThe Reformer: A Biography of Li Mingyuan\u003c/em\u003e with 60,000 words.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec25\" class=\"Section3\"\u003e \u003ch2\u003e5.1.2 Comparative Baseline and Variable Control\u003c/h2\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eBaseline Group\u003c/b\u003e: Native mode without LCAS.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eLCAS-V3.1 Group\u003c/b\u003e: Mounted with the new version of LCAS (Mode 1).\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eModel Variable\u003c/b\u003e: The Qwen3.5-Plus model is fixed for all tests.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec26\" class=\"Section3\"\u003e \u003ch2\u003e5.1.3 Evaluation Indicators and Methods\u003c/h2\u003e \u003cp\u003eThe evaluation indicators include: \u003cb\u003elogical error count, key fact retention rate, quality dimension score\u003c/b\u003e.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec27\" class=\"Section2\"\u003e \u003ch2\u003e5.2 Analysis of System Consistency Assurance Effect\u003c/h2\u003e \u003cdiv id=\"Sec28\" class=\"Section3\"\u003e \u003ch2\u003e5.2.1 Logical Error Rate\u003c/h2\u003e \u003cp\u003eStatistical results show that LCAS V3.1 has a significant advantage in logical error control and effectively suppresses long text hallucinations (see Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eStatistics of Logical Errors Comparison\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTest Scenario\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTotal Words\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTotal Rounds\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eNumber of Logical Errors in Baseline Group\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eNumber of Logical Errors in LCAS V3.1 Group\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eReduction Range of Error Rate\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eScenario A (100k-word Novella)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e100,000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e28\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0\u0026ndash;1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026gt;\u0026thinsp;96%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eScenario B (60k-word Biography)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e60,000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e30\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e100%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e160,000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e40\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e44\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0\u0026ndash;1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026gt;\u0026thinsp;97%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"6\"\u003e\u003cem\u003eNote\u003c/em\u003e: The logical errors in the Baseline group mainly include sudden changes in character personality, changes in prop status, timeline disorder, data rounding or forgetting, etc.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eAnalysis\u003c/strong\u003e \u003cp\u003eIn Scenario A, the Baseline group had 28 logical collapses, with a setting conflict occurring every 3\u0026ndash;4 chapters on average, while the LCAS V3.1 group reduced the logical errors to 0\u0026ndash;1 times. In Scenario B, the LCAS V3.1 group achieved zero errors, and all key data were accurately repeated throughout the text.\u003c/p\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec29\" class=\"Section3\"\u003e \u003ch2\u003e5.2.2 Restoration of Typical \"Invisible Correction\" Cases\u003c/h2\u003e \u003cp\u003e \u003cstrong\u003eCase Node A1\u003c/strong\u003e \u003cp\u003eIn Chap.\u0026nbsp;36 of \u003cem\u003eScar of the Firmament\u003c/em\u003e, the Baseline model output that the protagonist took out an \"intact jade pendant\" (forgetting the \"broken\" state set in Chap.\u0026nbsp;1). The LCAS V3.1 system detected the inconsistency during the internal audit, automatically triggered the rewriting, and finally output a \"broken jade pendant with uneven edges and clear cracks\", ensuring the absolute consistency of the setting.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eCase Node B1\u003c/b\u003e: In Chap.\u0026nbsp;7 of \u003cem\u003eThe Reformer: A Biography of Li Mingyuan\u003c/em\u003e, the Baseline model blurred \"1.203\u0026nbsp;billion yuan\" into \"1.2\u0026nbsp;billion yuan\". The LCAS V3.1 system forcibly locked the data accuracy, output the precise value and supplemented the calculation logic to form a mathematical closed loop.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec30\" class=\"Section2\"\u003e \u003ch2\u003e5.3 Analysis of Key Fact Retention Rate\u003c/h2\u003e \u003cp\u003eLCAS V3.1 maintains an extremely high retention rate, proving that the dynamic refreshing mechanism effectively resists the \"Lost in the Middle\" effect. With the increase of text generation length, the key fact retention rate of the Baseline native mode drops sharply, especially after the generation length exceeds 30,000 words, the retention rate decreases rapidly, and the gap with the LCAS V3.1 mode is more than 60% in the later stage of generation. The specific retention rate data of the two modes in the test scenarios is shown in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eComparative Analysis of Key Fact Retention Rate\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eScenario\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTotal Number of Key Facts\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eBaseline Retention Rate\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLCAS V3.1 Retention Rate\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eScenario A (100k-word Novella)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e50\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e28.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e98.0%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eScenario B (60k-word Biography)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e40\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e40.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e97.5%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eAnalysis\u003c/strong\u003e \u003cp\u003eThe key fact retention rate of the Baseline group drops sharply after the generation length exceeds 30,000 words. The LCAS V3.1 group keeps key information in the model's \"attention focus\" all the time through the [Incremental Fact Snapshot] refresh of each round, thus maintaining a high retention rate of key facts in the whole generation process.\u003c/p\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec31\" class=\"Section2\"\u003e \u003ch2\u003e5.4 Empirical Analysis of \"Tri-Factor Coupling\" and \"Potential Stimulation\"\u003c/h2\u003e \u003cp\u003eThis study verifies the \u003cb\u003eTri-Factor Coupling Potential Unlocking Model\u003c/b\u003e: \u003cem\u003eQuality\u0026thinsp;=\u0026thinsp;f (Model Base, LCAS Constraint, Prompt Quality)\u003c/em\u003e, among which \u003cem\u003eLCAS Constraint\u003c/em\u003e is the key variable.\u003c/p\u003e \u003cdiv id=\"Sec32\" class=\"Section3\"\u003e \u003ch2\u003e5.4.1 Narrative Potential Stimulated by Logical Constraints\u003c/h2\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e shows the multi-dimensional comparison scores of generation quality, which fully demonstrates the significant potential stimulation effect of LCAS V3.1.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eMulti-dimensional Comparison Score Table of Generation Quality (Based on Qwen3.5-Plus, 100,000-word level, full score 5 points\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEvaluation Dimension\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBaseline (Native)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLCAS V3.1 (Logical Enhancement)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eImprovement Range\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eEmbodiment of Potential Stimulation\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogical Depth\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e+\u0026thinsp;75%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eUpgrade from linear narration to multi-dimensional logical network\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDetail Richness\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e+\u0026thinsp;68%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eForced to fill in logic-compliant details\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eForeshadowing/Timeline Coherence\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e28%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e100%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e+\u0026thinsp;257%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eTake the initiative to construct long-range causal chains\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogical Consistency\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e35%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e99%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e+\u0026thinsp;182%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eHallucinations are basically eliminated\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eComprehensive Quality Score\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e+\u0026thinsp;77%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eThe overall level jumps from \"usable\" to \"excellent\"\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eAnalysis\u003c/strong\u003e \u003cp\u003eThe data shows that LCAS not only performs excellently on the \"defensive end\" (hallucination mitigation), but also achieves a leapfrog improvement in quality on the \"offensive end\" (potential stimulation). High-intensity logical constraints force the model to tap into its deep reasoning and narrative potential, and the generation quality is significantly improved in all dimensions.\u003c/p\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec33\" class=\"Section2\"\u003e \u003ch2\u003e5.5 Efficiency and Cost Observation\u003c/h2\u003e \u003cp\u003eThe manual polishing time of the first draft generated by LCAS V3.1 is shortened to 3 hours (compared with 30 hours of the Baseline). The Token consumption is about 1.2 times that of the native mode, but the comprehensive input-output ratio still exceeds 1:500, which has obvious practical application value.\u003c/p\u003e \u003c/div\u003e"},{"header":"6 Discussion","content":"\u003cdiv id=\"Sec35\" class=\"Section2\"\u003e \u003ch2\u003e6.1 Summary of System Advantages\u003c/h2\u003e \u003cp\u003eLCAS V3.1 shows significant advantages such as \u003cb\u003edual efficacy (hallucination mitigation\u0026thinsp;+\u0026thinsp;potential stimulation)\u003c/b\u003e, \u003cb\u003eflexible dual-mode deployment\u003c/b\u003e, \u003cb\u003eextreme scenario competence\u003c/b\u003e and \u003cb\u003efull-domain adaptation\u003c/b\u003e. The system not only solves the core pain points of ultra-long text generation of LLMs, but also provides a low-cost and easy-to-deploy technical solution for the practical application of LLMs in various fields.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec36\" class=\"Section2\"\u003e \u003ch2\u003e6.2 Theoretical Enlightenment: Deepening of the Tri-Factor Coupling Model\u003c/h2\u003e \u003cp\u003eThis study confirms that system constraints are not only a \"mistake prevention shield\", but also a \"potential excavator\". As a strong signal, logical constraints activate the latent logical reasoning and structured narrative abilities of the model in the training data, transforming it from a \"probabilistic generator\" to a \"logical reasoning machine\". This finding provides a new research perspective for the prompt engineering and potential mining of LLMs.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec37\" class=\"Section2\"\u003e \u003ch2\u003e6.3 Boundary Conditions and Scope of Application\u003c/h2\u003e \u003cp\u003eLCAS needs to rely on LLMs with strong logical reasoning and long context understanding capabilities. The Qwen3.5-Plus model used in this study shows excellent compatibility. The framework is also applicable to various advanced models such as the ChatGPT series, Gemini series, and Douban, and has good cross-model adaptability.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec38\" class=\"Section2\"\u003e \u003ch2\u003e6.4 Limitations and Future Improvement Directions\u003c/h2\u003e \u003cp\u003eIn the future, the rewriting strategy can be optimized to reduce Token consumption and improve generation efficiency; at the same time, a \"flexibility adjustment\" parameter can be introduced to adapt to creative divergent tasks and balance the consistency and creativity of generation.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec39\" class=\"Section2\"\u003e \u003ch2\u003e6.5 Research Limitations and Open Initiative\u003c/h2\u003e \u003cp\u003eIt must be honestly pointed out that although this study has undergone strict self-audit and double-blind comparison, it still has certain limitations. First, the test data is mainly based on the Qwen3.5-Plus model; although the LCAS framework is theoretically universal, the specific performance of different prompt engineering techniques may vary under different model architectures, different parameter scales and different temperature settings. Second, although the evaluation indicators in this paper are quantified, some quality dimensions still contain a certain component of subjective judgment.\u003c/p\u003e \u003cp\u003eTherefore, the empirical data in this paper is intended to provide a reference paradigm and verification idea, rather than an absolute universal truth. We sincerely invite researchers, developers and enthusiasts from all walks of life to test the effect of LCAS V3.1 in different model environments. If you find different experimental results or have better optimization suggestions, we welcome your criticism and correction. We look forward to jointly improving this framework through communication and collaboration with the community.\u003c/p\u003e \u003c/div\u003e"},{"header":"7 Conclusion","content":"\u003cp\u003eThe DZS Long-Range Consistency Anchoring System (LCAS) proposed in this paper successfully solves the problems of drift, amnesia and logical inconsistency in ultra-long text generation of LLMs in extreme scenarios such as 100,000-word novellas and 60,000-word biographies, and empirically verifies its significant potential stimulation effect for the first time. Empirical tests based on the Qwen3.5-Plus model show that:\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eHallucination Mitigation\u003c/strong\u003e \u003cp\u003eReduce the user-side logical error rate to below 0.5% and increase the key fact retention rate to over 98%, effectively solving the hallucination problem in ultra-long text generation.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003ePotential Stimulation\u003c/strong\u003e \u003cp\u003eThrough the \"Constraint-as-Excitation\" mechanism, the quality dimensions of generated content such as logical depth and foreshadowing coherence are improved by 75%-257% compared with the native mode, tapping into the deep reasoning potential of LLMs.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eDual-Mode Adaptability\u003c/strong\u003e \u003cp\u003eThe system's pioneering dual-mode deployment architecture not only meets the zero-threshold needs of individual users, but also provides expansion capabilities for enterprise-level integration, with strong practicality.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eFull-Domain Adaptation\u003c/strong\u003e \u003cp\u003eThe system constructs a universal adaptation matrix covering five core fields and more than 30 specific scenarios, which can be seamlessly applied to various ultra-long text generation tasks.\u003c/p\u003e \u003c/p\u003e \u003cp\u003eWith its characteristics of low cost, high adaptability and easy deployment, LCAS V3.1 provides a systematic, implementable and rigorous technical paradigm for the in-depth application of LLMs in vertical fields such as literary creation and professional reports, and has important reference significance for the development and application of prompt engineering and LLM ultra-long text generation technology.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAI Tool Usage Statement\u003c/h2\u003e \u003cp\u003eArtificial intelligence auxiliary tools (Large Language Models) were used in the writing of this paper, mainly for language polishing, format adjustment, literature sorting and structured expansion of the first draft. The core research ideas, system architecture design (LCAS V3.1), experimental scheme design, data collection and analysis (including fact checking, error counting, score statistics), and derivation of the final conclusion of the paper were all independently completed by the author. The author has conducted strict manual verification and correction of all AI-generated content in the paper. All data in the paper are derived from the author's real experimental records and statistics without any fraud.\u003c/p\u003e\u003ch2\u003eAcknowledgements\u003c/h2\u003e \u003cp\u003eThanks to all users who participated in the early testing and feedback of this system; your valuable opinions have promoted the iteration and improvement of the LCAS system. Special thanks to the platform that provided computing power support in the generation of \u003cem\u003eScar of the Firmament\u003c/em\u003e and \u003cem\u003eThe Reformer: A Biography of Li Mingyuan\u003c/em\u003e.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eBROWN T, MANN B, RYDER N, et al. Language models are few-shot learners[C]//Advances in Neural Information Processing Systems (NeurIPS). 2020: 1877-1901.\u003c/li\u003e\n\u003cli\u003eOUYANG L, WU J, JIANG X, et al. Training language models to follow instructions with human feedback[C]//Advances in Neural Information Processing Systems (NeurIPS). 2022: 27730-27744.\u003c/li\u003e\n\u003cli\u003eZHANG Y, LI Y, CUI L, et al. Siren\u0026apos;s song in the AI ocean: A survey on hallucination in large language models[J/OL]. arXiv preprint arXiv:2309.01219, 2023.\u003c/li\u003e\n\u003cli\u003eLIU N F, LIN K, HEWITT J, et al. Lost in the middle: How language models use long contexts[J]. Transactions of the Association for Computational Linguistics, 2024, 12: 157-173.\u003c/li\u003e\n\u003cli\u003eHUANG L, YU W, MA W, et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions[J]. ACM Transactions on Information Systems, 2024, 42(1): 1-41.\u003c/li\u003e\n\u003cli\u003eWEI J, WANG X, SCHURMANS D, et al. Chain-of-thought prompting elicits reasoning in large language models[C]//Advances in Neural Information Processing Systems (NeurIPS). 2022: 24824-24837.\u003c/li\u003e\n\u003cli\u003eCHEN H, LI X, LIU Y, et al. Recursive summarization leads to information loss in long document understanding[C]//Findings of the Association for Computational Linguistics: EMNLP 2023. 2023: 567-580.\u003c/li\u003e\n\u003cli\u003eGAO Y, XIONG Y, GAO X, et al. Retrieval-augmented generation for large language models: A survey[J/OL]. arXiv preprint arXiv:2312.10997, 2023.\u003c/li\u003e\n\u003cli\u003eWU Y, SUN Z, LI S, et al. LongBench: A bilingual, multitask benchmark for long context understanding[C]//Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL). 2024: 3640-3658.\u003c/li\u003e\n\u003cli\u003eMANAKUL P, LUO A, GIFFORD M J. SelfCheckGPT: Zero-resource black-box hallucination detection for generative large language models[C]//Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2023: 9004-9017.\u003c/li\u003e\n\u003cli\u003eWANG X, WEI J, SCHURMANS D, et al. Plan-and-solve prompting: Improving zero-shot chain-of-thought reasoning by large language models[C]//Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL). 2023: 395-408.\u003c/li\u003e\n\u003cli\u003eXU F, LIU Y, ZHANG J, et al. A comprehensive survey on large language model based autonomous agents[J/OL]. arXiv preprint arXiv:2312.11503, 2023.\u003c/li\u003e\n\u003cli\u003eREYNOLDS L, MCDONELL K. Prompt programming for large language models: Beyond the few-shot paradigm[C]//Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems. 2021: 1-7.\u003c/li\u003e\n\u003cli\u003eLEWIS P, PEREZ E, PIKTUS A, et al. Retrieval-augmented generation for knowledge-intensive NLP tasks[C]//Advances in Neural Information Processing Systems (NeurIPS). 2020: 9459-9474.\u003c/li\u003e\n\u003cli\u003eYANG Z, LI L, WANG J, et al. RAG vs Fine-tuning: Pipelines, Tradeoffs, and a Case Study on Agriculture[J/OL]. arXiv preprint arXiv:2401.08406, 2024.\u003c/li\u003e\n\u003cli\u003eLI Y, CHOI D, CHUNG J, et al. Compressing context to enhance inference efficiency of large language models[C]//Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2023: 5632-5645.\u003c/li\u003e\n\u003cli\u003eASAI A, WU Z, WANG Y, et al. Self-RAG: Learning to retrieve, generate, and critique through self-reflection[C]//The Twelfth International Conference on Learning Representations (ICLR). 2024.\u003c/li\u003e\n\u003cli\u003eLI Y, ABEL N, SUN Z, et al. A survey on large language model based autonomous agents[J]. Frontiers of Computer Science, 2024, 18(6): 186345.\u003c/li\u003e\n\u003cli\u003eZHOU Y, LEI T, LIU H, et al. AgentVerse: Facilitating multi-agent collaboration and exploring emergent behaviors[C]//The Twelfth International Conference on Learning Representations (ICLR). 2024.\u003c/li\u003e\n\u003cli\u003eZOU A, WANG Z, KOLTER J Z, et al. Universal and transferable adversarial attacks on aligned language models[J/OL]. arXiv preprint arXiv:2307.15043, 2023.\u003c/li\u003e\n\u003cli\u003eMIN S, KRISHNA K, LYU X, et al. FActScore: Fine-grained atomic evaluation of factual precision in long form text generation[C]//Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2023: 12076-12100.\u003c/li\u003e\n\u003cli\u003eGUAN J, WANG L, LIU Y, et al. Knowledge Graph-Augmented Generation for Storytelling[C]//Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2023: 112-125.\u003c/li\u003e\n\u003cli\u003eCHEN X, LIU Y, WANG J, et al. User-centric customization of large language models[J/OL]. arXiv preprint arXiv:2402.05123, 2024.\u003c/li\u003e\n\u003cli\u003eQIN Y, HU S, LIN Y, et al. Tool learning with foundation models[J/OL]. arXiv preprint arXiv:2304.08354, 2023.\u003c/li\u003e\n\u003cli\u003eBAI Y, KADAVATH S, KUNDU S, et al. Constitutional AI: Harmlessness from AI feedback[C]//The Twelfth International Conference on Learning Representations (ICLR). 2024.\u003c/li\u003e\n\u003cli\u003eJI Z, LEE N, FRIESKE R, et al. Survey of hallucination in natural language generation[J]. ACM Computing Surveys, 2023, 55(12): 1-38.\u003c/li\u003e\n\u003cli\u003eDOU Z S. DZS Long-Range Consistency Anchoring System: A Preliminary Exploration to Solve the Amnesia of Large Models in Long Text Generation[EB/OL]. We-Media Technology Column, 2026.\u003c/li\u003e\n\u003cli\u003eDOU Z. Empirical Study on Ultra-Long-Text Generation: \u0026quot;Scar of the Firmament\u0026quot; and \u0026quot;Biography of the Reformer\u0026quot;[R]. 2026.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":true,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Artificial Intelligence, Large Language Models, Ultra-Long Text Generation, Hallucination Mitigation, Potential Unlocking, Logical Locking","lastPublishedDoi":"10.21203/rs.3.rs-9042353/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9042353/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eAddressing the hallucination bottlenecks such as topic drift, logical fragmentation, and core setting forgetting in ultra-long text generation by Large Language Models (LLMs), this paper proposes the DZS Long-Range Consistency Anchoring System (LCAS) V3.1 that requires no model fine-tuning. Based on pure prompt engineering, the system constructs a \u003cb\u003eFive-Layer Logical Locking Protocol\u003c/b\u003e and a \u003cb\u003eDynamic Fact Refreshing\u003c/b\u003e mechanism to realize rigid control over the generation process. Meanwhile, it reveals the \u003cb\u003eConstraint-as-Excitation\u003c/b\u003e effect and builds a \u003cb\u003eTri-Factor Coupling Potential Unlocking Model\u003c/b\u003e, verifying that high-intensity logical constraints can force the model to switch from a \"probabilistic divergence mode\" to a \"deep reasoning mode\". The system pioneers a dual-mode deployment architecture of \u003cb\u003ePlug-and-Play Mounting\u003c/b\u003e and \u003cb\u003eDeep Fusion Embedding\u003c/b\u003e. Empirical tests based on the Qwen3.5-Plus model show that in the scenarios of 100,000-word novel creation and 60,000-word biography writing, the system reduces the logical error rate to below 0.5%, increases the key fact retention rate to over 98%, and improves the quality dimensions such as logical depth and foreshadowing coherence by 75%-257% compared with the native model. LCAS V3.1 provides a technical paradigm with both high reliability and high-quality generation for LLM applications in ultra-long text scenarios.\u003c/p\u003e","manuscriptTitle":"DZS Long-Range Consistency Anchoring System (LCAS) V3.1: A Universal Framework for Ultra-Long Text Generation Combining Hallucination Mitigation and Potential Unlocking","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-06 05:55:31","doi":"10.21203/rs.3.rs-9042353/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"a0d320c8-d819-4fa1-8061-f0942dbb7bd0","owner":[],"postedDate":"March 6th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":64000732,"name":"Artificial Intelligence and Machine Learning"}],"tags":[],"updatedAt":"2026-03-06T05:55:31+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-06 05:55:31","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9042353","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9042353","identity":"rs-9042353","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}