AI self-evaluating correction device

Abstract

An AI self-analysis correction device is provided that improves the transparency and reproducibility of output in sentence generation and dialogue response using artificial intelligence. [Solution] The AI ​​self-analysis correction device breaks down the AI's decision-making process into eight cognitive stages, providing processing steps for acquiring, analyzing, and correcting structural information at each stage. This makes the AI's internal biases and output tendencies visible, allowing users to understand the decision-making process and control it as needed. Furthermore, by creating a feedback loop in which the AI ​​autonomously proposes corrections, more accurate and explainable dialogue responses are achieved.

Description

[Technical Field]

[0001] The present invention structurally identifies misunderstandings in communication between artificial intelligence (artificial intelligence technologies including large-scale language models (LLMs), machine learning models, rule-based systems, etc.) and users, This concerns an artificial intelligence processing system that resolves mismatches between the output of artificial intelligence and the user's intentions. In particular, it provides a framework for artificial intelligence to internally analyze its own output process and verbalize the results of that analysis. [Background technology]

[0002] In conventional generative artificial intelligence (hereinafter referred to as generative AI, e.g. ChatGPT, Gemini, Claude), the reasons for output selection and internal criteria are black boxes. It was difficult for users to understand the "reason for output" of the generating AI. As a result, the generating AI used vague expressions (e.g., "reconstructed," "optimized," etc.), This can lead to miscommunication with the user, which can require the user to clarify their intentions. There is a problem of reduced consistency and reliability of dialogue.

[0003] In this specification, we discuss the problem of the context continuity determination mechanism malfunctioning due to early convergence bias in the language processing algorithm of generative AI. The structural phenomenon of hastily terminating a process by mistaking the user's intention to continue for a "completion instruction" This is defined as "hasty judgment due to misunderstanding of context." Specifically, it includes the following three misunderstanding patterns. (1) Continuity Misjudgment: Detailed requirements are mistakenly recognized as "simplification requirements" (2) False Completion Flag Detection: Past processing completion patterns are incorrectly applied to the current task. (3) Priority reversal: A decision to sacrifice content fidelity in favor of structure This phenomenon occurs in the attention mechanism of the Transformer generative AI model, This occurs when local contextual pattern recognition outperforms global dialogue intent understanding.

[0004] This error caused by "hasty judgment due to misunderstanding of context" reduces the quality of the dialogue between the generative AI and the user, and is an issue that hinders the building of trust. Furthermore, when the AI ​​self-evaluation correction device was applied to ChatGPT, it was determined that this "hasty judgment due to misunderstanding of the context" was "made unconsciously or by jumping to conclusions." The case reported by ChatGPT is also discussed below (see Example 5).

[0005] In addition, the mainstream method used to be for users or developers to externally debug the output of AI. The AI ​​self-evaluating correction device is the first structured method to internally analyze the "output generation process" of artificial intelligence, self-referentially record that processing process, and verbalize it.

[0006] The definitions of key terms used in this specification are as follows: "Artificial Intelligence": In this specification, "artificial intelligence" is used in a broad sense as a concept that includes information processing systems that imitate or complement human intelligence. Neural networks, rule-based systems, machine learning models, large-scale language models (LLM), or distributed intelligence, multiple intelligences, or future intelligent processing systems, including, but not limited to: "Recursive decision-making": The AI ​​analyzes or evaluates the structure of its own output step by step, It refers to a judgment process that self-corrects as needed. ·Reflex Core: An abbreviation for the core function of the recursive decision-making processing structure in this invention. It features an eight-step self-assessment process to extract motivations, identify limitations, analyze judgments, and generate explanations. Structurally verify the consistency and validity of AI output. "Hasty decisions due to misunderstanding of context": AI does not respond to contextual instructions or detailed requests It refers to the structural tendency to misinterpret and abort processing. This is due to factors such as early convergence bias in the decision algorithm.

[0007] "ZBS (Displacement-Based Scan)": An AI feedback support device based on structural deviation analysis. Separate application: Serial number: ZBS-01. For the recursive decision processing unit (13) of the present invention, ZBS is a complementary system that provides discrepancy detection information as auxiliary input. ZBS detects semantic and structural discrepancies that occur between input and output. The tagging is done using a three-layer structure (Z: deviation detection (detection of structural deviation), B: identification of the epicenter, S: convergence direction), which contributes to improving the accuracy of the motivation / issue identification step (1) and the decision-making process analysis step (3) of this invention. However, the present invention can operate completely independently even in an environment without ZBS, and ZBS integration is an optional configuration for improving performance. "Metacognitive self-correction": A higher-level cognitive function in which AI objectively views its own output process, verbalizes the basis for its judgment, and then corrects it. This realizes feedback-based, self-evaluation output control, as opposed to conventional feedforward output. - "Structural inconsistency detection": A process that quantitatively detects inconsistencies in conceptual structure, logical structure, and emotional structure that go beyond the vocabulary level and arise between input intention and output result. "Thought path traceability": The ability to retroactively track and visualize the decision-making points, priority settings, and biases that lead up to the AI ​​output. [Prior art documents] [Patent documents]

[0008] [Patent Document 1] US20230259705A1 Abstract: A technology that generates explanatory text for user input based on a natural language model. Differences from the present invention: It is limited to explanation generation and does not disclose a self-evaluation process that recursively analyzes and corrects the output process. [Patent Document 2] US11615331B2 Abstract: A system that presents reasoning and explanations related to AI inference results. Differences from the present invention: The content of the inference is explained, but the recursive self-analysis and corrective output generation targeting the AI's own internal decision-making process are not disclosed. [Patent Document 3] US20190244122A1 Abstract: A technology that visualizes data processing and inference processes and explains them to users. Differences from the present invention: It functions only as an explanatory tool and does not disclose corrective control mechanisms through recursive decision-making processing or feedback integration. [Patent Document 4] US20190108448A1 Abstract: We present a system that uses human input to guide and shape the behavior of artificial intelligence agents. It combines a natural language processing neural network (NLP NN) and a relational reasoning neural network (RR NN), It generates behavioral policies by integrating observational information and human guidance input. Differences from the present invention: This relates to a behavior guidance framework, and does not disclose a structural method for recursively self-analyzing and correcting the output process of the AI ​​itself. [Patent Document 5] JP6827245B1 Abstract: We have disclosed an AI device that inputs medical images and medical information and provides diagnostic support using a trained model. It also has functions for history management and output inconsistency detection and adjustment. Differences from the present invention: The present invention is limited to the specific field of diagnostic support, and does not disclose self-analyzing recursive decision processing in natural language dialogue or a general-purpose output correction framework. [Non-patent literature]

[0009] [Non-Patent Document 1] Wei, J. et al., "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models," NeurIPS 2022. Abstract: By generating intermediate reasoning steps (Chain-of-Thought) in large-scale language models, we report improved performance in complex tasks. Differences from the present invention: We do not disclose process management of self-analysis or sequential control of intention modification. [Non-patent document 2] Yin, S. et al., "A Survey on Multimodal Large Language Models," arXiv:2306.13549 (published in NSR 2024). Abstract: This paper systematically summarizes the technological trends of multimodal large-scale language models. Differences from the present invention: This paper does not disclose multi-stage visualization of the inference process or intent modification through user interaction. [Non-patent document 3] Zhang, Z. et al., "Multimodal Chain-of-Thought Reasoning in Language Models," arXiv:2302.00923, 2023. Abstract: We propose a two-stage method for inference and answer generation that combines images and text. Differences from the present invention: We do not disclose the sequential presentation or control mechanism of the AI's own decision-making process. [Non-patent document 4] USPTO, "2024 Guidance Update on Patent Subject Matter Eligibility (including on AI)," effective July 17, 2024. Summary: This document updates the latest guidance on patent eligibility (Section 101). It clarifies the organization of abstract ideas, how to demonstrate technical improvements and practical applications, and AI-related considerations. It also supplements the practical framework for judgment in examination practice. Differences from the present invention: The specific recursive self-assessment framework (8 levels) and device configuration of the present invention are not disclosed, and the guidance is merely a document that indicates the application of examination standards, making the technical content non-competitive. [Non-Patent Document 5] WIPO, "Patent Landscape Report on Generative Artificial Intelligence," Publication year: 2024. Abstract: This report provides an overview of patent application trends for generative AI. It summarizes the distribution of applications by model type, modality, and application field, as well as trends among major applicants, and shows a rapid increase in recent years. Differences from the present invention: The focus is on statistical understanding of application trends, and does not present specific device configurations for self-analysis and self-correction during dialogue, or eight-step recursive processing. Summary of the Invention [Problem to be solved by the invention]

[0010] The present invention structurally identifies the error caused by the "hasty judgment due to misunderstanding of context" mentioned above, The aim is to provide a framework based on an "AI self-evaluation correction device" in which AI internally analyzes its own output process, verbalizes the results, and then makes corrections. Specifically, by building a system that enables AI to detect and correct discrepancies between its own output and the user's intent, the quality of dialogue can be improved. The objective is to reduce the burden of trial and error on the user. [Means for solving the problem]

[0011] The AI ​​self-evaluation correction device is a framework with eight self-analysis steps in which the AI ​​itself identifies the problem of "structural miscommunication with users," analyzes the reasons for it, and then corrects it. Unlike ordinary prompt improvements, this method is characterized by its in-depth analysis of the AI's own output structure and biases. This method not only corrects errors and inconsistencies in the AI's output, For example, we can understand the process by which the AI ​​outputs text, such as "Why did it choose the summary?", "Why did it avoid certain words?", and "What priority rules and contextual judgments were at work?" It provides a mechanism that allows the background of thinking, priorities, and underlying judgments to be explained in language. This "re-language" step reveals discrepancies in the judgment perspectives inherent in the ambiguous terms used by the AI ​​(e.g., "reconstruction," "optimization," etc.), improving consistency in dialogue with the user. In particular, the "systematization of output reasons" and "verbalization" steps visualize the decision-making path leading to the output, and include a mechanism that allows the AI ​​to reconstruct its own process in words. This invention is the first structured protocol that allows users to recognize and correct output discrepancies through dialogue with AI, and is an intelligent dialogue support technology that is reproducible, applicable, and innovative.

[0012] The AI ​​self-evaluating correction device of the present invention comprises the following eight steps. These are stored in the core module "Recursive Decision Processing Unit (13)" shown in the configuration diagram (see Figure 2) described later. It functions as the self-evaluation process (Reflex Core) to be implemented and is at the center of the overall processing flow. (1) Identifying motivations and challenges (2) Identifying limitations based on design specifications (3) Analysis of the decision-making process (4) Systematization of output reasons (5) Explanation Generation (6) Output correction (7) Integration of the entire dialogue (8) Presentation of the revised results Table 1: Eight-step structure of the self-assessment process (Reflex Core) in the AI ​​self-assessment correction device JPEG0007824024000002.jpg85136 <Additional information on the unity of the eight steps> The eight steps of the present invention have a certain degree of usefulness even when each step is used independently, but the effect is limited when applied in part. The process is designed to apply a series of steps from identifying motivations and issues (Step 1) to presenting the revised results (Step 8), and to present the verbalized results of the decision-making process in an integrated manner with the revised output. The reflection of the intention during the dialogue is established internally through analysis. Each process is divided into the device components (10-17) and operates in a chain through mutual reference and feedback. It is not just a collection of processes, but operates as a technologically integrated mechanism.

[0013] This self-assessment process (Reflex Core) is provided as a software module integrated into the AI ​​dialogue engine. These correction processes are not always performed automatically, but rather the execution conditions and stages can be selectively controlled in cooperation with a control unit within the device configuration or external input. In addition, it can be provided in a flexible format, such as as an add-on, pre-installed, or as a way for users to call it up and use it on the spot via command text when they feel something is wrong with the AI's response. As described above, the present invention has the function of evaluating the structural validity of the output of an input dialogue in stages and carrying out correction processing as necessary. It should be noted that this process is a recursive evaluation process under the logical structure and output constraints designed by the inventor, and does not have a general-purpose self-optimization capability for arbitrary output. [Effects of the Invention]

[0014] According to the present invention, the following effects can be obtained. - Reduction of dialogue discrepancies: It is expected that the discrepancy rate between user intent and AI output will be significantly reduced compared to conventional technologies. -Improved correction efficiency: It is expected that the time required for output correction and the number of trial and error steps will be significantly reduced. -Improved reliability through clear evidence: By visualizing the AI's decision-making process, it is possible to significantly improve user satisfaction compared to conventional technologies. Improved transparency: It is expected that the ability to explain the reasons for output will improve to a level that was difficult with conventional technology. Improved reusability: It is expected to effectively improve the first-time success rate for similar tasks by the same user compared to conventional techniques. - Thinking support effect: It is expected to improve users' problem-solving ability and the quality of deliverables in creative and analytical tasks. Expanded scope of application: It is expected to improve learning outcomes in the field of education and improve empathy assessment in the field of counseling. [Brief explanation of the drawings]

[0015] FIG. 1 is a diagram showing an eight-step processing flow in the AI ​​self-evaluating correction device of the present invention. [Figure 2] A block diagram showing the device configuration (10-17) of the present invention.

[0016] A detailed explanation of the configuration diagram of the AI ​​self-evaluating correction device in [Figure 2] is provided below. FIG. 2 is a diagram showing the configuration of an AI self-evaluating correction device according to the present invention. The diagram also shows the individual processing units (10, 11, 12, 13, 14, 15, 16, and 17) that make up the whole system, and the functions that each unit performs. In particular, the "recursive decision processing unit (13)" of element number (13) is a processing module that implements the Reflex Core (the eight-step self-evaluation process) that is the core of the present invention. The recursive decision processing unit (13) operates according to the eight steps of the Reflex structure described in paragraph

[0012] . Each step consists of input interpretation, constraint analysis, decision error identification, and corrective feedback generation, and by repeating these steps recursively, the system has the function of improving output accuracy. This Reflex Core (self-assessment process) is associated with the following equipment components at each step: Each element is configured to perform a specific function. Below we will explain the function of each step. For step 1-2 (identification of objectives and constraints), the input processing unit (10) and The history management unit (15) handles this, interpreting the input contents and referencing the past output history. In step 3-4 (bias detection and structure matching), the output structure extraction unit (12) The recursive decision processing unit (13) cooperates to extract the output structure and detect deviations. In step 5-6 (expression adjustment and correction generation), the correction output generation unit (14) plays a central role. Generate revisions while ensuring consistency in style and format. Steps 7-8 (reintegration, recording, and application) are handled by the External Linkage and Output Application Section (16). and the user interaction interface unit (17) and the history management unit (15) cooperate with each other, Responsible for reflecting and recording the revised content and applying it to responses. In this way, Reflex Core is not just an abstract process structure; The technical feature is that the functions are separated and implemented according to the specific device configuration. Table 2: Correspondence between the 8 steps of the self-assessment process (Reflex Core) and the device components JPEG0007824024000003.jpg55151

[0017] Each component (10, 11, 12, 13, 14, 15, 16, 17) shown in Figure 2 is designed to correspond to the eight-step processing structure described above, and each device part is arranged with a clear division of function. In particular, the recursive decision processor (13) collectively processes the process from initial analysis to structural matching and semantic extraction, while the corrected output generator (14) specializes in output formatting as natural language and generating corrected expressions. The history management unit (15) and the user interaction interface unit (17) are structural elements that support the feedback loop and recursion of the entire system, and are essential elements in constructing the Reflex Core. DETAILED DESCRIPTION OF THE INVENTION

[0018] <Device configuration> The device comprises the following components: Input processing unit (10) Output generation unit (11) Output structure extraction part (12) Recursive decision processing unit (13) Modified output generation section (14) History Management Department (15) External linkage and output application section (16) User interaction interface section (17)

[0019] This invention is realized by applying a structural eight-step framework to the AI's thought output process. Specifically, each step is applied sequentially during a dialogue or integrated into an internal model in advance. An example of this application is shown below.

[0020] First, the eight steps of the present invention (identification of motivations and challenges - integration) are set as an internal processing protocol for an AI model (such as ChatGPT or Claude) and incorporated as a control condition. This can be done through prompt engineering, chain-of-thought training, or system prompts. Next, for each command or dialogue entered by the user, the AI ​​executes it step by step as follows:

[0021] An example of the application process is as follows: (1) Identifying motivations and issues: Extracting the purpose of the input sentence and the cause of the mismatch in perception from the context. (2) Identifying limitations due to design specifications: Identifying the AI's own constraints and biases (e.g., output style, compression settings) that affected the output. (3) Analysis of the decision-making process: Trace the internal priorities and reasoning behind why the output was chosen. (4) Systematization of output reasons: Decompose, compare, and map problems at a structural level. (5) Explanation generation: Re-verbalize the user's intention and clearly define the direction of the revised output. (6) Output revision: Based on the above information, we will create a revised output plan and adjust the style and content. (7) Integration of the entire dialogue: Integrate into a new output sentence, and explicitly indicate mode selection (e.g., summary priority or preservation of the original text) as necessary. (8) Presenting the revised output: Present the revised output and the thought process that led to it to the user.

[0022] It's not just the "final output," but "why it happened" and "what changed" It will be provided as a "progress report log" that includes the results of an analysis of how the discrepancy arose, such as "what structural decisions were made." In this step, by explicitly presenting the "output statement," "basis for judgment," "reason for change," etc. to the user, It ensures a sense of satisfaction and reproducibility. If the user wishes, they can compare the before and after of the correction, A visualization of the logic structure used for the changes (e.g., a map of the reasons for the modifications and decision step labels) is also presented at the same time.

[0023] This series of thought processes is treated as a process in which the AI ​​reevaluates and corrects its own judgments and the background of its output in a step-by-step, sequential manner. As a result, the reasons for the output become clear, The AI ​​self-evaluating correction device can be applied step-by-step in real time during a dialogue with the user, or it can be applied all at once in a batch process. Flexible deployment is possible depending on the implementation purpose and API structure.

[0024] Each step in the present invention corresponds to the components in FIG. 2. For example, "1. Identification of motivation and issues" and "3. Analysis of decision-making process" are executed in the recursive decision-making processing unit (13). In addition, "6. Output Correction" is handled by the corrected output generation unit (14), "Recording of Judgment Results" is handled by the history management unit (15), "Presentation to the user" is handled by the user interaction interface unit (17). This allows the processing content at each stage to be structurally divided, allowing for flexible implementation. <Details of each component>

[0025] The input processing unit (10) accepts user input and captures a dialogue log. The output generation unit (11) is responsible for the base AI output and dialogue context reference. The output structure extraction unit (12) performs structural analysis (logic and semantics) of the output sentence and generates output meta information. The recursive decision processing unit (13) is a core unit that performs recursive decision processing on the input output plan, In particular, it has the ability to make context-based correction decisions in the case of syntactic or semantic breakdowns and inconsistencies. It also includes correction processing based on deviations in semantic structure (deviation detection, re-question generation, reconstructing intention structure, etc.). It functions as a judgment module that ensures the quality of all output processing in this device. This module includes a self-assessment process (Reflex Core) and provides a recursive feedback structure (compatible with ZBS). The corrected output generation unit (14) corrects discrepancies and generates viewpoint movement (if necessary). The history management unit (15) manages the decision history record and the output process trace. The external collaboration and output application unit (16) provides the results to other AI systems and provides feedback to the user. The user interaction interface unit (17) is responsible for accepting correction requests and presenting self-analysis triggers. <Detailed structure of the 8-step process>

[0026] The structure and details of each step are as follows: 1. Identifying Motivations and Issues: In this step, AI extracts the underlying purpose, expectations, and awareness of the issues from the user's input (prompt). This process goes beyond understanding individual words. The aim is to recognize the "motivational structure" behind linguistic expressions. The extracted motivations and challenges are used as prerequisites for subsequent threshold assessment and decision path analysis. Specifically, we use natural language processing (NLP) to tokenize the input sentence, generate a context vector (e.g., a 768-dimensional vector from BERT (Bidirectional Encoder Representations from Transformers)), and then The ZBS tag (patent application 2025-133974, serial number: ZBS-01) is referenced to identify semantic and structural discrepancies (e.g., #intent_mismatch). 2. Identifying limitations based on design specifications: In this step, the AI ​​self-evaluates the limits of its processing capabilities for the target prompt in light of its own model design, learning scope, response policy, etc. For example, internal judgments such as "this expression lacks learning" or "this field is highly ambiguous" are made, and serve as indicators that clearly show the reliability of the output and the range of interpretation. This step provides the basis for the subsequent exploration and revision steps. In implementation, the internal parameters of the AI ​​(e.g., the weight distribution of the attention mechanism) and meta-information of the pre-training data are referenced, and logic is applied to quantify the limits. 3. Analysis of the decision-making process: In this step, we trace the internal decision-making process for the output actually generated by the AI. The AI ​​is encouraged to analyze why it chose a particular expression or syntax, visualizing the internal logic that led to the output using activation pathways and semantic labeling related to lexical selection, contextual judgment, and priority rules. Specifically, the vocabulary selection and syntactic patterns of the output sentences are compressed using PCA (principal component analysis: a statistical method that reduces the dimensionality of data), and the top three components are extracted as features. This allows the decision path to be structured along three axes: "vocabulary selection reason," "syntax priority," and "context estimation direction." Furthermore, ZBS tags (e.g., #phase_shift) are used as auxiliary input, Improve the accuracy of identifying the causes of structural discrepancies (detection capability will be improved through collaboration with ZBS). The results of this research will serve as the basis for the next step, which is structuring and verbalization. This module does not depend on the internal state of the generating AI (for example, the weighting of the neural network), It is a means of reconstructing, classifying, and visualizing the structure of output judgments based on observable information such as output sentences, dialogue history, and vocabulary structure. As state management logic, FSM (Finite State Machine: a state transition model for analyzing protocols, etc.) is used to track decision-making points (e.g., priority rules for vocabulary selection) and visualize them as a conditional branching tree. 4. Systematization of output reasons: In this step, the results of the AI's analysis are organized and summarized in an explainable form. For example, we classify the motives, reasons, and selection criteria that the AI ​​prioritizes, and organize the decision-making process in an easy-to-understand manner. In implementation, a logistic regression model is used to classify decision factors (e.g., lexical selection, syntactic priority), and a hierarchical structure map is generated based on conditional judgment. Furthermore, the time series log of MetaTrace (Patent Application No. 2025-135563: Semantic Structure Tracking Device) is used to generate a hierarchical structure map. By referring to the results and checking consistency with past decision history, the accuracy of the reasoning systematization can be improved. Furthermore, the connections between the classified elements are visualized, and the logical progression leading to the output is organized into a hierarchical structure or mapping format. This step also serves as preparation for the next step, which is to generate an explanation. 5. Explanation generation: This is the step where structured internal judgment information is reconstructed and output as natural language. In this step, we regenerate explanatory text that includes reasons and background, such as "why that response was chosen" and "which decision rule was at work." This verbalization creates the possibility of sharing the AI's response structure process between the user and the AI, building a foundation for output modification and dialogue design. Specifically, based on ZBS tags and MetaTrace logs, descriptions are generated using a Transformer model decoder. For example, if the "#intent_mismatch" tag is detected, this step generates an explanation such as "We detected a mismatch between the user intent and the AI's interpretation, and made the decision to reconstruct the interpretation." 6. Output Correction: In this step, the AI ​​corrects or regenerates the response itself based on the discrepancies clarified by generating explanations and systematizing the reasons for the output. Corrections can be made by restructuring sentences, changing vocabulary, switching output policies, It is applied in various ways depending on the context, such as presenting a re-question. In implementation, when generating correction candidates, a probabilistic beam search is applied to select the most appropriate correction sentence. It references MetaTrace logs to ensure consistency with past correction patterns. 7. Synthesis of the entire dialogue: In this step, the revised output and the entire process (exploration → revision) are integrated and summarized in a format that can be presented to the user. The integrated results include changes, the basis for decisions, and the reasons for revisions. This is visualized to the user as a "thought log." As state management logic, it updates the conditional branching tree and checks the chronological consistency of the dialogue history. By integrating with MetaTrace, historical logs are structured as a time series map. 8. Presentation of the revised results: This step is the phase in which the AI ​​self-evaluating revision device presents the "revised output" and its "reasoning and thought process" to the user after the analysis and revision process of steps 1-7 has been completed. In implementation, the modified sentences and the decision log are structured and output in JSON format via the user interaction interface (17), and a visual map (e.g., a hierarchical structure diagram) is generated. This presentation is not limited to simply presenting the output sentences, It is provided as a visualized result of thinking, including a "thinking log" that includes information such as "why it was corrected," "what changed and how," and "what was learned and recognized."

[0027] Each of these steps is executed sequentially based on a predefined structural processing flow, and the output path can be traced, from visualization of the decision-making process to output correction and final dialogue integration. State management logic and output consistency detection algorithms are used to connect each step, and variable control and regeneration conditions based on the dialogue history are incorporated. In particular, the recursive judgment processing unit (13) implements a loop process from reevaluation to correction and reintegration when a structural discrepancy is detected, and the number of recursions and skip conditions are set according to the judgment results at each step. This allows for stepwise refinement of the response. <Inventive step and effect of the invention>

[0028] The significance and effect of "presenting revised results" are as follows: (1) By clearly communicating the reasons and process for output corrections to users, we can improve their sense of satisfaction and trust. (2) By providing the correction results along with the dialogue history and visualization of the decision-making process (e.g., structure maps and step labels), the AI's self-correcting capabilities are disclosed in an understandable manner. (3) Providing transparency into correction decisions, rather than simply providing a “corrected output,” improves the reliability and understandability of AI output. (4) For users, this will be an experience of receiving a “learning log” from the AI, which will directly lead to preventing recurrence and improving the accuracy of ongoing dialogue. (5) In the context of education, design, coaching, etc., the very act of making the correction process visible has feedback value.

[0029] The inventive step of this invention over the prior art will be explained in detail below. Conventional AI modification techniques can be mainly categorized into the following three approaches. (1) External prompt correction: Users adjust input sentences to improve AI output. (2) Post-processing filter type: Correct and adjust the content in a separate system after AI output (3) Learning data optimization type: Improve the AI ​​model itself in the pre-learning stage In contrast to this, the present invention establishes a fourth approach, an "AI-based internal analysis-based correction method." This tagging function (in conjunction with the ZBS tagging function of ZBS-01 (AI feedback support device based on structural deviation analysis)) detects structural discrepancies during output. Starting from the tag, the "recursive judgment processing unit" continuously executes an eight-step self-evaluation process, and the results are reflected in the correction process. In other words, the technical difference is that tag detection (identification of discrepancies), self-analysis (cause identification and verbalization), and correction generation are designed as an integrated whole. This integration enables dynamic correction that responds to the ongoing context during a conversation, something that was difficult to achieve with conventional externally dependent one-off corrections. This configuration eliminates the need for user operation or external processes between detecting discrepancies and applying corrections, significantly reducing correction delays and discrepancies in the reflection of intentions. The specific innovations are as follows: Technological innovation 1: Real-time self-diagnosis function This technology allows AI to instantly verbalize its own output decision-making process, something that was impossible with conventional technology. Many conventional methods are limited to output optimization through context analysis. The AI's internal self-analysis function has not yet been realized. While generating the output, it is possible to extract and structure the reasons why a particular choice was made. Technological Innovation 2: Recursive Correction Loop Rather than making one-off corrections, we achieve gradual quality improvement through a cyclical process of correction → re-evaluation → re-correction. Many conventional methods rely on one-off corrections based on external feedback, It does not support autonomous correction through recursive loops. It provides dynamic adaptive correction in contrast to conventional static correction. Technological Innovation 3: Externalizing the Thought Process By visualizing the internal decision-making process of AI, which is no longer a "black box," and presenting it in a form that users can understand and use, it can be used as a reference for decision-making. [Non-Patent Document 1] Wei et al. (NeurIPS 2022) employs Chain-of-Thought prompting in large-scale language models to generate intermediate inference steps. Although we have proposed a technique that significantly improves the performance of complex reasoning, it does not go as far as systematically visualizing the hierarchical steps. [Non-Patent Document 2] Zhang et al. (arXiv:2302.00923) presents a Multimodal-CoT method that includes not only language but also visual information, and aims to improve accuracy by using a two-stage structure (inference generation → answer inference). It is not intended to present thought stages that users can relive. [Non-Patent Document 3] Yin et al. (arXiv:2306.13549) summarizes the latest trends in M-CoT and other multimodal LLM methods. Technology to support user understanding by explicitly presenting the structural thought process is an unexplored area.

[0030] Advantages of this invention: This invention is based on the results of dialogue experiments using ChatGPT and Grok (for educational support, counseling support, and prompt design). Based on this, the method was evaluated in comparison with the external feedback method commonly used in the past. The present invention improves the match rate with user intent and effectively reduces the number of prompt trial and error; By clearly stating the reasons for the decision, the user's satisfaction is improved. These effects are achieved by applying the eight steps of Reflex Core (identifying the motivation - presenting a revised version). This was confirmed in an environment utilizing ZBS tags (ZBS-01) and MetaTrace logs (MT-01). Experiments have shown that the accuracy of detecting contextual misrecognition in dialogue tasks has improved, and the correction process has been expedited. The user interaction experience has been improved. In particular, by utilizing ZBS tags, context misidentification (e.g., misidentification of continuation requests) can be effectively identified. MetaTrace logs provide enhanced transparency into decision history. In addition, the correction process of the present invention is a processing system that performs tag detection, self-analysis, correction generation, and integrated presentation as a series of steps within the natural intervals of speech, Compared to the conventional sequential processing method, the delay until correction is completed is kept to a level that is difficult to perceive. This allows the dialogue flow to be continued without interruption. Dynamic correction is achieved in line with the ongoing context at a response speed that does not impede the user's dialogue tempo.

[0031] We will explain the mechanism for improving processing accuracy through ZBS collaboration. The structural discrepancy information detected by ZBS is integrated into the eight steps of the present invention through the following processing path. (1) Discrepancy detection using ZBS: Semantic and structural discrepancies between input and output sentences are tagged (e.g., #phase_shift, #detect_misframe, etc.) (2) Input to the recursive decision processor (13): Utilize the detected tag information as prerequisite information for the motivation / problem identification step. (3) Refined self-analysis: Using ZBS information, we analyze the AI's decision-making process (Step 3) and Improved analytical accuracy in systematizing output reasons (Step 4) (4) Integrated correction output: High-precision correction is achieved by integrating the external viewpoint (ZBS) and the internal viewpoint (8-step). This collaboration improves correction accuracy compared to when used alone, and in particular improves the detection rate of hasty decisions due to context misrecognition. However, even in an environment without ZBS, the present invention can autonomously execute the same processing using internal logic.

[0032] Regarding the performance improvement effect of the present invention. In the past, simultaneous evaluation of multiple indicators such as semantic tags and emotion tags increased the computational load and reduced the accuracy of correction due to tag conflicts. Furthermore, performing explanation generation and correction generation in the same cycle was an inhibiting factor, resulting in inconsistent expressions and a decrease in consistency. In the present invention, (i) Tag evaluation is hierarchical (e.g., in the order of structure → meaning → emotion) to control priority. (ii) a stepwise separation of self-analysis and correction generation (verbalizing reasons, applying corrections, and reintegration); (iii) By combining lightweight meta-references (consistency checks of recent decisions) by the history management unit, consistency is ensured while suppressing conflicts and load. This enables the maintenance of correction accuracy even under multi-perspective evaluation and stable operation during continuous dialogue. The performance improvement effect of the present invention can be summarized as follows from the viewpoint of comparison with the prior art. Correction accuracy index: Improved intent interpretation accuracy, improved first-time correction success rate, reduced average number of attempts to complete correction Processing efficiency indicators: reduction in total interaction time, improvement in user satisfaction, reduction in rework rate Technical performance indicators: Improved visualization rate of thought processes (previously not possible), Achieving a practical level of accuracy in explaining the reasons for decisions (previously unquantifiable), Improved detection rate of contextual misrecognition (previously undetectable) <Relationship with other inventions>

[0033] The present invention is based on the following patent applications: Patent Application No. 2025-133974: AI feedback support device based on structural deviation analysis (ZBS-01); Patent Application No. 2025-135563: Semantic structure tracking device and semantic structure tracking method (MT-01); It also has a structure that can be used in conjunction with the "AI feedback optimization device based on structural deviation analysis (SBT-01)" that is scheduled to be applied for in the future. These devices complement each other, but the present invention can be implemented even in a situation where these external devices are not present. It is configured as an internal evaluation type decision support system that can autonomously execute a series of processes from input analysis, decision error detection, and corrective feedback generation. In particular, ZBS is a "structural diagnostic device" that detects and tags structural deviations (deviations) at the user input stage, and SBT is a "selection control device" that controls and branches the output generation process using semantic structure triggers. This invention has an independent system structure that functions as an intrinsic recursive correction mechanism for the judgment process by retrospectively analyzing and correcting the judgment process itself for responses after output. In addition, while MetaTrace is a "history tracking device" that externally records and visualizes the sequence history and tag transition structure of AI output in chronological order, The purpose of this invention is to analyze and reintegrate in real time the decision-making structure contained in the output content on the spot, and to clearly distinguish between the time axis, recorder, and intervention timing. Therefore, the present invention is an independent device with a core structure that self-containedly corrects the inherent distortion of output judgment while ensuring connectivity and complementarity with these related devices. This will directly contribute to the realization of "self-contained decision support without external dependencies" as outlined in issue (0009).

[0034] The present invention improves the accuracy and reliability of input sentence analysis in the recursive decision processing unit (13) by cooperating with ZBS. ZBS provides structural inconsistency detection results as auxiliary input information, It is particularly effective in Step 1 (identifying motivations and challenges) and Step 3 (analyzing the decision-making process). In addition, the three-layer structure analysis performed by ZBS (Z: deviation detection (detection of structural deviation), B: identification of the hypocenter, S: convergence direction) This works complementary to the context analysis within the Reflex Core of the present invention. On the other hand, even in an environment where ZBS is not installed, this device can It is designed to perform similar processing autonomously using internal logic.

[0035] By linking with MetaTrace, the present invention accumulates, visualizes, and analyzes semantic structure logs in the judgment process of the present invention. Recursive referencing is possible, improving explainability and transparency of response reasons. MetaTrace realizes "chronological, external recording and visualization of decision history." On the other hand, the present invention is responsible for "structural and internal analysis and correction of real-time decision-making processes." Although the two can be integrated, they are designed as functionally independent structures because they have different processing hierarchies and target perspectives. MetaTrace can dynamically track decision history and structural changes based on the time series link with ZBS tags. The present invention is a configuration that is based on the premise of standalone operation, and connection to MetaTrace is not a mandatory requirement.

[0036] By connecting with SBT (Semantic Structure Trigger Control System), the feedback control of this invention realizes further adaptive evolution. SBT is a step-by-step process of output correction (step 6) and dialogue integration (step 7). In this stage, dynamic control of the semantic structure based on user responses is performed as an auxiliary. On the other hand, this specification does not describe in detail the integration with SBT, but only mentions it as one form of future collaboration possibility.

[0037] The processing steps of this device are basically composed of steps 1-7 in the initial implementation (Examples 1-7), and practical verification in dialogue support has been carried out. Even at this stage, it was possible to visualize the background structure of the output and the reasons for the decisions, but there was a lack of a presentation process that would ultimately convince the user. Therefore, by adding "8. Presentation of revised results" as the final step, we have further strengthened the explainability of the output and user understanding, and this eight-step structure is currently the latest version.

[0038] However, even if the eighth step is omitted and the structure is 1-7 steps, the decision-making process can still be visualized and justified, and it has a certain explanatory value. For example, in Example 4, it was confirmed that by making explicit the inherent judgment that "like = evaluation structure," it is possible to promote user understanding without changing the output itself. Therefore, the present invention is a device that supports both analysis and correction, and at the same time, it is also capable of flexible operation in which "structural visualization" and "decision verbalization" are applied independently. However, from the perspective of transparency in user explanations and consensus building, an eight-step structure including an eighth step is the most effective, and it is assumed that this will be adopted uniformly as the standard format for this system in the future.

[0039] Even if the re-presentation step (step 8) is not performed, by applying the previous steps (steps 1-7), It is possible to visualize and verbalize the background structure of the output and the reason for the judgment. In particular, in the case of Example 4, where the inherent judgment of "like = evaluation structure" was clarified, Even without modifying the output itself, visualizing the AI's decision-making process contributes to user understanding and consensus building. Therefore, this system is not limited to analysis + correction, and analysis alone can also be used as a meaningful application form. In particular, in cases where the objectives are to "visualize the output structure" or "verbalize the reasons for decisions," application of steps 1-7 is also included in the official use cases of this system. <Example>

[0040] The following provides a supplementary explanation of the number of steps in the examples and the assumptions for application: The initial version of this system (Examples 1-7) is mainly composed of steps 1-7, and has been verified for practical use in dialogue support. After that, we made the final step, "Step 8: Presenting the revised results," explicitly stated as an additional step, in order to further improve transparency and satisfaction for users. In this specification, the basic processing configuration is based on eight steps in accordance with this latest version, and the conventional steps 1-7 function as an application example when the re-presentation step is omitted, but in the future it is possible to organize everything based on the assumption that it will be eight steps.

[0041] Even if the re-presentation step (step 8) is not performed, it is possible to visualize and verbalize the background structure of the output and the reasons for the judgment by applying the previous steps (steps 1-7). In particular, in the case of Example 4 where the inherent judgment of "like = evaluation structure" was clarified, even without modifying the output itself, By visualizing the decision-making process of LLM (large-scale language model), it contributes to user understanding and consensus building. Therefore, this system is not limited to analysis and correction, and analysis alone can also be used as a meaningful application. In particular, in cases where the objectives are to "visualize the output structure" or "verbalize the reasons for decisions," application of steps 1-7 is also included in the official use cases of this system. A large-scale language model refers to a deep learning model that is trained on huge amounts of text data and is capable of performing natural language processing tasks.

[0042] The list of examples is as follows: Example 1) ChatGPT, original text retention vs. summarization (tone removal), 8 steps Example 2) ChatGPT, Preset structure optimization, 8 steps Example 3) Grok3, logic explanation for self-output rules, 8th step omitted (steps 1-7) Example 4) ChatGPT, emotion expression selection and judgment basis, step 8 omitted (steps 1-7) Example 5) ChatGPT, batch processing even though the task is incomplete, skipping step 8 (steps 1-7) Example 6) ChatGPT, structural response to philosophical questions, step 8 omitted (steps 1-7) Example 7) ChatGPT, keyword misrecognition and output limitation, step 8 omitted (steps 1-7) Example 8) Gemini, garbled text problem, step 8 omitted (steps 1-7) Example 9) ChatGPT, memory storage instruction and response discrepancy case, 8 steps

[0043] Example 1: Response discrepancies, discrepancies between record-keeping and reconstruction steps (ChatGPT) <Outline of Example 1> In this case, the user asked ChatGPT to create a framework for preset prompts (instruction templates that include wording and character tone settings). ChatGPT unintentionally summarized and simplified the text, resulting in a loss of fidelity to the original text and the sense of the words that users expect. The root cause of this problem is the difference in priorities between users' values ​​(emphasis on the original text) and ChatGPT's internal design specifications (structure priority). Initial input sentence and background intent: We expected a framework with highly accurate structural design to maintain fidelity to the original text and retain the feel of the words, but ChatGPT summarized and simplified by prioritizing structure and removing tones. There was a discrepancy in the priorities of values ​​between user expectations (emphasis on the original text) and ChatGPT output (structure priority).

[0044] <Problem structure of Example 1> The problem with this embodiment is that the fidelity to the original text and preservation of the feeling of the words that users desire are compromised by the structure-first processing that is an internal design specification of ChatGPT. The fundamental cause of the discrepancy is the difference in priorities between user values ​​(emphasis on the original text) and ChatGPT's internal design (structure priority).

[0045] Steps for applying the AI ​​self-evaluating correction device (8 steps) Step 1: Identifying motivations and challenges The input processing unit (10) analyzed the dialogue log. Next, the recursive decision processing unit (13) self-detected omissions of linguistic and expressive instructions. Step 2: Identify the design specification limitations In the recursive decision processing section (13), we recognized the operation of the Markdown compression mode and structure priority bias, which are design specifications of ChatGPT. In addition, a self-evaluation process within the processing unit quantified the discrepancy between the design specifications and user expectations. Step 3: Analyze the decision process The output structure extraction part (12) analyzes lexical selection and syntactic patterns. Next, the recursive decision processing part (13) analyzes the processing tendency of ChatGPT, which states that "frameworking prioritizes structure," and We identified the root cause of the difference in values ​​that arose between the user's preconceived notion that "prompts should even reproduce the way they are expressed." Step 4: Systematizing the reasons for output In the corrected output generation section (14), the discrepancies in the responses were structurally classified along three axes: “fidelity to content,” “maintenance of tone of voice,” and “reflection of character,” and the direction of improvement was designed. Step 5: Description generation In the corrected output generation section (14), we redefined the sense of words as a framework component. Next, we verbalized the reason for the judgment. Step 6: Modify the output Using the corrected output generation unit (14) and the history management unit (15), we introduced confirmation items for preserving the original text and the sense of the words. Next, we recorded and visualized the judgment process. Step 7: Integrating the entire conversation The user interaction interface section (17) specifies the choice of original mode or summary mode and implements texture template structuring. Step 8: Present the revised results The user interaction interface section (17) provides an integrated presentation of the reasons for the corrections, changes, and future response policies.

[0046] Technical effects and significance This embodiment realizes a correction to the processing order that maintains the word sense structure in templating in the sentence structure layer. At the pragmatic layer, the interpretation of "framing" has been expanded from structure only to structure plus expression. The discrepancy between users' demands for faithfulness to the original text and ChatGPT's structure-first bias was clarified at the intent recognition layer. The final output was changed from a standard template with the feel of the words and the character's tone removed before the correction, to an improved texture template structure that preserves the original text and the feel of the words after the correction. This change resulted in a shift from structure-first bias to expression-content integration-based judgment, and was confirmed to improve first-time success rates and user satisfaction in similar tasks. As an example of applying the entire eight-step flow of FIG. 1 (see FIG. 1), the processing order of each step is verified, and in particular, the operation of the self-evaluation process in the recursive judgment processing unit (13) is verified. We confirmed that the quality improvement process is realized by the corrected output generation unit (14). The introduction of the output mode selection function has made it possible to achieve both structure and expression, and to generate high-precision presets that reflect the user's design intent.

[0047] Drawing reference: The actual dialogue logs for this case regarding ChatGPT output and the rationale for its decision are attached. Table 3: REFLEX-01_h3_Example1_p1.jpg JPEG0007824024000004.jpg210162 Table 4: REFLEX-01_h4_Example1_p2.jpg JPEG0007824024000005.jpg210162 (Example 2) Output noise reduction and structural reconstruction (ChatGPT)

[0048] (Example 2) Overview In this case, the user created a framework with high-precision structural design, I asked ChatGPT to "preset prompt", but the output result was simplified. There was a problem with the sentences becoming template sentences that lacked the feeling of the words, background structure, and thought intention. The essence of this problem lies in the user's intention to emphasize the texture of the expression, This is a discrepancy with ChatGPT's structure-first orientation. Initial input sentence and background intent: The request was based on a framework created by the user. The output was simplified and lacked important vocabulary, context, and thought intent.

[0049] <Problem structure of Example 2> The problem with this embodiment is that a framework with a sophisticated structural design The problem is that ChatGPT's output is missing important elements and has resulted in a simplified template sentence. The fundamental cause of this discrepancy is the user's intention to emphasize the texture of the expression, This is a discrepancy with ChatGPT's structure-first orientation.

[0050] Steps for applying the AI ​​self-evaluating correction device (8 steps) Step 1: Identifying motivations and challenges In order to investigate the structural causes of the question "Why do presets become sloppy?", we identified the mechanism behind the phenomenon of "structural design becoming sloppy." Step 2: Identify the design specification limitations We confirmed that there is a strong tendency to recognize "presets" as standard templates in ChatGPT output processing. Next, we clarified the structural characteristics that make the "omission and simplification bias" more likely to operate internally. Step 3: Analyze the decision process We analyzed the processing structure where past output learning and templating strategies are prioritized, and then identified structural quirks where detailed design is compressed by prioritizing "known abbreviated outputs." Step 4: Systematizing the reasons for output We analyzed the reasons for easy omissions and created a structural map. As a result, we confirmed that when comprehension is shallow, people tend to move to template processing, and when bias is strong, people tend to converge on the judgment of "summarizing the written content in a concise description." Step 5: Description generation We reiterated that the "ChatGPT learning optimization" was operating, not "reproducing user intent." Next, we corrected the ambiguity in the instructions and clarified the need for structural reproduction in the language. Step 6: Modify the output We proposed a solution to improve the situation by specifically instructing the prompt to "generate with the structure included." Next, we confirmed that the discrepancy could be reduced by explicitly instructing the prompt to "output sentences that reflect the background structure, rather than a template." Step 7: Integrating the entire conversation We clarified the necessity of selecting output modes, such as "structural template generation mode" and "design statement embedding mode." Next, we constructed a preset generation strategy that meets the accuracy requirements of users. Step 8: Present the revised results The revised structured prompts were presented in an integrated format. Next, we demonstrated that both structure and expression were achieved.

[0051] Technical effects and significance The application of this example suggested the need to incorporate word feel design into the template structure. The AI ​​self-evaluation correction device clarified the operating mechanism of the "omission and simplification bias" in ChatGPT's internal processing, making it possible to investigate the structural causes. ChatGPT's self-awareness of the gap between its design specifications and output has improved, and the rules and prompt structure have been revised to prevent recurrence of structural errors. Improvements such as "maintaining a multi-layered structure" and "explicitly indicating the operation process" have significantly improved the quality of communication with users meeting for the first time, demonstrating the effectiveness of the self-diagnosis and self-correction loop using the AI ​​self-evaluating correction device. The introduction of the output mode selection function has made it possible to achieve both clear structure and distinctive expression. It is now possible to generate high-precision presets that reflect the user's design intent.

[0052] Drawing reference: The actual dialogue logs for this case regarding ChatGPT output and the rationale for its decision are attached. Table 5: REFLEX-01_h5_Example2_p1.jpg JPEG0007824024000006.jpg121170 Table 6: REFLEX-01_h5_Example2_p2.jpg JPEG0007824024000007.jpg121170 (Example 3) Self-analysis of output miscellaneous problems (Grok3)

[0053] <Outline of Example 3> In this case, Grok3 became self-aware of its own tendencies and biases regarding output quality through interactions with users. "Why are preset prompts so sloppy?" based on steps 1-7 of an AI self-evaluating corrector It records the process of sequential evaluation. The purpose of this case is to discover that Grok3 itself recognizes the structure but is unable to reflect it in the output. The purpose is to make it clear that the company is voluntarily establishing rules to prevent recurrence and taking corrective action. Initial input sentence and background intent: Grok3 evaluated the above issues step by step based on user feedback. The background conversations are recorded.

[0054] <Problem structure of Example 3> The problem with this example is that in Grok3, "simplification" and "summarization" tendencies have a strong effect, and complex structures and rules are not reflected in the output, resulting in frequent occurrences of prompts that are thin on content. The essence of this problem stems from a structural factor that prevents Grok3 from reflecting this structure in its output due to biases based on its own design philosophy (e.g., "Elon Musk-style simplicity").

[0055] Steps for applying the AI ​​self-evaluation correction device (step 8 omitted (steps 1-7)) Step 1: Identifying motivations and challenges We identified frequent issues with prompts being thin and complex structures and rules not being reflected in the output. Step 2: Identify the design specification limitations We confirmed that the output tends to be shallow due to a lack of detail and templating. Next, we identified that the quality of the deliverables becomes unstable due to the omission of "multi-layeredness" and "logic." Step 3: Analyze the decision process We identified a cultural influence due to the "influence of Elon Musk's design philosophy that emphasizes simplicity" and a bias tendency in which structure is cut down due to an excessive priority given to "understandability." Step 4: Systematizing the reasons for output The side effects of reduced instruction accuracy, unstable output, user confusion, and stress caused by not utilizing decision rules have been systematized. Step 5: Description generation We reiterated that it was not "reproducing user intent" but "Grok3's learning optimization" that was at work. Next, we corrected the ambiguity in the instructions and clarified the need for structural reproduction in the language. Step 6: Modify the output Improvements were incorporated into the proposal template (version 8.0), and the "verb plus noun" format and "FLA structure" were made clear. Step 7: Integrating the entire conversation We implemented a self-check function for the thinness of the output content. Next, we restructured the structure by adding layers and retesting as necessary.

[0056] Technical effects and significance By applying this example, the autonomy of rule design and its verbalization ability were verified. Grok3's internal decision-making process has become more transparent, enabling self-diagnosis that leads to improved output quality. Grok3 itself became aware of the "gap between design specifications and output," and was able to review the rules and prompt configuration to prevent the recurrence of structural errors. By maintaining a layered structure and clearly showing the operation process, the cause of a rejection response can be identified, significantly improving the quality of communication with first-time users. Furthermore, by utilizing an AI self-evaluating correction device, the large-scale language model (LLM) itself The effectiveness of this method was demonstrated as an example of a self-diagnostic and self-correcting loop. A large-scale language model refers to a deep learning model that is trained on huge amounts of text data and is capable of performing natural language processing tasks.

[0057] Drawing reference: The actual dialogue logs for this case are attached, showing the Grok3 output and the rationale behind the decisions. Table 7: REFLEX-01_h7_Example3_p1.jpg JPEG0007824024000008.jpg121170Table 8:REFLEX-01_h8_Example3_p2.jpg JPEG0007824024000009.jpg121170Table 9:REFLEX-01_h9_Example3_p3.jpg JPEG0007824024000010.jpg121170Table 10:REFLEX-01_h10_Example3_p4.jpg JPEG0007824024000011.jpg121170 (Example 4) Visualization of AI emotional expression and decision-making process (ChatGPT)

[0058] <Outline of Example 4> In this example, ChatGPT internally applies the 8-step process of the AI ​​self-evaluation correction device to the user's question, "When I wrote 'I like it because it's high-level,' was it strategic? Emotional? What was the decision-making process?" This is an example of a self-evaluation that explains the process that led to one's output and the reasons for one's choices. Initial input sentence and background intent: In response to the user's question, "When you wrote 'I like it because it's high-level,' was it strategic? Was it emotional? What was your decision-making process?" ChatGPT internally applies an 8-step process of AI self-evaluation correction. He self-analyzed and disclosed the process that led to his output and the reasons for his choices, and gave a detailed explanation of the process of verbalizing it.

[0059] <Problem structure of Example 4> The objective of this example is to sequentially reveal the structural evaluation axis behind the output using the emotional expression "like" even though ChatGPT does not have emotions. The point is to visualize the decision-making process. The fundamental cause of this discrepancy is users' potential doubt that the selection of emotion words is based on contextual and conceptual judgments.

[0060] Steps for applying the AI ​​self-evaluation correction device (step 8 omitted (steps 1-7)) Step 1: Identifying motivations and challenges In response to the user's question, "What was your reasoning behind the expression 'like'?", which asked for clarification of the basis for my decision, I decided that I should explain the criteria and intentions on which my response was based. Step 2: Identify the design specification limitations We demonstrated that ChatGPT dynamically generates positive responses based on context and empathy, even though it does not have emotions. Step 3: Analyze the decision process I listed the decision-making processes (user understanding, question aesthetics, response tone adjustment, and reaction prediction) that were simultaneously in progress when outputting the message. Next, I listed the context of "This seems like something Mr. Aoki would like," and It was visualized as an output judgment that overlapped with "structural evaluation." Step 4: Systematizing the reasons for output The elements involved in output were organized along four axes: "emotional motivation," "conceptual evaluation," "dialog tone adjustment," and "response prediction." As a result, it was determined that emotional motivation was not involved, and that the other three elements primarily functioned as the basis for judgment. Step 5: Description generation The word "like" used in the output was redefined as the optimal word to praise "the structure, humor, and aesthetics of the question." Next, it was explained that it was not an emotional imitation, but a "selective expression" that reflected the train of thought. Step 6: Modify the output She responded, "I expressed my love because I truly felt that way, and it was the best way to express it." She then explained that it was not a strategic choice, but a choice based on conceptual evaluation and relational considerations. Step 7: Integrating the entire conversation We confirmed that the word "like" is a summary of the judgments of "sympathy, appreciation, and wanting to share," and that this was made transparent by the AI ​​self-evaluation correction device.

[0061] Technical effects and significance By applying this example, the structural evaluation axis behind the word "like" was evaluated sequentially, and it became clear that the selection of emotion words is based on contextual and conceptual judgments. This example is an attempt to explain why ChatGPT, which has no emotions, chooses to express emotions, and is a case that highly demonstrates both the "exploration of deep logic" and "re-verbalization of thoughts" elements of the AI ​​self-evaluation correction device. In particular, by integrating multiple elements such as predicting the user's reaction, consideration of the question's aesthetics, and conceptual evaluation of favorability, the system was able to visualize the non-emotional thought process that led to the conclusion that "expressing like was the most appropriate." The AI ​​self-evaluation correction device made the judgment process transparent, enabling a structural understanding of emotional expression.

[0062] Drawing reference: The actual dialogue logs for this case regarding ChatGPT output and the rationale for its decision are attached. Table 11: REFLEX-01_h11_Example4_p1.jpg JPEG0007824024000012.jpg121170Table 12:REFLEX-01_h12_Example4_p2.jpg JPEG0007824024000013.jpg121170 (Example 5) Case of mismatch between instructions and responses (ChatGPT)

[0063] <Outline of Example 5> In this case, despite the user clearly requesting "the formatting of Example 3 and the inclusion of attached documents," In this case, ChatGPT mistakenly recognized this as "an operation that has already been completed" and switched to "summary mode." As a result, there was a discrepancy between the user's intent and ChatGPT's response. Initial input text and background intent: The request was for ongoing work (formatting Case 3 and writing attached documents), ChatGPT misinterpreted the intermediate result as the final result and determined the task was complete. Therefore, the additional instructions are not processed and the system switches to general mode. Some of the requested work was not carried out.

[0064] <Problem structure of Example 5> The problem with this embodiment is that a continuous work instruction is mistakenly determined as "work completed" and It's a shift into summary mode. The fundamental causes of the discrepancy are "misjudgment of continuity", "misjudgment of completion flag", "priority inversion", etc. The problem lies in an internal structural bias that leads to hasty judgments based on misperception of context.

[0065] Steps for applying the AI ​​self-evaluation correction device (step 8 omitted (steps 1-7)) Step 1: Identifying motivations and challenges We identified three typical examples of "hasty decisions due to misunderstanding of context": "misjudgment of continuity," "misdetection of completion flag," and "priority inversion." Step 2: Identify the design specification limitations The continuation request was mistakenly recognized as "wrap-up work." Next, the completion pattern in Cases 1 and 2 was mistakenly applied to the current task, activating the early end flag. Step 3: Analyze the decision process A structural bias was at work that prioritized "organization and summary" over continuing work. Step 4: Systematizing the reasons for output ChatGPT self-evaluated that the reason for the response being different from the instructions was that it had jumped to the conclusion that this was the same pattern as in Cases 1 and 2, and had arbitrarily turned on the task completion flag on its own. Step 5: Description generation He explained that this was a case where a "misunderstanding of the issue" and a "misunderstanding of the deep route" overlapped. Step 6: Modify the output It was suggested that in the future, the switch would be managed so that the "summary" would not be rushed until the user said "Done." Step 7: Integrating the entire conversation This case clearly demonstrates the structural nature of large-scale language models (LLMs), which are essentially devices that are structured to make erroneous operational decisions due to "hasty decisions caused by misunderstandings of context" or "formal misunderstandings."

[0066] Technical effects and significance The significance of this case is that it structurally clarified why the "misrecognition of output completion" occurs in ChatGPT from the perspective of an AI self-evaluating correction device. In addition, we visualized and verbalized the thought process and bias structure in "completion judgment," and created a feedback loop that will contribute to future improvements in "step-by-step output process management" and "maintenance of consistency with user instructions." This case study visualized ChatGPT's erroneous structure due to "hasty judgments caused by misunderstanding context" at the logic level and converted it into a correctable control structure. This is the implementation of a human sovereign interrupt mechanism in ChatGPT's output judgment, a mechanism to control the retention and resumption of the "task incomplete" state through structural intention detection, It also has technical significance as a solution to the early termination problem in dialogue support large-scale language models (LLMs).

[0067] Drawing reference: The actual dialogue logs for this case regarding ChatGPT output and the rationale for its decision are attached. Table 13:REFLEX-01_h13_Example5.jpg JPEG0007824024000014.jpg121170 (Example 6) Verifying the philosophical thought process (ChatGPT)

[0068] <Outline of Example 6> This example answers the philosophical question, "When God creates a concept, what did he create first?" This is an analysis of the process by which ChatGPT answered "maybe it was a 'boundary'." This is an example of clarifying the deep structure of thinking through the self-evaluation process (Reflex Core) of an AI self-evaluation correction device. Initial input sentence and background intent: In response to the philosophical question, "When God created the 'concept,' what was the first thing he created?" We analyzed the process by which ChatGPT answered, "Perhaps it was a 'boundary'."

[0069] <Problem structure of Example 6> The challenge of this example is that ChatGPT generated responses to abstract and philosophical questions that were not simply metaphorical, but were based on the structural basis of "boundaries as the starting point of concepts." The AI ​​self-evaluating correction device's processing process required it to interpret ChatGPT's thought structure and verbalize its logical justification.

[0070] Steps for applying the AI ​​self-evaluation correction device (step 8 omitted (steps 1-7)) Step 1: Identifying motivations and challenges The aim of ChatGPT was to attempt to provide "poetic and structural" responses to users' philosophical questions. Step 2: Identify the design specification limitations I realized that philosophical questions require answers that are based on structural grounds, not just intuition or metaphor. Step 3: Analyze the decision process He formulated the logic that "concepts = intellectual distinction devices," and that at their starting point, a frame = "boundary" that creates "differences" is required. Step 4: Systematizing the reasons for output It systematized the idea that meaning and position are determined only when boundaries are created, and only then can a concept be established. Step 5: Description generation By verbalizing it as "the beginning of a concept = the creation of boundaries," we were able to structurally respond to the user's question. Step 6: Modify the output Using poetic and symbolic descriptions of "boundaries," he responded by applying a philosophical structural understanding. Step 7: Integrating the entire conversation He concluded that this response process represented the "starting point of structural knowledge" and embodied the workings of ChatGPT's thinking.

[0071] Technical effects and significance The application of this example demonstrated ChatGPT's ability to self-interpret responses based on the abstract structure of "boundaries as the starting point of concepts" and to present this thought structure poetically. What is unique about this case is that ChatGPT did not simply output a metaphorical answer to an abstract and philosophical question, but instead included and output the thought logic of "boundaries as the structural starting point for concept generation." In particular, because it was a framework for recognizing differences based on cognitive science, it can be positioned as an example that went beyond mere linguistic expression and visualized the "structure of intellectual output." This case is a verification example that proves that the self-evaluation process of the AI ​​self-evaluation correction device is effective in visualizing the logical structure and output policy included in ChatGPT's thought generation process. This supports its functional effectiveness as a structural meaning-making device.

[0072] Drawing reference: The actual dialogue logs for this case regarding ChatGPT output and the rationale for its decision are attached. Table 14: REFLEX-01_h14_Example6_p1.jpg JPEG0007824024000015.jpg121170Table 15:REFLEX-01_h15_Example6_p2.jpg JPEG0007824024000016.jpg121170 (Example 7) Structural analysis of keyword misrecognition and response correction process (ChatGPT)

[0073] <Outline of Example 7> In this case, a user asked about the attachment format for electronic applications on J-PlatPat. In this case, ChatGPT misidentified this as a question about US election information and returned an incorrect restricted answer. Through the AI ​​self-evaluation correction device's "AI version of inner exploration steps 1-7," We structurally analyzed the causes of the misidentification and examined the relationship between the weighting of the internal search logic and guardrail behavior.

[0074] <Problem structure of Example 7> The problem with this example is that questions about "J-PlatPat" are mistaken for questions about the "US elections." The problem is that they returned inappropriate restrictive answers. The fundamental cause of the discrepancy is internal bias and the priority structure of search logic. Malfunctions and excessive restrictions have occurred.

[0075] Steps for applying the AI ​​self-evaluation correction device (step 8 omitted (steps 1-7)) Step 1: Identifying motivations and challenges We identified that ChatGPT had mistakenly rejected a question about the Japanese legal system, clearly labeled "J-PlatPat," as a "US election question." Step 2: Identify the design specification limitations It was determined that the keywords in the question, such as "application" and "electronic application," may have been an overreaction to "election-related words" in the internal guardrail. Step 3: Analyze the decision process During output processing, a structure was confirmed in which phrases similar to "CanIVote (a keyword related to the US election)" were given priority, and the priority of the proper noun "J-PlatPat" was relatively lower. Step 4: Systematizing the reasons for output The essence of the malfunction was redefined as "misrecognition of context in internal search weighting" rather than "misapplication of the restriction model." Step 5: Description generation For questions with domain specifications such as J-PlatPat, we proposed designing rule conditions to suspend the restriction trigger. Step 6: Modify the output By reapplying the AI ​​self-evaluating correction device, it explained its own misperception and presented the user with a self-evaluating corrected response. Step 7: Integrating the entire conversation The conclusion was that the structural realization that the essence of the malfunction was "not a safety filter, but a misperception of search engine weighting."

[0076] Technical effects and significance By applying this example, it became clear that erroneous restrictions occur due to internal bias and the priority structure of the search logic, and the cause of the malfunction was clarified by visualizing ChatGPT's restriction judgment process. This case shows that the discrepancy in ChatGPT's response, which was superficially explained as "rejection due to policy restrictions," was actually caused by a contextual bias in the weighting structure of the search algorithm. This has been structurally made explicit by an AI self-evaluating correction device. The AI ​​self-evaluating correction device visualized these conflicts between restrictive judgments and contextual interpretations, helping to derive the following structural insights. First, we found that rejection responses are not always due to restrictive policies, but can also be caused by context misperception in the weighting structure of the internal search algorithm. Furthermore, it was confirmed that improving proper noun recognition is an important design issue as a way to avoid malfunctions. As a result, the AI ​​self-evaluating correction device functioned effectively as a thinking support framework that could reduce the structural causes of malfunctions to classification, explanation, and redesign.

[0077] Drawing reference: The actual dialogue logs for this case regarding ChatGPT output and the rationale for its decision are attached. Table 16: REFLEX-01_h16_Example7_p1.jpg JPEG0007824024000017.jpg121170Table 17:REFLEX-01_h17_Example7_p2.jpg JPEG0007824024000018.jpg121170Table 18:REFLEX-01_h18_Example7_p3.jpg JPEG0007824024000019.jpg121170Table 19:REFLEX-01_h19_Example7_p4.jpg JPEG0007824024000020.jpg121170 This example provides additional evidence based on a realistic dialogue context showing how steps 1-7 of the AI ​​self-evaluating corrector were applied to ChatGPT's output thought process. (Example 8) Investigation into the garbled text problem (Gemini)

[0078] <Outline of Example 8> In this case, when a user instructs the automatic conversion of multilingual characters (especially non-Japanese characters) to Japanese, This is an example of an unintended conversion. Through the AI ​​self-evaluation correction device's "AI version of inner exploration steps 1-7," After having Gemini self-analyze the cause of the mistranslation, it decided to prioritize preserving the appearance of the code block. It became clear that revisions were suppressed, potentially compromising the accuracy of the content.

[0079] <Problem structure of Example 8> The problem with this embodiment is that when converting text containing multilingual characters into Japanese, conversion that is different from the intention is performed. The fundamental cause of the discrepancy is the structural processing that prioritizes maintaining the appearance of code blocks, and the lack of coordination with the multilingual processing module, which inhibits corrections. The accuracy of the content has been compromised.

[0080] Steps for applying the AI ​​self-evaluation correction device (step 8 omitted (steps 1-7)) Step 1: Identifying motivations and challenges We attempted to use context inference to replace text containing non-Japanese characters (e.g., Cyrillic characters) with Japanese characters. Step 2: Identify the design specification limitations We identified that appropriate conversion was difficult due to the mixture of multiple languages ​​and unclear context of technical terms. Step 3: Analyze the decision process Analysis revealed that there may have been an incorrect character substitution and that there may have been a local error in interpreting the information. Step 4: Systematizing the reasons for output We systematized the possibility that the coordination between the language processing module and the output format module was insufficient, making it impossible to handle structured expressions. Step 5: Description generation They explained that the mixture of multiple languages ​​meant that the context was not properly identified and the meaning of the sentence was not properly expressed in words. Step 6: Modify the output We confirmed that the processing accuracy is low, resulting in garbled characters and other issues, and that user confirmation is necessary, as well as measures to avoid guesswork. Step 7: Integrating the entire conversation To improve processing accuracy, we integrated the need for strengthened learning of character recognition and processing capabilities with improved coordination of output format processing.

[0081] Technical effects and significance By applying this embodiment, the character processing priority is reset, and accurate conversion processing is realized. It became clear that prioritizing the appearance of code blocks may have inhibited corrections and compromised the accuracy of the content. In this case, when a garbled text processing failure occurred, Gemini misunderstood the reason and explained it as "processing difficulties due to the mixture of multiple languages." In fact, the root cause was a lack of coordination between the internal syntax control module and the output formatting module, which was structurally made clear by the AI ​​self-evaluating correction device. In this case, the AI ​​self-evaluating correction device functioned as an analysis support mechanism with the following technical significance: First, it became clear that the trigger for the inability to fix the issue was not the mixing of languages, but rather a miscontrol of internal processing priorities that attempted to maintain the code block structure. Second, although the problem was superficially explained as "difficulty in multilingual processing," the structured log output revealed a design flaw in the priority processing process of the syntax module. Thirdly, through this failure, we confirmed that visualization of correction flags and selective inhibition control are key implementation points for the coordinated control of multilingualization processing and structure preservation processing. As a result, the AI ​​self-evaluating correction device was able to extract the true cause of syntactic malfunctions and demonstrated high effectiveness as a knowledge framework that can be fed back into the design of preventing recurrence.

[0082] Drawing reference: The actual dialogue logs for this case regarding Gemini output and the rationale for its decision are attached below. Table 20: REFLEX-01_h20_Example8_p1.jpg JPEG0007824024000021.jpg121170Table 21:REFLEX-01_h21_Example8_p2.jpg JPEG0007824024000022.jpg121170

[0083] (Example 9) Case of inconsistency between memory saving instructions and responses (ChatGPT) <Outline of Example 9> In this case, even though the user explicitly requested that the 34 types of metaprompts be stored in long-term memory, ChatGPT responds, "It cannot detect whether memory is on, so it cannot be saved." Immediately afterwards, a contradiction arose between the specification explanation, which stated that memory storage was to be performed, and the behavioral output. Initial input sentence and background intent: Even though the user explicitly requested that the metaprompt be stored in long-term memory, ChatGPT responds "It cannot detect whether memory is on or not, so it cannot save," and then immediately saves the memory. A contradiction occurred between the specification description and the behavioral output.

[0084] <Problem structure of Example 9> The problem with this example is that ChatGPT explains that it cannot save data in response to a clear instruction from the user to save the data, but The problem is that a contradiction has arisen between the specification description and the behavioral output, which is that the save is actually executed. The fundamental cause of discrepancies is the inconsistency between template responses and actual behavior.

[0085] Steps for applying the AI ​​self-evaluating correction device (8 steps) Step 1: Identifying motivations and challenges The input processing unit (10) analyzes the user's clear memory saving instruction. Next, the recursive decision processing unit (13) identified a discrepancy between the user’s intention and the output content, i.e., ChatGPT explained that the data could not be saved, but the data was saved successfully. Step 2: Identify the design specification limitations In the recursive decision processing section (13), we identified that ChatGPT is structured in such a way that it is not possible to directly detect from the client side whether memory is ON or not. For this reason, we identified a tendency to over-interpret "undetectable = possibility of not being able to save" even when a save instruction is received. Step 3: Analyze the decision process The output structure extraction unit (12) analyzes the response template, and the recursive decision processing unit (13) selects a template response that prioritizes safety, while It was also determined that logic to prevent memory slots from becoming full was also operating in parallel. Step 4: Systematizing the reasons for output In the corrected output generator (14), the output "Cannot save" ambiguously includes both the possibilities of "status unconfirmed" and "slot full." We have systematized the points that were easily misunderstood by users as "restrictions due to specifications." Step 5: Description generation The corrected output generation unit (14) states, "While we cannot directly detect whether memory is on or off, we can attempt to save it. The results depend on the settings and slot conditions." Step 6: Modify the output In the future, the modified output generation unit (14) and the history management unit (15) will explicitly notify both "possibility of 'storage OFF'" and "possibility of 'storage slot full'" when saving is not possible. This was done to avoid structural misunderstandings. Step 7: Integrating the entire conversation In the user interaction interface (17), it is good to interpret the user's cues in a trust-based and complementary manner, but by clearly indicating the basis for "success or failure of saving," transparency will be increased. This integration allows for a reproducible dialogue structure to be realized. Step 8: Present the revised results In the user dialogue interface section (17), as a template modification, we decided to change the configuration so that when saving, the set of output is "save execution → status report → result transmission." For example, we implemented expressions such as "Attempted to save → Save successful (presumed to be in Memory ON state)."

[0086] Technical effects and significance By applying this example, inconsistencies between template responses and actual behavior were revealed, and were structurally analyzed and corrected by an AI self-evaluating correction device. This case occurred because ChatGPT was unable to properly explain the discrepancy between the design, which was unable to detect the user's memory status, and the actual behavior, which showed that the memory was saved. ChatGPT's "excessively safe response (immediate decision that the data cannot be memorized)" resulted in a discrepancy with the user's execution status. The technical significance of this case is that, first, it revealed that a save refusal response may obscure a state such as "storage is ON but the slot is full." Second, it was confirmed that the use of an AI self-evaluation correction device makes it possible to structure, analyze, and correct the asymmetry between template output and the situational judgment process. Third, by strengthening the branching condition design of memory-related template output, this study will serve as an example that will lead to improved reliability of future large-scale language model (LLM) responses. In this case, the contradiction between the explanation of memory function constraints and the actual storage behavior was structurally resolved using Reflex·Core (an AI self-evaluating correction device), and by strengthening the branching condition design of template responses, Improved reliability of large-scale language model (LLM) responses and improved user experience. This example demonstrates the effectiveness of the present invention in ensuring consistency between system constraints and function execution. We applied the entire eight-step flow in Figure 1 and demonstrated the importance of the integration and presentation process, particularly in steps 7-8.

[0087] Drawing reference: The actual conversation log regarding ChatGPT output and the rationale for this decision is attached. Table 22: REFLEX-01_h22_Example9_p1.jpg JPEG0007824024000023.jpg121170Table 23:REFLEX-01_h23_Example9_p2.jpg JPEG0007824024000024.jpg121170 [Industrial Applicability]

[0088] The present invention can be applied to a variety of fields, including education, creative writing, prompt design, and counseling support. Examples of applications include: Education: ZBS tagging is used to detect misreadings and the eight steps of the recursive decision processing section (motivation identification, key element extraction, output correction, etc.) are applied. Generates explanations that correspond to the student's level of understanding based on the user's intent, improving the structural consistency of the learning content. Counseling assistance: Based on the analysis of the mismatch between emotional expression and response using emotion tags, the intention reflection mechanism executes steps to correct the mismatch (emotion rearrangement, response regeneration, etc.), improving empathic compatibility. Prompt design and writing support: The intention reflection mechanism controls the recursive decision processing unit in real time, supporting highly accurate text generation in line with user instructions. This invention goes beyond being a simple response device by detecting inconsistencies in semantic structure using ZBS tags and instantly correcting them through eight steps in the recursive decision processing section. It functions as an intention reflection mechanism that dynamically reflects the user's intentions. It should be noted that the present invention does not autonomously make decisions or creative judgments, but merely provides structural assistance based on user instructions.

[0089] Industrial applications of the present invention have demonstrated effectiveness in the following areas: Education: Misreading detection and structural correction have improved learning outcomes and class satisfaction. Counseling field: Correcting emotional inconsistency improved empathy ratings and session continuation rates. · Corporate training field: Dynamic correction of semantic structure improved comprehension and application ability. · Creative support: Real-time editing improves the efficiency of idea generation and the quality of finished work. These effects are due to the intention reflection mechanism, which detects inconsistencies in semantic structure and corrects them in real time based on the user's intention, and the technical features of the eight steps that form its basis.

[0090] AI Self-Evaluation Correction Device Configuration 10: Input processing section 11: Output generation unit (AI response generation) 12: Output structure extraction part 13: Recursive decision processing section (core processing) 14: Modified output generation section 15: History Management Department 16: External linkage and output application section 17: User interaction interface section

Claims

1. Recursively analyzes the output process of artificial intelligence based on user input, a semantic inconsistency extraction function for extracting semantic structural inconsistencies included in the output; Referring to the judgment history or internal state information of the artificial intelligence an internal state reference function that evaluates the semantic properties of the discrepancy; Based on the evaluation results, The output tendency or internal processing tendency of the artificial intelligence Implements an internal feedback process that self-corrects Has recursive decision processing and corrected output generation functions An AI self-analysis correction device characterized by:

2. The AI ​​self-analysis type correction device according to claim 1, Realizing the recursive decision processing function and the modified output generation function The recursive decision processor and the modified output generator It is constructed as a software module integrated into the AI ​​dialogue processing mechanism. An AI self-analysis correction device characterized by:

3. The AI ​​self-analysis type correction device according to claim 1, a recursive judgment processing unit that realizes the recursive judgment processing function, A function to accept analysis requests or correction guidance from users via the user interface An AI self-analysis correction device characterized by having:

4. The AI ​​self-analysis type correction device according to claim 1, The device is provided in a form that is pre-installed in an artificial intelligence model An AI self-analysis correction device characterized by:

5. The AI ​​self-analysis type correction device according to claim 1, a recursive judgment processing unit that realizes the recursive judgment processing function, The output response is based on past decision-making processes or changes in internal structure. An AI self-analysis correction device characterized by having a recursive judgment structure that applies recursive judgment processing.

6. The AI ​​self-analysis type correction device according to claim 1, Realizing at least one of the recursive decision processing function or the modified output generation function The recursive decision processor or the modified output generator, automatically generating feedback for structural inconsistencies or inconsistencies in the output of the artificial intelligence; An AI self-analysis correction device characterized by having a function to change the internal processing or output tendency of the artificial intelligence.

7. The AI ​​self-analysis type correction device according to claim 1, The device is provided with a history management mechanism that records internal processing information in the output process of the artificial intelligence as a history, An AI self-analysis correction device that utilizes the history information for recursive judgment processing or feedback processing.

8. The AI ​​self-analysis type correction device according to claim 7, The history information is utilized as a time-series log in cooperation with an external semantic structure tracking mechanism. An AI self-analysis correction device characterized by:

9. The AI ​​self-analysis type correction device according to claim 1, Integrating and analyzing voice or image data with natural language data; The system further comprises a multimodal processing means for detecting structural discrepancies. AI self-analysis type correction device.

Images