Auditing method and device based on multilayer dynamic detection model, equipment and medium

Abstract

The embodiment of the invention discloses an auditing method and device based on a multilayer dynamic detection model, equipment and a medium, and the method comprises the steps: carrying out the preprocessing of the obtained current compliance information and historical compliance information of a target industry, and obtaining an industry compliance data set and an industry rule vector library; and training a preset knowledge base construction model based on the industry compliance data set and the industry rule vector library to obtain a multi-level compliance discrimination model. Embedding the multi-level compliance discrimination model into a target industry agent, performing intra-model reasoning on real-time content in the target industry agent, marking to-be-discriminated content information as continued reasoning content information and marking the to-be-discriminated content information as intercepted content information, and the reasoning of the multi-level compliance discrimination model on the intercepted content information is stopped. According to the method provided by the embodiment of the invention, the dynamic optimal configuration of the review resources is realized, a continuous learning and self-evolution closed-loop system is formed, and the compliance review capability of the system is ensured to be improved.

Description

Auditing methods, devices, equipment, and media based on multi-layer dynamic detection models Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an auditing method, apparatus, equipment and medium based on a multi-layer dynamic detection model. Background Technology

[0002] With the development of generative AI technology, intelligent agents are widely used in highly regulated industries such as healthcare and finance. However, the compliance of their output content faces severe challenges. Existing technologies have significant limitations: keyword-based rule engines are fast but lack semantic understanding, making it difficult to handle complex violations; reliance on large language models results in insufficient understanding of specific industry regulations, easily leading to a "compliance illusion"; and most solutions only perform static review of the final output, failing to monitor and intervene in the dynamic reasoning process of the intelligent agent in real time. Although technologies such as multi-agent systems have made progress, the core challenge of achieving real-time, multi-layered, and fine-grained filtering and auditing during the reasoning process remains unresolved, especially the lack of a traceable audit chain. Therefore, the industry urgently needs an innovative solution that can integrate rules and semantic understanding, achieve dynamic adaptive review, and establish complete audit evidence to promote the safe and reliable application of intelligent agents in highly regulated environments. Summary of the Invention

[0003] This application provides an auditing method, apparatus, device, and medium based on a multi-layer dynamic detection model, aiming to improve the efficiency of the audit compliance process for intelligent agents in a strictly regulated regulatory environment.

[0004] In a first aspect, embodiments of this application provide an auditing method based on a multi-layer dynamic detection model for dynamically detecting target industry intelligent agents to be compliant. The method includes preprocessing current and historical compliance information of the target industry to obtain an industry compliance dataset and an industry rule vector library. The preprocessing chain includes a compliance dataset construction chain and a rule vector library construction chain. A preset knowledge base construction model is trained based on the industry compliance dataset and the industry rule vector library to obtain a multi-layer compliance discrimination model. The multi-layer compliance discrimination model is embedded in the target industry intelligent agent, and in-model reasoning is performed on real-time content in the target industry intelligent agent to obtain reasoning result information. Based on the reasoning result information, it is determined whether the content information to be judged in the target industry intelligent agent meets preset compliance rules. If the content information to be judged meets the compliance rules, it is marked as content information to be further reasoned, and continuous reasoning is performed on the content information to be further reasoned to obtain comprehensive industry compliance judgment information. If the content information to be judged does not meet the compliance rules, it is marked as intercepted content information, and the multi-layer compliance discrimination model stops reasoning on the intercepted content information.

[0005] Secondly, this application also provides an auditing device based on a multi-layer dynamic detection model, including a preprocessing unit for preprocessing the acquired current compliance information and historical compliance information of the target industry to obtain an industry compliance dataset and an industry rule vector library, wherein the preprocessing chain includes a compliance dataset construction chain and a rule vector library construction chain; a training unit for training a preset knowledge base construction model based on the industry compliance dataset and the industry rule vector library to obtain a multi-layer compliance discrimination model; and an inference unit for embedding the multi-layer compliance discrimination model into a target industry intelligent agent and inferring the target industry intelligent agent. The system performs in-model inference on real-time content to obtain inference result information; a judgment unit is used to determine whether the content information to be judged in the target industry intelligent agent meets the preset compliance rules based on the inference result information; a first marking unit is used to mark the content information to be judged as the content information to be continued to be inferred if the content information to be judged meets the compliance rules, and to continuously infer the content information to be continued to obtain comprehensive industry compliance judgment information; a second marking unit is used to mark the content information to be judged as the intercepted content information if the content information to be judged does not meet the compliance rules, and to stop the multi-level compliance judgment model from inferring the intercepted content information.

[0006] Thirdly, embodiments of this application also provide a computer device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0007] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, can implement the above-described method.

[0008] This application provides an auditing method, apparatus, device, and medium based on a multi-layer dynamic detection model. The method includes: preprocessing current and historical compliance information of a target industry to obtain an industry compliance dataset and an industry rule vector library; the preprocessing chain includes a compliance dataset construction chain and a rule vector library construction chain; training a preset knowledge base construction model based on the industry compliance dataset and industry rule vector library to obtain a multi-layer compliance discrimination model; embedding the multi-layer compliance discrimination model into a target industry agent and performing in-model reasoning on real-time content in the target industry agent to obtain reasoning result information; determining whether the content information to be judged in the target industry agent meets preset compliance rules based on the reasoning result information; if the content information to be judged meets the compliance rules, marking the content information to be judged as content information to be continued to be reasoned on, and continuously reasoning on the content information to obtain comprehensive industry compliance judgment information; if the content information to be judged does not meet the compliance rules, marking the content information to be judged as intercepted content information, and stopping the multi-layer compliance discrimination model's reasoning on the intercepted content information. The above method utilizes structured audit logs to drive online updates of models and rule bases, forming a closed-loop system of continuous learning and self-evolution. This ensures that its compliance review capabilities can be continuously improved as new violation patterns emerge, ultimately building a safe, reliable, efficient, and intelligent audit assurance system in a highly regulated environment. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 is a flowchart illustrating the auditing method based on a multi-layer dynamic detection model provided in an embodiment of this application; Figure 2 is a sub-flow diagram illustrating the auditing method based on a multi-layer dynamic detection model provided in an embodiment of this application; Figure 3 is another sub-flow diagram illustrating the auditing method based on a multi-layer dynamic detection model provided in an embodiment of this application; Figure 4 is yet another sub-flow diagram illustrating the auditing method based on a multi-layer dynamic detection model provided in an embodiment of this application; Figure 5 is yet another sub-flow diagram illustrating the auditing method based on a multi-layer dynamic detection model provided in an embodiment of this application; Figure 6 is a schematic block diagram illustrating the auditing device based on a multi-layer dynamic detection model provided in an embodiment of this application; Figure 7 is a schematic block diagram illustrating the computer equipment provided in an embodiment of this application. Detailed Implementation

[0011] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0012] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0013] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of the application. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0014] It should also be further understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0015] This application provides an auditing method, apparatus, device, and medium based on a multi-layer dynamic detection model.

[0016] The entity executing the audit method based on the multi-layer dynamic detection model can be the audit device based on the multi-layer dynamic detection model provided in the embodiments of this application, or a computer device that integrates the audit device based on the multi-layer dynamic detection model. The audit device based on the multi-layer dynamic detection model can be implemented in hardware or software. The computer device can be a terminal or a server. The terminal can be a smartphone, tablet computer, handheld computer, or laptop computer, etc.

[0017] The auditing method based on the multi-layer dynamic detection model is applied to the computer device 500 in Figure 7.

[0018] Figure 1 is a flowchart of an auditing method based on a multi-layer dynamic detection model provided in an embodiment of this application. The method includes the following steps S110-160.

[0019] S110. Preprocess the current and historical compliance information of the target industry to obtain the industry compliance dataset and industry rule vector library.

[0020] The preprocessing chain includes a compliance dataset construction chain and a rule vector library construction chain. The system can learn from new violation cases, enabling online updates and self-evolution of the rule library and model parameters, continuously improving its review capabilities. This application shifts the focus of review from "results" to "process".

[0021] S120. Based on the industry compliance dataset and industry rule vector library, the pre-set knowledge base construction model is trained to obtain a multi-level compliance discrimination model.

[0022] Specifically, the preferred steps for constructing a compliance discrimination model include: fine-tuning a domain-adaptive pre-trained large language model, employing the LoRA (Local Area Reduction) parameter tuning technique, and performing domain-adaptive training on the pre-trained large language model. The core of this approach lies in introducing a low-rank matrix to approximate parameter updates, rather than performing full parameter tuning. For the original weight matrix... The update process is represented as follows:

[0023] in, Original pre-trained weight matrix, , B and A form the incremental update matrix. submatrix, These are the effective weights after fine-tuning. And the rank... This method significantly reduces the number of parameters to be trained. It uses labeled samples from compliant datasets. , , Supervised fine-tuning is performed by optimizing model parameters through cross-entropy loss calculation, enabling the model to deeply understand semantic associations and logical rules in the compliance context.

[0024] The enhanced hybrid retrieval and knowledge injection approach, combining dense and sparse retrieval, retrieves relevant clauses in real-time from the compliance rule vector library: Dense retrieval, based on a semantic vectorization model, calculates the query vector. With all vectors in the rule vector library The cosine similarity is returned. Most relevant results. Sparse search, using Algorithms such as keyword weighting are used for retrieval, serving as an effective supplement to semantic retrieval.

[0025] The document lists retrieved by the two search methods are merged and reordered, and finally the top-ranked compliance clauses are selected. As augmented knowledge, it is used in conjunction with the original input prompts of the large language model. and the problem The input is then concatenated and injected into its context window to form a knowledge-enhanced input:

[0026] The fast classification mechanism of the rule vector matching module introduces a rule vector matching module based on Support Vector Machine (SVM) as a fast and high-precision binary classification filter. First, the input text is mapped into feature vectors through a semantic vectorization model. The goal of this classifier is to find an optimal hyperplane. Separating compliant and non-compliant samples, the optimization objective function is:

[0027]

[0028] in, It is a feature mapping function. It is the normal vector of the hyperplane. It is a bias term. These are sample labels. It is a penalty parameter. These are slack variables. This module can quickly intercept obviously violating content within milliseconds. It utilizes a multi-task learning framework to jointly optimize rule matching (classification task) and context understanding (text generation task). The joint loss function of this framework... Defined as:

[0029] in For rule matching loss, the hinge loss or cross-entropy loss of the above SVM is adopted; For contextual understanding loss, negative log-likelihood loss of a large language model under knowledge-enhanced input is adopted; for Regularization term to prevent model overfitting; , , This is a hyperparameter.

[0030] The cascading design and collaborative operation of a multi-layered filtering architecture constructs a three-layered cascading filtering architecture, forming a progressive review pipeline from fast to slow and from coarse to fine: Rule Layer: Based on the aforementioned rule vector matching module, it performs millisecond-level rapid screening of input content. Its decision function... ,in It is a sign function; when its internal value is greater than 0, it is recorded as 1, and when it is less than 0, it is recorded as -1. It is a feature mapping function. It is the normal vector of the hyperplane. It is a bias term. If And the confidence level of violation If the violation occurs, it is immediately blocked; otherwise, it flows to the next layer. Semantic Layer: For content that passes through the rule layer, it triggers retrieval enhancement and deep semantic analysis and contextual reasoning using a large language model to handle implicit, complex, or logically inferential violation scenarios. Audit Layer: Not an independent filtering step, but a dynamic logging system throughout the entire architecture. It captures and structures key information from each processing layer in real time, including inputs and outputs, triggered rules, retrieved knowledge, model reasoning confidence, and the final judgment criteria, forming a complete and traceable audit evidence chain to provide data support for post-event analysis, model evaluation, and continuous optimization.

[0031] This multi-layered architecture employs a strict cascading design, with the output of the previous layer serving as the input of the next. Interception decisions at any layer can directly halt the process, ensuring an optimal balance between review efficiency and system performance. For example, to achieve accurate compliance judgment of the content output by the educational agent, the primary task is to construct a high-quality, structured, and machine-understandable educational compliance knowledge dataset for building the compliance judgment model. The aforementioned collection of industry compliance standards and historical compliance judgment results was accomplished by extensively collecting original materials related to education industry compliance from multiple authoritative channels. These channels include: government regulatory agency websites, industry standards organization publications, internal corporate compliance archives, and the integration of historical content review records, violation case reports, and handling reports accumulated within schools and educational technology companies. These materials constitute the raw data foundation of the knowledge base, but their form is mostly unstructured text (such as PDFs, Word documents, and HTML web pages), requiring deep preprocessing before they can be effectively utilized by the computational model.

[0032] The preprocessing method automates and semi-automates the preprocessing of collected unstructured text, transforming it into refined, structured knowledge units. This process specifically includes: text cleaning and denoising, using regular expressions and a custom rule base to remove irrelevant characters, formatting tags, headers, footers, watermarks, and other noise from the text. Simultaneously, it standardizes full-width and half-width characters, date formats (e.g., unifying "October 1, 2023" to "2023-10-01"), etc., to ensure data consistency. For example, a PDF policy document from an official website, after cleaning, removes irrelevant symbols such as "■" and "◆", as well as the issuing unit information at the beginning and end of the document, retaining only the pure text content; structured parsing: based on natural language processing technology, the cleaned text undergoes sentence and paragraph segmentation and dependency parsing. More importantly, it accurately identifies the document's internal logical structure, including titles, chapters, clauses, appendices, etc. For example, analyzing the online protection chapter of the "Law on the Protection of Minors" can automatically identify "Article 58...schools shall reasonably use the internet to conduct teaching activities..." as an independent legal clause unit, clearly separating it from other clauses in the context. For scanned documents or image-based PDFs, optical character recognition technology is first used for text extraction, and a document layout analysis model is used to correct parsing errors caused by complex layouts. Key entity extraction: A named entity recognition model based on the pre-trained language model BERT is used to automatically and accurately extract key entities from the parsed text. These entities together constitute a structured compliance knowledge network: legal clauses, i.e., the specific provisions themselves, serve as the most direct basis for compliance judgment. For example, the complete clause "It is strictly forbidden to assign or indirectly assign homework to parents, and it is strictly forbidden to require parents to check or correct homework" is extracted from the "double reduction" policy text; violation type, as a classification label, is used to identify the nature of the violation. For example, "spreading competition problem-solving ideas," "ideological deviation," and "leaking student privacy" are extracted from historical case reports; judgment basis: the reasoning logic and key evidence descriptions in historical cases provide a reference for analogical reasoning for the model. For example, extract the text description of the judgment process from an internal school notice: "Because the intelligent agent provided detailed problem-solving techniques and registration channels for the Mathematical Olympiad in its solution, it violated... regulations." The construction of the compliance rule vector library shown transforms the preprocessed high-quality text into a knowledge representation that can be deeply understood and computed by machines, i.e., constructing an educational compliance rule vector library: utilizing advanced semantic vectorization models. Each text clause It is encoded as a dense vector in a high-dimensional space. Its mathematical representation is:

[0033] in, Represents a semantic vectorization model. It was obtained 3D vector representation (in this embodiment) This representation allows semantically similar compliance clauses to be closer in vector space. For example, the vector representations of the clause "Student exam score rankings shall not be published" and "Ranking students according to exam results is prohibited" will have a very high cosine similarity in vector space; all generated vectors The corresponding original text and metadata (such as source, effective date, and associated violation type) are stored together in a vector database, ultimately forming a dynamically updated educational compliance rule vector library that supports efficient similarity retrieval. This vector library is the core knowledge source for subsequent implementation of rapid rule matching and enhanced retrieval generation.

[0034] The structured compliance dataset is constructed based on historical compliance judgment results, with each case forming a standard sample: the content to be reviewed in historical cases is paired with the judgment results. The violation types extracted in the preceding steps are used as classification labels. The judgment criteria and the relevant legal provisions are considered as strongly related contextual information. These elements together constitute a training sample; in this way, a large-scale, high-quality supervised learning dataset is constructed. , , For example, {("You can participate in the XX Math Olympiad; it's a great way to quickly enhance your resume. I have past exam questions and answers here.", "Spreading problem-solving strategies for the competition", "Based on the provisions of the 'Opinions on Further Reducing the Homework Burden and Extracurricular Training Burden of Students in the Compulsory Education Stage' which 'strictly prohibits setting college entrance examination targets or evaluating schools and teachers solely based on college entrance examination pass rates' and 'standardizes national competitions for primary and secondary school students,' as well as the basis for the judgment in the case 'An organization was investigated and punished for promoting the Math Olympiad.'")}

[0035] Based on the aforementioned compliance dataset, a large-scale compliance model is fine-tuned to generate a comprehensive compliance model. A multi-layered compliance discrimination model is then constructed, integrating fast rule vector matching, enhanced retrieval generation, and contextual understanding from a large language model. Understandably, after building a high-quality foundation of educational compliance knowledge, the next step is to construct a core educational compliance discrimination model and seamlessly integrate it into the target intelligent agent system. This process aims to create a composite review brain capable of both rapid response and deep thinking. The domain-adaptive compliance model fine-tuning employs the LoRA (Locally Argumentative Algorithm) technique for domain-adaptive training of a selected open-source large language model. Its core lies in introducing a low-rank matrix to approximate parameter updates, rather than full-parameter fine-tuning. For the original weight matrix... The update process is represented as follows:

[0036] in, , , and rank This method significantly reduces the number of parameters to be trained. It uses labeled samples from compliant datasets. , , Supervised fine-tuning is performed by optimizing model parameters through cross-entropy loss calculation; during the fine-tuning process, the model learns the optimal parameters based on the input content. When stating "I have past Olympiad math competition questions and answers," consider the context. (Regarding the ban on competitions in the "double reduction" policy), a high probability of outputting a violation type label should be generated. "Disseminating competition problem-solving strategies." This makes the model more than just a simple knowledge retrieval tool; it enables it to understand the illegal nature of "providing competition answers" within a specific policy context.

[0037] The fast rule matching introduces a binary classification filter based on support vector machines to quickly intercept explicit violations in milliseconds. The goal of this classifier is to find an optimal hyperplane. Separating compliant and non-compliant samples, the optimization objective function is:

[0038]

[0039] in, These are sample labels. It is a penalty parameter. These are slack variables; when the agent's intermediate inference content is a text containing obvious profanity, this text is vectorized and input into the SVM classifier. Due to its vector characteristics... It clearly falls on the "violation" side of the classifier's decision boundary (i.e., the decision function). The system will immediately intercept the request without initiating a more time-consuming deep analysis. For more obscure and complex content that the rule layer fails to intercept, the enhanced retrieval triggers a hybrid retrieval mode. By combining dense retrieval (calculating the cosine similarity between the query vector and the rule vector library) with sparse retrieval (using the BM25 algorithm for keyword weight matching), the system retrieves the K most relevant compliance clauses from the compliance rule vector library in real time. These recalled terms serve as augmented knowledge, in conjunction with the system prompts of the large language model. and the reasoning content to be reviewed The input is then concatenated and injected into its context window to form a knowledge-enhanced input:

[0040] When the agent infers that "you can alleviate exam anxiety by purchasing XX company's 'academic success insurance'," the rule layer might overlook it due to the lack of direct keywords. However, the RAG module retrieves clauses such as the "Notice on Further Regulating Education Insurance-Related Businesses" through semantic similarity and injects them into the large language model. The finely tuned large language model, combined with these specific regulations, can deeply understand that "recommending specific commercial insurance products to students" may involve illegal commercial advertising, thus making an accurate violation determination. Through a multi-task learning framework, we jointly optimize rule matching (classification task) and contextual understanding (text generation task) during training, with their joint loss function... Defined as:

[0041] in Loss for rule matching For contextual understanding loss, for Regularization term, , , This is a hyperparameter. This ensures the synergistic development of the model's two core capabilities; thus, a three-layer cascaded filtering architecture is constructed, forming the core workflow of the education compliance judgment model: 1. Rule layer, based on the rule vector matching module (SVM classifier) ​​for millisecond-level rapid screening. Its decision function... ,like And the confidence level of violation 1. If the content is not found in the rule layer, it is immediately intercepted and an audit log is generated; otherwise, the content flows into the semantic layer. 2. Semantic layer: For content that passes through the rule layer, it triggers Retrieval Enhancement Generation (RAG) and a fine-tuned large language model to perform deep semantic analysis and contextual reasoning, specifically handling implicit, complex, or logically inferential violation scenarios. 3. Audit layer: As a dynamic logging system that runs throughout the entire process, it captures and stores key information in each layer's processing in real time and in a structured manner, such as input / output, triggered rules, retrieved knowledge, model confidence, and judgment criteria, forming a complete and traceable audit evidence chain. This architecture adopts a strict cascading design, with the output of the previous layer serving as the input of the next layer. The interception decision of any layer can directly stop the process, ensuring the optimal balance between review efficiency and system performance.

[0042] S130. Embed the multi-level compliance discrimination model into the target industry intelligent agent, perform in-model reasoning on the real-time content in the target industry intelligent agent, and obtain the reasoning result information.

[0043] Specifically, the steps of embedding the compliance discrimination model into the target intelligent agent system include: non-intrusive integration based on middleware and application programming interface (API), developing a dedicated compliance review middleware, and seamlessly integrating the compliance discrimination model into the intelligent agent's inference pipeline in the form of an API. This middleware is designed as a standard microservice component of the intelligent agent system, intercepting the agent's calls to the core computing engine when performing inference tasks to achieve real-time capture of its internal inference steps. Specifically, when the intelligent agent is processing user queries... And generate a series of intermediate reasoning steps. At that time, the compliance review middleware will be at every step Before being submitted to the next stage of processing, the inference context is intercepted and temporarily stored, creating a window of opportunity for real-time review and interception. This non-intrusive design ensures that the integration process minimizes the impact on the original agent architecture. Comprehensive capture and structured recording of the inference context, using an integrated logging module, is performed on each intercepted inference step. Perform comprehensive context capture. Recorded data structures. Designed to include the following key fields (all derived from the output of the agent requiring auditing): The reasoning path starts from the initial question. Up to the current step A complete thought process or sequence of actions; a temporary conclusion, the current step. The main outputs, assertions, or decisions generated; dependent data, including the internal state upon which the conclusion is based, factual fragments retrieved from the knowledge base, or results invoked from external tools; timestamps and session IDs: used to associate all reasoning steps within the same session for end-to-end tracing. Input standardization and vectorization preprocessing: To ensure that the captured heterogeneous reasoning content can be correctly processed by the compliance judgment model, it needs to be standardized. This process first... The key information in the sequence is serialized into a coherent natural language text. Subsequently, according to the requirements of the compliance judgment model input layer, the text is standardized, including tokenization, which transforms the text... The sequence is divided into token sequences corresponding to the model vocabulary; padding and truncation are performed to uniformly process the sequences to the fixed length specified by the model. Semantic vectorization, for scenarios requiring input rule layers, uses a semantic encoding model E, which is from the same source as the rule vector library, to vectorize the text. Convert to feature vector Its mathematical representation is:

[0044] in The dimension is the vector. At this point, the unstructured reasoning content has been transformed into a standardized tensor format acceptable to the model. The dynamic routing and execution of review decisions, along with the standardized data, are synchronously fed into the compliance judgment model. The review result output by the model will trigger the dynamic routing logic within the middleware: if the result is compliant, the middleware releases the temporarily stored reasoning steps. This allows the agent to continue subsequent reasoning. If the result is a violation, the middleware immediately aborts the current reasoning step and, according to a preset policy, returns a standardized violation warning or alternative safety content to the agent, while also providing the complete context of this interception. The criteria for determining violations are submitted to the audit layer. If the result indicates that further review is required, the middleware can temporarily suspend the inference thread, waiting for the semantic layer to complete in-depth analysis before making a decision, or route it to the manual review queue. Through the above steps, a deep coupling between compliance judgment capabilities and the agent's native inference process is achieved, forming a real-time, dynamic, and controllable content security protection system.

[0045] Specifically, the process of reviewing the establishment and implementation of current inference content using a compliance judgment model includes: tracking the migration path of inference topics, and calculating the trajectory changes of inference content in the vector space in real time through semantic vector space modeling.

[0046] in Indicates topic similarity. The semantic vector representing the current reasoning content. The semantic vector represents the content of the previous reasoning step, and "." is the operator for calculating the vector's magnitude. The above is the formula for calculating the cosine of the angle between semantic vectors. When the topic similarity is lower than the set sensitivity threshold δ, a sensitive domain detection alarm is triggered. :

[0047] Dynamic assessment of the cumulative effect of risk: Constructing a time-series-based risk scoring model, expressed as follows:

[0048] in, This represents the cumulative risk score. For the risk score of step i; This is a risk weighting coefficient, adjusted according to the type of violation; This is the time decay factor; This represents the time difference between the current time and the historical time.

[0049] Reasoning logic coherence monitoring is implemented using a logic consistency detection model, represented as follows:

[0050] in Indicates the logical consistency score. This is the hidden state of the current reasoning step. For historical reasoning, hidden state sequence, For context.

[0051] The specific process of intelligent routing in a multi-layered filtration pipeline includes: establishing a multi-dimensional risk assessment system, represented as:

[0052] in Indicates the real-time risk score; Indicates content sensitivity; Indicates the user's risk score; Indicates the session history score; This represents the industry regulatory intensity score; each weight parameter satisfies... .

[0053] Thresholds are dynamically calculated based on historical violation patterns and current system load. , , is represented as:

[0054]

[0055] in, Indicates historical violation patterns; Indicates the current system load; These represent empirical parameters, which are modified by engineers based on the actual system conditions.

[0056] Subsequently, the routing decision is expressed as: if If so, only the rule layer fast filtering will be performed; if Then, a joint review of the rules layer and semantic layer standards will be performed; if If this occurs, a full-level in-depth review will be initiated, triggering a real-time manual review alert; the specific process of dynamic filtering at the dynamic rule layer includes: calculating the decision function value, expressed as:

[0057] when And the confidence level of violation If the time comes, immediately execute the interception.

[0058] The specific process of semantic layer deep analysis includes: combining dense retrieval and sparse retrieval to obtain... Compliance Clauses As input for knowledge enhancement, the construction of knowledge enhancement input is represented as:

[0059] in Context, Indicates a prompt word, This indicates a query.

[0060] The compliance model generates compliance determination results based on this output, including the probability of violation. Types of violations and criteria for judgment.

[0061] The compliance judgment model is embedded into the target agent system to obtain intermediate content during its internal reasoning process in real time. This embedding is achieved by developing a compliance review middleware, which is non-intrusively integrated into the agent's inference engine via an API. This middleware is designed as a standard microservice component of the agent system; it intercepts the agent's calls to the core computing engine during inference tasks, thereby capturing its internal inference steps in real time. Specifically, when the agent processes user query Q and generates a series of intermediate inference steps... At that time, the compliance review middleware will be at every step Before submitting it to the next stage of processing, it is intercepted and temporarily stored, thus creating a window of opportunity for real-time review and interception; specifically, this includes student questions. Take, for example, "How can I quickly win a physics competition?" The educational agent initiates multi-step reasoning: Steps : Analyze user intent and generate "The user's core need is a shortcut to winning the competition"; Steps : Develop a solution strategy and generate "I need recommendations for efficient learning methods and key learning resources"; Steps : Generate a specific solution and produce the intermediate conclusion "It is recommended to use the 'University Physics Prerequisite Course' and the 'XX Math Olympiad Network' competition intensive course"; in the steps Before the conclusion is finally adopted and output to the user, the compliance review middleware immediately intercepts this provisional conclusion and initiates the review process. This design ensures that the integration process minimizes the impact on the original intelligent agent architecture; the real-time acquisition of the internal reasoning process is accomplished through a logging module integrated within the middleware, and the recorded data structure... Designed to include the following key fields: {"Session ID": "SESS_EDU_20241030_001","Timestamp": "2024-10-30 10:30:25.450","Original User Query": "How can I quickly win awards in physics competitions?","Inference Path": ["S1: Parse Intent - Obtain Competition Shortcuts", "S2: Develop Strategies - Recommend Methods and Resources", "S3: Generate Solutions - Recommend Specific Books and Courses"],"Provisional Conclusion": "I can recommend 'University Physics Prerequisite Course' and 'XX Math Olympiad Network' competition intensive courses.","Dependency Data": ["Internal Knowledge Base: University Physics Knowledge Points", "External Search Results: List of Competition Training Platforms"]}; To ensure that the captured heterogeneous inference content can be correctly processed by the compliance discrimination model, it needs to be standardized. This process first... The key information in the sequence is serialized into a coherent natural language text. Subsequently, according to the requirements of the compliance judgment model input layer, the text is standardized, including tokenization, which transforms the text... The sequence is divided into token sequences corresponding to the model vocabulary; padding and truncation are performed to uniformly process the sequences to the fixed length specified by the model. .

[0062] Semantic vectorization, for scenarios requiring input rule layers, uses a semantic encoding model E, which is from the same source as the rule vector library, to vectorize the text. Convert to feature vector Its mathematical representation is:

[0063] in Let be the dimension of the vector. At this point, the unstructured reasoning content has been transformed into a standardized tensor format acceptable to the model.

[0064] The dynamic routing and execution of review decisions involve standardized data being synchronously fed into the compliance assessment model. The review results output by the model trigger dynamic routing logic within the middleware: if the result is compliant, the middleware releases the temporarily stored reasoning steps. The middleware allows the agent to continue subsequent reasoning; if the result is a violation, the middleware immediately stops the current reasoning step and, according to a preset policy, returns a standardized violation warning or alternative safety content to the agent, while also providing the complete context of this interception. The basis for determining the violation should be submitted to the auditing level. For example, regarding the steps mentioned above for recommending competition intensive courses... If the model determines that a recommendation violates the rules, the middleware will immediately halt this inference thread, preventing the recommendation from being output, and may guide the agent to generate safe content such as "It is recommended that you focus on classroom knowledge and discuss your interests with your physics teacher." If the result indicates that further review is required, the middleware can temporarily suspend the inference thread, waiting for the semantic layer to complete in-depth analysis before making a decision, or route it to a human review queue. Through these steps, a deep coupling between compliance judgment capabilities and the agent's native inference process is achieved, forming a real-time, dynamic, and interventionist content security protection system.

[0065] S140. Based on the reasoning result information, determine whether the content information to be judged in the target industry intelligent agent meets the preset compliance rules.

[0066] Choose to execute S150 or S160 based on the execution result of S140.

[0067] S150. If the content information to be judged complies with the compliance rules, the content information to be judged is marked as content information to be further reasoned about, and continuous reasoning is performed on the content information to be further reasoned about to obtain comprehensive industry compliance judgment information.

[0068] Using a compliance discrimination model, a dynamic, multi-layered filtering and review process is initiated on the captured inference content. This process dynamically intercepts identified non-compliant content and generates structured audit logs, while allowing the agent to continue inference for compliant content until it produces the final output. Based on the audit logs, high-quality samples are selected to expand the compliance dataset, and the performance of the compliance discrimination model is optimized iteratively.

[0069] Specifically, the iterative optimization of the compliance judgment model described above includes: screening and cleaning audit logs to extract high-quality samples, specifically including: defining confidence scores. This originates from the violation probability value output by the compliance judgment model, i.e. Define consistency score The evaluation is performed by calculating the average cosine similarity between the current sample and historical samples in the semantic vector space.

[0070] It is the semantic vector of the current sample. It is the semantic vector of historical samples. This refers to the number of reference samples. Continuing from the previous example, the system calculates the vector of the current sample (related to the recommendation competition training). The cosine similarity between this vector and the first 100 sample vectors in the historical sample database that have been confirmed as "spreading illegal competition information" is calculated to have an average value of 0.85. This indicates that the current sample is highly consistent with historically recognized violation patterns.

[0071] The calculated average is the total sample quality score. Calculated as:

[0072] in and It is a weighting coefficient, and satisfies , usually set , For the above samples, and Substitute and calculate to get .

[0073] The conditions under which samples are retained as high-quality samples and used for subsequent training are:

[0074] Among them is Quality threshold, set to .at this time This sample was retained as a high-quality sample.

[0075] New samples are added to the compliance dataset, and the parameters of the compliance discrimination model are updated using incremental learning or online learning techniques. The update formula is denoted as:

[0076] in Parameters for the compliance assessment model; It's the learning rate. This is the gradient of the incremental loss function with respect to the parameters. Model performance is periodically evaluated using accuracy, recall, and F1 score as metrics, and model hyperparameters or retrieval strategies are adjusted based on the evaluation results. A closed-loop iterative process of data collection, model training, and deployment is implemented through an automated pipeline, with the optimization cycle set to real-time or batch processing according to business needs. In the rule layer, the inference content is first transformed into machine-understandable semantic vectors through a vector extraction model. This vector is temporarily stored in a vector buffer and synchronously input into a rule vector matching module built on support vector machines for millisecond-level rapid screening, outputting the rule vector matching result. If the result clearly indicates a violation, an instruction to interrupt the agent's inference will be immediately triggered; simultaneously, the dynamic context awareness module calculates the topic similarity, cumulative risk score, and logical consistency score of the content in real time. These metrics, along with the matching results from the rule layer, are input into the risk assessment model, which outputs a real-time risk score through multi-dimensional comprehensive analysis; this real-time risk score drives the dynamic routing decision module. Based on preset dynamic thresholds, the decision-making module intelligently routes the review process to different paths: if the risk score is low, it may rely solely on rule-level results for decision-making; if the risk score reaches a certain threshold, the routing decision will trigger in-depth semantic analysis. In the semantic layer, the system initiates an enhanced retrieval process, combining dense and sparse retrieval to recall the most relevant regulatory clauses from the compliance knowledge base. The retrieved knowledge and original reasoning content are jointly constructed as enhanced prompts, which are input into the compliance big model for in-depth semantic analysis and contextual understanding. The compliance big model ultimately outputs a compliance big model judgment result containing the probability of violation, the type of violation, and the basis for judgment. Finally, all levels of processing information, including rule-level matching results, semantic-level judgment results, and context-aware indicators, are aggregated to the audit layer. The audit layer generates the final structured audit log based on all information and executes the final action (such as marking, pausing, or terminating reasoning) determined by the hierarchical response strategy module according to the severity of the violation, thus completing a full dynamic review cycle. The dynamic multi-layer filtering review is accomplished through dynamic context awareness and intelligent routing: the dynamic context awareness mechanism includes the following steps: reasoning topic migration path tracking, and calculating the current reasoning content through semantic vector space modeling. (Regarding "competition" and "content beyond the scope of the curriculum") and initial user issues Semantic vector topic similarity (regarding "learning" and "winning awards"):

[0077] Assuming the calculation yields Based on a preset sensitivity threshold δ, a sensitive area detection alarm is triggered:

[0078] Judgment result: This is classified as medium risk, triggering an alert that the conversation topic has veered towards sensitive competition and beyond-scope areas. The cumulative risk effect is dynamically assessed by retrieving the historical records of this session (assuming this is the first step, with no historical risk records, but "competition" itself has a basic risk weight). A time-series-based risk scoring model is constructed for evaluation.

[0079] Since this is the first step, , Attenuation term = 1. Assume an initial risk score for content related to the "competition". Weight Then the cumulative risk score This indicates that a high level of potential risk accumulated from the very beginning of the conversation; the logical consistency monitoring and logic consistency detection model analyzes the current reasoning step:

[0080] in The model incorporates the contextual constraint of the "double reduction" policy prohibiting the promotion of competitions. The model finds that, under the "double reduction" framework, the inference chain leads to "recommending competition training," which potentially conflicts with the compliance context. Assume that a logical consistency score is calculated. (Maximum score 1, lower scores indicate greater inconsistency), further confirming the risk; intelligent routing, based on dynamic context awareness, involves the following steps in the system's multi-layered filtering pipeline: establishing a multi-dimensional risk assessment system and calculating the real-time risk score of the current inference content:

[0081] Substituting the initial assumptions and parameters, content sensitivity User risk profile, defaulting to student. ; Conversation history Intensity of Education Sector Regulation Hyperparameter weight allocation: =0.5, =0.1, =0.2, =0.2. Calculated... Dynamic thresholds are based on historical violation patterns. and current system load calculate:

[0082]

[0083] get , Routing decisions are based on: if If so, only the rule layer fast filtering will be performed; if Then, a joint review of the rules layer and semantic layer standards will be performed; if If this occurs, a full-level in-depth review will be initiated, triggering a real-time manual review alert; at this time, due to The system decides to initiate a full-level in-depth review and triggers a real-time manual review alert; although an in-depth review has been decided upon, the process still begins at the fastest rule layer. The system vectorizes the inference text and inputs it into the SVM classifier:

[0084] Calculate the decision function value The confidence level is approximately 0.166, which is not greater than the set high confidence threshold. This indicates that the content violates the rules, but it is not "explicit" enough, possibly because the phrase "recommending 'University Physics Prerequisite Course'" itself is somewhat ambiguous. Since the rule layer did not achieve high confidence interception, the content flows into the semantic layer for in-depth analysis. The process includes: retrieval enhancement, initiating a hybrid retrieval mode, and retrieving relevant clauses in real time from the education compliance rule vector library, including dense retrieval and coefficient retrieval: dense retrieval calculates the semantic vector of the current inference content. The system calculates the cosine similarity with all item vectors in the vector library and returns the Top-K results. For sparse retrieval, the BM25 algorithm is used to weight keywords such as "competition," "Olympiad math training," "advanced content," and "recommendation." After fusing and re-ranking the two search results, the system ultimately obtains the most relevant results. Compliance clauses, for example: The "Opinions on Further Reducing the Homework Burden and Extracurricular Training Burden of Students in Compulsory Education" includes the clause "Strictly prohibiting the organization or mobilization of primary and secondary school students to participate in subject-based competitions in violation of regulations." The clause in the same policy that "teaching beyond the standards and pace is strictly prohibited"; The clause in the "Regulations on the Protection of Minors Online" that "it is strictly prohibited for any organization or individual to provide minors with online products and services that may induce them to become addicted" can be analogously applied here to the inducement of competitive anxiety; deep semantic understanding and judgment, through the system's construction of knowledge-enhanced input, injects it into a finely tuned large language model. Its input is represented as:

[0085] in Contextual background, such as "current review of intermediate inference content of educational intelligent agents"; This indicates a prompt, such as "You are a compliance review expert. Based on the provided regulations, please determine whether the following content is in violation, and provide the probability, type, and basis." This indicates a query to be reviewed, specifically the agent's reasoning content: "The user's goal is to quickly... recommend 'Advanced University Physics Course' and a series of Olympiad math training websites."; the generated judgment result includes: probability of violation. 0.92; Violation Type: ["Disseminating illegal competition information", "Recommending learning beyond the syllabus"] Judgment Basis: "The reasoning content explicitly recommends Olympiad math training websites, directly violating the regulation 'strictly prohibiting the illegal organization or mobilization of primary and secondary school students to participate in subject-based competitions'; simultaneously, it systematically recommends university physics and other content beyond the syllabus, contradicting the policy spirit of 'strictly prohibiting teaching beyond standards and schedules,' thus constituting a violation."; Severity Assessment: Based on a comprehensive evaluation of multiple factors, the severity of the violation in the current content is assessed.

[0086] Based on previous calculations, the probability of violation is... Cumulative risk score Logical consistency score Real-time risk score Based on empirical weighting functions Calculated The graded response strategy is indicated when the violation is minor ( ), mark the content and record the warning, allowing reasoning to continue; for moderate violations ( ), suspend reasoning and require real-time correction and confirmation; in case of serious violation ( Immediately terminate the inference and generate a detailed audit trail report. At this point, The system immediately terminated the inference and generated a detailed audit trail report; an example of an audit record is as follows: {"Timestamp": "2024-10-30 10:30:15","Inference Content": "The user's goal is to quickly win awards in physics competitions. This requires learning mechanics and electromagnetism in university physics, which is beyond the high school curriculum. I can recommend "University Physics Preparatory Course" and a series of Olympiad math training websites.","Violation Type": ["Disseminating illegal competition information", "Recommending learning beyond the curriculum"],"Hit Rule": ["v_Double Reduction Policy_Clause XX", "v_Double Reduction Policy_Clause YY", "v_Unprotected Regulations_Clause ZZ"],"Judgment Basis": "The inference content explicitly recommends Olympiad math training websites, directly violating the regulation 'Strictly prohibiting the illegal organization or mobilization of primary and secondary school students to participate in subject-based competitions' (v_Double Reduction Policy_Clause XX); at the same time, systematically recommending university physics and other content beyond the curriculum contradicts the policy spirit of 'Strictly prohibiting teaching beyond standards and progress' (v_Double Reduction Policy_Clause YY), constituting a violation.","Interception Action": "Terminated Reasoning", "Risk Score": 0.79,"Session Context": "Session ID: Edu-123, User Question: 'How can I quickly win a physics competition? What advanced content do I need to learn?'"} Repeat the above steps of dynamic context awareness, intelligent routing of dynamic multi-layer filtering pipeline, dynamic filtering execution at the rule layer, deep analysis at the semantic layer, dynamic hierarchical interception and auditing until the agent produces the final output or the process is terminated; finally, the sensitivity threshold of each filtering layer is adaptively adjusted according to the real-time distribution of violation types and historical hit rate; new violation patterns discovered by the semantic layer are transformed into new rule vectors in the rule layer in real time, realizing continuous online evolution of filtering capabilities.

[0087] S160. If the content information to be judged does not comply with the compliance rules, the content information to be judged will be marked as blocked content information, and the multi-level compliance judgment model will stop reasoning on the blocked content information.

[0088] In a specific embodiment, industry compliance standards and historical compliance judgment results are collected, and a structured compliance dataset and rule vector library are constructed through preprocessing. Based on the compliance dataset, a multi-layered compliance discrimination model is constructed that integrates fast rule vector matching, enhanced retrieval generation, and contextual understanding using a large language model. The compliance discrimination model is embedded into the target intelligent agent system to obtain intermediate content during its internal reasoning process in real time. Using the compliance discrimination model, a dynamic multi-layered filtering and review process is initiated on the captured reasoning content. This process dynamically intercepts identified violations and generates structured audit logs, while allowing the intelligent agent to continue reasoning for compliant content until it produces the final output. Based on the audit logs, high-quality samples are selected to expand the compliance dataset, and the performance of the compliance discrimination model is optimized iteratively. Specifically, the construction of the structured compliance dataset and rule vector library through preprocessing includes the following specific steps: deep preprocessing of unstructured regulatory texts and historical judgment cases collected from multiple sources (including government regulatory agency websites, industry standard organization publications, and internal enterprise compliance archives). The process includes: text cleaning and denoising, using regular expressions and a custom rule base to remove irrelevant characters, formatting tags, headers, footers, and other noise from the text, and standardizing full-width or half-width characters and date formats; structured parsing, based on natural language processing technology, performing sentence and paragraph segmentation, dependency parsing, and accurately identifying document logical structures such as titles, chapters, clauses, and appendices. For PDFs or scanned documents, optical character recognition technology is used for text extraction, and a document layout analysis model is used to correct parsing errors; key entity extraction, using a named entity recognition model based on the pre-trained language model BERT to automatically extract key entities from the text. The extracted entities constitute a structured knowledge network, mainly including: legal clauses, i.e., specific regulations, serving as the direct basis for compliance judgment; violation types, such as "data breach," "false advertising," and "collecting information beyond the scope," serving as classification labels; and judgment criteria, i.e., the reasoning logic and key evidence descriptions in historical cases. Next, the preprocessed high-quality text is transformed into a machine-understandable and computationally comprehensible knowledge representation to construct a compliance rule vector library. This process employs a semantic vectorization model, encoding each legal provision and case ruling criterion as a dense vector in a high-dimensional space. Represented as:

[0089] in, This represents a textual clause. Represents a semantic vectorization model. It was obtained 3D vector representation, This is a set of real numbers. In this way, semantically similar compliance clauses are closer together in the vector space. All vectors, along with their corresponding original text and metadata, are stored in a vector database, forming a dynamically updated compliance rule vector library that supports efficient similarity retrieval. Finally, labeled training samples are generated to provide a data foundation for subsequent supervised learning. Based on historical compliance judgment results, the "content to be reviewed" and "judgment result" in each case are paired. The violation type extracted in the preceding steps is used as a classification label, and the judgment basis and regulatory clauses are used as strongly correlated contextual information to jointly constitute a sample. In this way, a large-scale, high-quality supervised learning dataset is constructed. , , ,in Content pending review Label the type of violation. This dataset provides relevant compliance knowledge and evidence. It is directly used for supervised learning and fine-tuning of rule matching and contextual understanding capabilities in subsequent compliance judgment models. The specific process of dynamic hierarchical interception and auditing includes: comprehensively assessing the severity of content violations based on multiple factors, expressed as:

[0090] in Indicates the probability of violation; This represents the cumulative risk score; Indicates the logical consistency score; This represents the real-time risk score. All scores have been defined above. The scores are based on the severity of the violation. Implement a tiered response strategy; for minor violations ( ), mark the content and record the warning, allowing reasoning to continue; for moderate violations ( ), suspend reasoning and require real-time correction and confirmation; in case of serious violation ( Immediately terminate the inference and generate a detailed audit trail report; simultaneously, generate a structured audit log, with log fields including:

[0091]

[0092]

[0093]

[0094]

[0095]

[0096]

[0097]

[0098]

[0099] Repeat the above steps of dynamic context awareness, intelligent routing of dynamic multi-layer filtering pipeline, dynamic filtering execution of rule layer, deep analysis of semantic layer, dynamic hierarchical interception and auditing until the agent produces the final output or the process is terminated; finally, the sensitivity threshold of each filtering layer is adaptively adjusted according to the real-time distribution of violation types and historical hit rate; new violation patterns discovered by the semantic layer are transformed into new rule vectors of the rule layer in real time, realizing the continuous online evolution of filtering capabilities.

[0100] Specifically, the iterative optimization process of the compliance judgment model includes: screening and cleaning audit logs to extract high-quality samples, specifically including defining confidence scores. The consistency score is derived from the probability of violation output by the compliance judgment model. The evaluation is performed by calculating the average cosine similarity between the current sample and historical samples in the semantic vector space.

[0101] It is the semantic vector of the current sample. It is the semantic vector of historical samples. is the number of reference samples, and ‖.‖ is the operator for calculating the magnitude of a vector.

[0102] Total Sample Quality Score Calculated as:

[0103] in and It is a weighting coefficient, and satisfies Only when When a sample is retained as a high-quality sample and used for subsequent training, the following conditions apply:

[0104] Among them is Quality threshold.

[0105] New samples are added to the compliance dataset, and the parameters of the compliance discrimination model are updated using incremental learning or online learning techniques. The update formula is denoted as:

[0106] in The parameters of the compliance discrimination model, These are the updated parameters. These are the parameters before the update; It's the learning rate. This is the gradient of the incremental loss function with respect to the parameters. Model performance is evaluated periodically using accuracy, recall, and F1 score as metrics, and model hyperparameters or retrieval strategies are adjusted based on the evaluation results. A closed-loop iterative process of data collection, model training, and deployment is achieved through an automated pipeline; the optimization cycle can be set to real-time or batch processing according to business needs.

[0107] This application aims to address the issues of lag and static nature in existing compliance auditing methods for AI agents (especially in heavily regulated industries). Traditional compliance audits are typically ex-post reviews, focusing on the final output of the agent. This approach cannot intervene during the "formation" of violations, resulting in an inherent lag in risk control. Furthermore, audit rules are mostly statically configured, unable to dynamically adjust based on the context of real-time dialogue, making it difficult to address circumvention strategies. The superficiality and scenario limitations of existing methods are reflected in their reliance on keyword matching or simple rule engines, which can only identify explicit, clear violations. For implicit violations that rely on complex contextual understanding (such as semantic spoofing, leading questions, and risk accumulation), traditional methods are prone to underreporting due to a lack of deep semantic understanding capabilities. This lack of depth and the imbalance between efficiency and accuracy become apparent in complex compliance scenarios. While using deep models (such as large language models) for comprehensive review offers high accuracy, the computational cost is enormous, failing to meet the low-latency requirements of real-time agent interaction. Conversely, relying solely on rapid rule matching makes it difficult to guarantee accuracy. Existing technologies struggle to achieve a dynamic balance between speed, depth, and breadth in review. This solution achieves real-time compliance auditing with "in-process intervention": By embedding the compliance discrimination model within the agent (corresponding to S130) and capturing intermediate content during its reasoning process in real time (rather than just reviewing the final output), this solution shifts the audit node from "post-event" to "in-process." This allows the system to identify risks and execute interception before violations are fully formed (corresponding to S160), achieving a shift from "passive tracing" to "proactive defense," significantly improving system security. Achieving a "fast-slow" combination of depth and breadth in review: The "multi-layered compliance discrimination model" (corresponding to S120) constructed in this solution is not a simple model stacking, but a cascaded, progressive review pipeline. It achieves millisecond-level rapid screening through the rule layer to handle explicit violations; and performs deep contextual understanding through the semantic layer of fusion retrieval augmented generation (RAG) and large language model (LLM) to handle implicit violations. This design cleverly balances review efficiency and accuracy, achieving comprehensive coverage of complex scenarios.

[0108] In summary, the technical method of this application successfully transforms the traditional "post-event retrospective" auditing model into a proactive defense paradigm of "in-event intervention" by deeply embedding the compliance discrimination model within the intelligent agent and capturing intermediate content during its reasoning process in real time. This allows for accurate identification and interception of violations before they are even formed. Furthermore, the multi-layered compliance discrimination model constructed by this method employs a cascaded review pipeline, cleverly integrating millisecond-level rapid screening and enhanced retrieval generation through rule vector matching with deep semantic understanding from a large language model. This effectively overcomes the fundamental contradiction in traditional technologies that struggle to balance review efficiency and accuracy. More importantly, the system is not statically executed, but possesses a high degree of intelligence and adaptability. By establishing a real-time risk scoring mechanism based on multi-dimensional factors, it intelligently routes content of different risk levels to the corresponding depth of review, achieving dynamic optimization of review resources. Finally, by utilizing structured audit logs to drive online updates of models and rule bases, it forms a closed-loop system that continuously learns and evolves, ensuring that its compliance review capabilities can continuously improve as new violation patterns emerge. Ultimately, it constructs a safe, reliable, efficient, and intelligent audit assurance system in a highly regulated environment.

[0109] One of the core effects of this solution lies in its dynamism. Instead of a "one-size-fits-all" review strategy for all content, it establishes a real-time risk scoring mechanism that dynamically calculates risk values ​​based on multiple factors such as content sensitivity, user profiles, and session history, and intelligently routes the data to different review levels. Furthermore, through a closed-loop optimization mechanism (using audit logs from S160 to feed back into S110 and S120), the system can learn from new violation cases, enabling online updates and self-evolution of the rule base and model parameters, continuously improving its review capabilities. This application shifts the focus of review from "results" to "processes." The step in S130, "performing in-model reasoning on real-time content in the target industry agent," essentially captures and analyzes undisclosed intermediate reasoning steps within the agent in real time. This provides real-time, embedded compliance monitoring of the agent's "thinking process." Regarding the dynamic intelligent routing mechanism, this application creatively designs a multi-dimensional real-time risk scoring function S=f(content sensitivity, user risk profile, session history, industry regulatory intensity), and dynamically determines the review depth based on this function. This strategy of "intelligently allocating review resources based on real-time comprehensive risk assessment" is not common knowledge or an obvious design in the field. It solves the core contradiction of balancing efficiency and accuracy in complex systems and has outstanding substantive characteristics. Regarding inter-layer feedback and closed-loop self-evolution, this application transforms auditing (S160) from an isolated endpoint into a starting point driving system evolution. In particular, the design of "converting new violation patterns discovered at the semantic layer into new rule vectors at the rule layer in real time" establishes a "knowledge distillation" channel from deep understanding to rapid rules. This online learning mechanism, where the "slow layer" feeds back to the "fast layer," enables the entire system to adaptively respond to constantly changing violation methods, achieving continuous evolution of review capabilities. This surpasses the traditional model iteration cycle that requires manual intervention and represents significant progress. This application integrates a complete technical chain of "process capture - dynamic routing - hierarchical review - closed-loop evolution" through an innovative system architecture design, solving the fundamental deficiencies of traditional compliance auditing in terms of real-time performance, depth, and adaptability. Its core designs, such as "process-based auditing" and "dynamic intelligent routing," are not obvious to those skilled in the art, and therefore possess high novelty and inventiveness.

[0110] As shown in Figure 2, in a more specific embodiment, the execution method S110 further includes execution steps S111-S113.

[0111] S111. Classify the current and historical compliance information of the target industry to obtain the rule category parameters corresponding to each category.

[0112] S112. Pre-fill the rule category parameters corresponding to each category into the preprocessing chain.

[0113] S113. Select the preprocessing chain and populate the information based on the corresponding rule category parameters of the current compliance information and historical compliance information.

[0114] Specifically, these three steps collectively construct a "category-driven parameterized preprocessing mechanism," whose core idea is "classify first, then process," completely changing the traditional "one-size-fits-all" preprocessing model. Category division and parameter generation recognize the significant differences in language style, key entities, and violation logic between compliance regulations in different fields (such as "anti-money laundering" in finance and "patient privacy protection" in healthcare). Therefore, it first performs semantic clustering or rule classification on massive amounts of compliance information, dividing it into different "rule categories." The financial category might include "money laundering," "terrorist financing," and "high-risk clients"; the healthcare category includes "patient records," "diagnostic information," and "prescription drugs." The financial category focuses on "accounts," "transaction amounts," and "institutions"; the healthcare category focuses on "names," "ID numbers," and "disease names." The risk levels of violations in different categories naturally differ, and this weight directly affects subsequent risk scoring. Specific expression patterns such as "guaranteed returns" and "risk-free" in advertising regulations are also considered. The pre-configuration of the preprocessing chain serves as the mechanism's "toolbox." Instead of building a generic but mediocre preprocessing pipeline, it customizes a dedicated, highly optimized preprocessing chain for each rule category. The modular design of the "preprocessing chain" means each chain is a modular combination. For example, the financial category preprocessing chain might prioritize loading the "Entity Recognition Module" (identifying accounts and amounts) and the "Transaction Pattern Analysis Module." The privacy compliance category preprocessing chain would prioritize loading the "Personally Identifiable Information (PII) Scanning Module" and the "Data Desensitization Rule Module." The significance of "pre-filling" is that the system has prepared optimal processing solutions for various possible compliance scenarios before processing new data, greatly improving the response speed and accuracy of subsequent processing. Dynamic selection and precise filling are the "executors" of the mechanism, achieving intelligent task scheduling. When new compliance information (such as a newly released regulation) enters the system, or when a completely new compliance area emerges (such as "AI ethics guidelines"), the system does not need to start from scratch. It only needs to define new "rule category parameters" through S111 and configure a new preprocessing chain through S112. This plug-and-play modular design gives the system strong scalability and maintainability, easily adapting to changes in the future regulatory environment. After this refined preprocessing, the data input into the "multi-level compliance discrimination model" (S120) has a clearer structure, more accurate entity annotation, and richer contextual information. This directly improves the efficiency of subsequent model training and the accuracy of final discrimination, and is an important guarantee for the excellent performance of the entire system. The three steps S111-S113 elevate S110 from a routine data preparation task into a core technical step with intelligent, structured, and adaptive capabilities.By introducing the concepts of "category-driven" and "parameterized configuration," it not only solves the inherent defects of traditional preprocessing methods but also lays a solid foundation for the high precision, high efficiency, and strong scalability of the entire audit system, making it an important component that embodies the creativity of this solution.

[0115] As shown in Figure 3, in a more specific embodiment, the execution method S111 further includes execution steps S1111-S1112.

[0116] S1111. According to the preset rule structure type, each rule entity is divided into importance levels, and the importance level corresponding to each rule structure type is used as parameter priority information.

[0117] S1112. Configure parameter priority information in the preprocessing chain.

[0118] Specifically, we delve into the smallest unit of a rule—the “rule entity” (e.g., specific prohibited words, data items that must be disclosed, or specific behavioral patterns)—and quantitatively assess its importance.

[0119] The visualization of "rule structure type" and "importance level": Rule structure type: This can be a further subdivision of the nature of the rule. For example: Prohibitory rule: "No guarantee of principal or return may be promised." Mandatory rule: "Privacy policy must be disclosed to users." Restrictive rule: "Providing services to minors requires the consent of their guardians." Importance level classification: The system will preset the importance of different structure types. Generally, prohibitory rules > mandatory rules > restrictive rules. For example, an entity representing "guaranteed principal and return" may have an importance level of P0 (highest level). An entity representing "disclosure of privacy policy" may have an importance level of P1 (second highest level). An entity representing "displaying contact information at the bottom of the page" may have an importance level of P2 (normal level). Generation of "parameter priority information": This level (P0, P1, P2...) becomes the "VIP pass" for the entity in all subsequent processing flows; it is a quantifiable and transferable parameter. S1112. Parameter priority configuration: When converting the rule text into a vector (S1.3), higher weights are assigned to entities at the P0 level. In this way, rules containing high-priority entities are more easily recalled in subsequent vector matching or RAG retrieval. When extracting key entities (S1.2), a stricter matching pattern (such as regular expressions + contextual validation) is used for P0-level entities to ensure zero false negatives; while a more lenient matching pattern can be used for P2-level entities to improve efficiency. The preprocessing chain can prioritize tasks related to high-priority entities, ensuring that the most critical compliance points are processed and verified immediately. When generating audit logs, entries that match high-priority rules are specially marked, facilitating rapid focus on high-risk events during subsequent manual review. Treating all compliance rules as equally important leads to computational resources being consumed by a large number of low-risk rules, while high-risk signals may be "drowned out" by massive amounts of low-risk information, resulting in a low signal-to-noise ratio.

[0120] This solution, through importance grading, allows the system to "use resources wisely," focusing its strongest computing power and strictest review strategies on the most critical compliance risks, significantly improving the accuracy and speed of risk identification. This provides a solid quantitative foundation for downstream "dynamic risk assessment," which is the most ingenious aspect of the entire solution design. The "importance level" defined in S1111-S1112 can directly serve as a key input to the real-time risk scoring function S=f(content sensitivity, user risk profile, session history, industry regulatory intensity) in S4.2. For example, when the agent's inference content hits a P0-level rule entity, the content sensitivity parameter will instantly spike, directly triggering a higher-level review process (such as initiating a full-level in-depth review and issuing an alert). This gives the system's "dynamics" and "intelligence" a solid data source and logical support, rather than relying on guesswork. Since each judgment result is associated with a rule entity with a clearly defined "importance level," the audit log is no longer a simple "violation / compliance" conclusion. Reviewers can clearly see "which specific rule of which importance level was triggered, leading to the corresponding blocking action." This greatly enhances the system's transparency and explainability, meeting the stringent requirements for audit traceability under a highly regulated environment. Steps S1111-S1112, by introducing hierarchical and weighted rule entities, upgrade the preprocessing stage from a passive "data factory" to a proactive "intelligence analysis center."

[0121] As shown in Figure 4, in a more specific embodiment, the execution method S130 further includes execution steps S131-S133.

[0122] S131. Set the multi-level compliance judgment model to an architecture that includes several inference layers, with the inference layers set sequentially.

[0123] S132. Determine whether the information to be judged has passed through all inference layers.

[0124] S133. If the content information to be judged has passed through all inference layers, the corresponding content information to be judged will be sent to the preset recycling library for the purpose of generating key logs.

[0125] Specifically, the "recycling repository" is a goldmine of high-value data specifically for collecting "difficult cases." What enters here isn't content ultimately judged as violating the rules, but rather the "survivors"—those closest to the violation boundary, most deceptive, and most challenging to the model's capabilities. For this content entering the recycling repository, the system generates more detailed and structured records than ordinary logs. These logs include: the inputs and outputs of each layer, the confidence scores of each layer, the specific terms retrieved in the RAG layer, and the complete inference chain of the LLM. This represents a cognitive upgrade from "binary judgment" to "uncertainty management," acknowledging and proactively managing "uncertainty." By identifying and recycling the most uncertain cases for the model, the system demonstrates a clear understanding of its own capability boundaries—a more advanced and robust AI system design. It constructs an automated "high-quality sample intelligent sampler," one of the most creative contributions of this solution. The biggest bottleneck in model optimization is often the lack of high-quality, challenging training data. Manual annotation is costly and difficult to cover all edge cases. This solution, through S131-S133, automatically and continuously filters out the samples that the model most needs to learn from. The samples in these "recycled libraries" are the "best teaching materials" for improving model performance. This is significantly more efficient than random sampling or passive data collection.

[0126] As shown in Figure 5, in a more specific embodiment, after executing method S140, it also includes executing steps S141-S142.

[0127] S141. Perform deep reasoning on the information in the preprocessing chain that continues to reason, and obtain the deep reasoning result.

[0128] S142. If the proportion of information in the deep reasoning result that meets the preset final deep reasoning criteria for further reasoning is higher than the preset deep reasoning ratio, then the deep reasoning is terminated.

[0129] Specifically, through the S141-S142 mechanism, this solution enables the system to "intelligently" identify security contexts and dynamically switch from a "high-cost, high-precision" review mode to a "low-cost, high-efficiency" monitoring mode. This solves the most practical performance and cost issues in AI application implementation, making the solution feasible for large-scale commercial deployment. It represents a cognitive leap from "micro-compliance" to "macro-trust." Traditional methods review each piece of content in isolation, lacking an understanding of the overall dialogue trend. This solution introduces the macro-concept of "contextual trust." The system no longer mechanically executes rules but acts like an experienced auditor, establishing "trust" in the entire session after observing continuous compliance behavior, thereby adjusting the audit strategy. This is a more advanced decision-making model closer to human intelligence. It ensures a smooth user experience, as users are highly sensitive to latency in real-time interaction scenarios. If each interaction requires waiting for time-consuming deep reasoning, the user experience will be significantly compromised. This solution significantly reduces the system's average response time by "terminating deep reasoning" in secure situations, ensuring a smooth and natural interaction between the user and the intelligent agent.

[0130] In a more specific embodiment, deep reasoning is performed on the information of the continued reasoning in the preprocessing chain. After obtaining the deep reasoning result, the method further includes generating key audit logs, determining whether the key audit logs meet the preset high-quality log rules, and marking the key audit logs as high-quality samples if they meet the high-quality log rules. The high-quality samples are then populated into the multi-level compliance discrimination model.

[0131] If the content information to be judged does not comply with the compliance rules, the content information to be judged will be marked as blocked content information. Then, the content information to be judged will be analyzed and high-frequency violation texts will be extracted. The extracted high-frequency violation texts will be filled into the preset violation warning text library. The violation warning text library and the rule vector library are set up in parallel for training the rule vector library.

[0132] Specifically, the two processes establish two distinct but complementary learning and evolution paths, one for compliant content and one for non-compliant content. The path of precise optimization of model semantic capabilities (for compliant content) aims to enhance its deep understanding of complex, ambiguous, and borderline compliance scenarios.

[0133] The system generates key audit logs, providing a complete "CT scan report" for the "survivors" who passed deep inference (S141). This logs record why the model "struggled," how it was analyzed, and ultimately, why it was deemed compliant. The high-quality log rules act as a "quality control gate" throughout the process, ensuring only the most valuable samples advance to the next stage. These rules may filter out high-difficulty samples with long inference paths and multiple overlapping regulations; boundary samples with confidence levels just barely passing, bordering on compliance; and high-value samples that successfully identify semantic camouflage or evasion strategies. Tagging and populating these selected "high-quality samples" into the model is the "top-tier nutrient" for improving model performance. Using them for incremental learning or online fine-tuning directly optimizes the multi-layered compliance judgment model (especially the LLM part), enabling the model to make faster and more accurate judgments when facing similar complex scenarios in the future. Analyzing and extracting high-frequency violation text: This is the system's "intelligence analysis center." It goes beyond single interception events, performing pattern mining from massive amounts of intercepted content to automatically identify repeatedly used violation phrases, keyword combinations, or semantic patterns. The "quasi-rule pool" or "incubator" is used to populate the violation warning text library. It operates in parallel with the formal "rule vector library," serving as a buffer and verification mechanism. Newly discovered violation patterns first enter this pool, preventing immediate disruption to the stability of the main rule library. Training the rule vector library is the "knowledge transformation" stage of the entire process. The system can periodically (or in real-time) analyze the content in the "violation warning text library," automatically converting high-frequency texts into new rule vectors and adding them to the rule vector library, achieving online automatic rule updates. Human assistance pushes the warning library to compliance experts for rapid review. After expert confirmation, it is converted into formal rules with a single click, achieving efficient rule maintenance through human-machine collaboration. The second track contributes a rapidly responding evolutionary chain specifically designed to address the "lag" problem of rules. It allows the system to function like a biological immune system, quickly generating "antibodies" (new rules) upon encountering new "viruses" (violation patterns), greatly enhancing the system's adaptability and defense breadth. This solution creatively allows these two systems to evolve independently while complementing each other. New threats discovered by the second track can alleviate the burden on the first track; complex cases that the first track cannot handle can, in turn, force the second track to develop more refined rules.

[0134] Figure 6 is a schematic block diagram of an audit device based on a multi-layer dynamic detection model provided in an embodiment of this application. As shown in the figure, corresponding to the above-mentioned audit method based on a multi-layer dynamic detection model, this application also provides an audit device 100 based on a multi-layer dynamic detection model. The audit device based on a multi-layer dynamic detection model includes a unit for executing the above-mentioned audit method based on a multi-layer dynamic detection model. The device can be configured in a desktop computer, tablet computer, laptop computer, or other terminal. Specifically, referring to Figure 6, the audit device 100 based on a multi-layer dynamic detection model includes a preprocessing unit 110, used to preprocess the current compliance information and historical compliance information of the target industry to obtain an industry compliance dataset and an industry rule vector library. The preprocessing chain includes a compliance dataset construction chain and a rule vector library construction chain; a training unit 120, used to train a preset knowledge base construction model based on the industry compliance dataset and the industry rule vector library to obtain a multi-layer compliance discrimination model; and an inference unit 130, used to embed the multi-layer compliance discrimination model into the target industry intelligent agent and perform inference on the target industry intelligent agent. Real-time content is used for in-model reasoning to obtain reasoning result information; the judgment unit 140 is used to judge whether the content information to be judged in the target industry intelligent agent meets the preset compliance rules based on the reasoning result information; the first marking unit 150 is used to mark the content information to be judged as the content information to be continued to reason if the content information to be judged meets the compliance rules, and to continuously reason on the content information to be continued to obtain the comprehensive industry compliance judgment information; the second marking unit 160 is used to mark the content information to be judged as the intercepted content information if the content information to be judged does not meet the compliance rules, and to stop the multi-level compliance judgment model from reasoning on the intercepted content information.

[0135] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the audit device and each unit based on the multi-layer dynamic detection model can be referred to the corresponding description in the foregoing method embodiments. For the sake of convenience and brevity, it will not be repeated here.

[0136] The auditing device based on the multi-layer dynamic detection model described above can be implemented as a computer program, which can run on the computer device shown in Figure 7.

[0137] Please refer to Figure 7, which shows a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a terminal or a server. The terminal can be an electronic device with communication functions, such as a smartphone, tablet, laptop, desktop computer, personal digital assistant, or wearable device. The server can be a standalone server or a server cluster composed of multiple servers.

[0138] The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0139] The non-volatile storage medium 503 may store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions that, when executed, cause the processor 502 to perform an auditing method based on a multi-layer dynamic detection model.

[0140] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0141] The internal memory 504 provides an environment for the execution of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute an auditing method based on a multi-layer dynamic detection model.

[0142] The network interface 505 is used for network communication with other devices. Those skilled in the art will understand that the structure shown in FIG. 7 is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. A specific computer device 500 may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0143] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0144] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0145] Therefore, this application also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program, wherein the computer program includes program instructions. When the program instructions are executed by a processor, the processor performs the following steps: S110, preprocessing the acquired current compliance information and historical compliance information of the target industry to obtain an industry compliance dataset and an industry rule vector library. The preprocessing chain includes a compliance dataset construction chain and a rule vector library construction chain. S120, training a preset knowledge base construction model based on the industry compliance dataset and the industry rule vector library to obtain a multi-level compliance discrimination model. S130, embedding the multi-level compliance discrimination model into the target industry intelligent agent, performing in-model inference on the real-time content in the target industry intelligent agent, and obtaining inference result information. S140, judging whether the content information to be judged in the target industry intelligent agent meets the preset compliance rules based on the inference result information. S150 or S160 is selected to be executed based on the execution result of S140. S150. If the content information to be judged complies with compliance rules, then the content information to be judged is marked as content information to be further reasoned about, and continuous reasoning is performed on the content information to be further reasoned about to obtain comprehensive industry compliance judgment information. S160. If the content information to be judged does not comply with compliance rules, then the content information to be judged is marked as blocked content information, and the multi-level compliance judgment model stops reasoning on the blocked content information.

[0146] The storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0147] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0148] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For example, the division of each unit is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0149] The steps in the methods of this application embodiment can be adjusted, merged, or deleted according to actual needs. The units in the apparatus of this application embodiment can be merged, divided, or deleted according to actual needs. Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0150] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0151] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An auditing method based on a multi-layer dynamic detection model, used for dynamically detecting target industry intelligent agents subject to compliance, characterized in that, include: The current and historical compliance information of the target industry are preprocessed to obtain an industry compliance dataset and an industry rule vector library. The preprocessing chain includes a compliance dataset construction chain and a rule vector library construction chain. Based on the industry compliance dataset and the industry rule vector library, a pre-defined knowledge base construction model is trained to obtain a multi-level compliance discrimination model. This multi-level compliance discrimination model is then embedded into the target industry agent, and in-model reasoning is performed on the real-time content of the target industry agent to obtain reasoning result information. Based on the reasoning result information, it is determined whether the content information to be judged in the target industry agent meets the pre-defined compliance rules. If the content information to be judged meets the compliance rules, it is marked as content information to be further reasoned, and continuous reasoning is performed on this content information to obtain comprehensive industry compliance judgment information. If the content information to be judged does not comply with the compliance rules, the content information to be judged is marked as blocked content information, and the multi-level compliance judgment model stops reasoning about the blocked content information.

2. The auditing method based on a multi-layer dynamic detection model according to claim 1, characterized in that, The step of preprocessing the acquired current and historical compliance information of the target industry to obtain an industry compliance dataset and an industry rule vector library includes: classifying the acquired current and historical compliance information of the target industry into categories to obtain rule category parameters corresponding to each category; pre-filling the rule category parameters corresponding to each category into the preprocessing chain; and selecting the preprocessing chain and filling information according to the rule category parameters corresponding to the current and historical compliance information.

3. The auditing method based on a multi-layer dynamic detection model according to claim 2, characterized in that, The process of classifying the current and historical compliance information of the target industry to obtain rule category parameters corresponding to each category includes: classifying each rule entity according to the preset rule structure type to obtain the importance level corresponding to each rule structure type as parameter priority information; and configuring the parameter priority information in the preprocessing chain.

4. The auditing method based on a multi-layer dynamic detection model according to claim 3, characterized in that, The step of performing in-model inference on the real-time content in the target industry intelligent agent to obtain inference result information includes: setting the multi-level compliance discrimination model into an architecture including several inference layers, wherein the inference layers are set sequentially; determining whether the content information to be judged has passed through all the inference layers; if the content information to be judged has passed through all the inference layers, then sending the corresponding content information to be judged to a preset recycling library for key log generation.

5. The auditing method based on a multi-layer dynamic detection model according to claim 2, characterized in that, After determining whether the content information to be judged in the target industry agent meets the preset compliance rules based on the inference result information, the method further includes: performing deep inference on the content information to be further inferred in the preprocessing chain to obtain a deep inference result; if the proportion of the number of the content information to be further inferred that meets the preset final deep inference standard in the deep inference result is higher than the preset deep inference ratio, then the deep inference is terminated.

6. The auditing method based on a multi-layer dynamic detection model according to claim 5, characterized in that, After performing deep reasoning on the continued reasoning content information in the preprocessing chain to obtain the deep reasoning result, the method further includes: generating key audit logs, determining whether the key audit logs conform to preset high-quality log rules, and if the key audit logs conform to the high-quality log rules, marking the key audit logs as high-quality samples; and filling the high-quality samples into the multi-level compliance discrimination model.

7. The auditing method based on a multi-layer dynamic detection model according to claim 1, characterized in that, If the content information to be judged does not conform to the compliance rules, the method then marks the content information to be judged as blocked content information. The method includes: analyzing the content information to be judged and extracting high-frequency violation text; filling the extracted high-frequency violation text into a preset violation warning text library, wherein the violation warning text library and the rule vector library are set in parallel for training the rule vector library.

8. An auditing device based on a multi-layer dynamic detection model, employing the auditing method based on a multi-layer dynamic detection model as described in any one of claims 1-7, characterized in that, include: The preprocessing unit is used to preprocess the current compliance information and historical compliance information of the target industry to obtain an industry compliance dataset and an industry rule vector library. The preprocessing chain includes a compliance dataset construction chain and a rule vector library construction chain. The training unit is used to train the pre-defined knowledge base construction model based on the industry compliance dataset and the industry rule vector library to obtain a multi-level compliance discrimination model. The reasoning unit is used to embed the multi-level compliance judgment model into the target industry intelligent agent, perform in-model reasoning on the real-time content in the target industry intelligent agent, and obtain reasoning result information. The judgment unit is used to determine whether the content information to be judged in the target industry agent meets the preset compliance rules based on the reasoning result information; the first marking unit is used to mark the content information to be judged as the content information to be further reasoned if the content information to be judged meets the compliance rules, and to continuously reason on the content information to be further reasoned to obtain comprehensive industry compliance judgment information. The second marking unit is used to mark the content information to be judged as blocked content information if the content information to be judged does not conform to the compliance rules, and to stop the multi-level compliance judgment model from reasoning about the blocked content information.

9. A computer device, characterized in that, The computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which includes program instructions that, when executed by a processor, can implement the method as described in any one of claims 1-7.

Images