An intelligent decision-making method and system based on five-layer nested closed loops

Abstract

This invention, based on the academic innovation theory—the endogenous introspective governance paradigm—proposes an intelligent decision-making method and system based on a five-layer nested closed loop. It addresses the shortcomings of existing intelligent decision-making systems, such as the lack of independent closed-loop feedback, hierarchical scheduling, and insufficient adaptive optimization capabilities for functional units. Unlike existing decision-making systems that rely on external rule constraints for optimization, this solution proactively performs multi-dimensional pre-review before generating decision schemes by evaluating the motivational purity of value-anchored units, enabling the system to possess self-verification and evolutionary capabilities. The invention includes: acquiring decision request data and forming decision element data; processing units, scheduling the system, de-identifying and coordinating, providing feedback updates, and generating comprehensive outputs through a micro-loop layer, a large-loop layer, an ecological collaboration layer, an adaptive optimization layer, and a collaborative symbiosis layer; and generating decision schemes, early warning information, or freeze information. This system supports both decision-making and evaluation modes, sharing the same underlying five-layer nested closed-loop architecture and a four-dimensional review engine.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, data processing, intelligent decision support, and risk warning technologies, and particularly to an intelligent decision-making method based on a micro-circulation layer, a macro-circulation layer, an ecological collaboration layer, an adaptive optimization layer, and a synergistic symbiosis layer. Background Technology With the development of artificial intelligence, data processing, and intelligent decision support technologies, intelligent decision-making systems have been gradually applied in scenarios requiring the analysis, judgment, and output of processing results from multi-source data, such as government governance, tax collection and management, urban management, emergency command, industry supervision, enterprise operation, and education and training. Existing intelligent decision-making systems typically acquire business data, historical data, and rule data through a data acquisition module. They then utilize preset models, rule bases, or case libraries to identify, classify, match, and evaluate the input data, and output corresponding decision suggestions, risk warnings, or solutions based on the evaluation results. Some systems can also adjust model parameters, rule thresholds, or recommendation results to a certain extent based on user feedback, historical execution results, or changes in business indicators.

[0002] In existing technologies, common intelligent decision-making systems often adopt a single-layer process architecture, which completes one-time processing in the order of data input, data analysis, rule judgment, and solution output. This type of architecture can meet the data processing and auxiliary judgment needs in relatively fixed scenarios, but its functional modules typically only form a one-way call relationship. The results of module operation are mostly used for subsequent processes, making it difficult to form an independent closed loop within each module that combines goal setting, state calibration, boundary checking, logic execution, effect feedback, and parameter updates. Therefore, when input data, business scenarios, or judgment rules change, existing systems often require manual reconfiguration of rules, adjustment of models, or maintenance of knowledge bases, and the modules themselves do not adequately utilize feedback on operational deviations and processing effects.

[0003] Meanwhile, existing intelligent decision-making systems typically rely on fixed rule bases, historical case libraries, or preset templates for matching and output when generating solutions. For complex scenarios involving multiple constraints, multiple business entities, or multiple processing objectives, the system often can only search and call upon existing solutions, making it difficult to dynamically combine and adaptively adjust different functional logics based on current scenario data, constraint boundaries, historical feedback, and execution results. This leads to a mismatch between the system's output and the actual scenario, and a lack of unified scheduling, data feedback, and collaborative correction mechanisms between different functional modules.

[0004] Furthermore, in multi-entity or cross-departmental collaborative decision-making scenarios, different entities typically maintain their own data, rules, and historical processing records. While existing systems can achieve some data exchange through interfaces, they lack sufficient capabilities for anonymized interaction of cross-entity data, comparison of rule differences, complementary methodological experience, and utilization of processing result feedback. This makes it difficult to achieve collaborative optimization between entities while ensuring data security. For historical data, feedback data, and business change data generated during long-term operation, existing systems mostly remain at the level of statistical analysis or manual review, lacking continuous calibration of full-process operational data, identification of rule vulnerabilities, updating of method libraries, and processing for medium- and long-term trend projection.

[0005] Furthermore, risk identification in existing intelligent decision-making systems is mostly concentrated on verification after the solution is output or during execution, mainly used to detect anomalies that have occurred or are about to occur. This approach makes it difficult to pre-screen input data, path selection, rule boundaries, and execution methods during the decision-making solution generation stage, and it is also difficult to reverse the risk assessment results to the solution generation, rule correction, and model parameter adjustment processes, thus limiting the system's continuous optimization and closed-loop processing capabilities in complex decision-making scenarios.

[0006] In summary, existing technologies for intelligent decision-making systems suffer from several shortcomings. These include a lack of independent closed-loop feedback and adaptive iteration capabilities among functional modules; a lack of unified scheduling and data feedback mechanisms between different levels; insufficient cross-entity data collaboration and rule complementarity; and a lack of proactive linkage between risk identification and solution generation. Furthermore, existing intelligent decision-making systems are generally efficiency-driven or externally regulated in their technical approach. This means the system passively optimizes itself through preset rules, external annotation feedback, or manual configuration. Its decision review mechanism primarily focuses on compliance verification after the decision solution is generated, lacking an endogenous mechanism for proactively reflecting on decision-making motivations and pre-calibrating the decision path before decision initiation. It also lacks the closed-loop driving capability to automatically transform execution effect feedback into parameter iterations for each processing unit.

[0007] This invention proposes an endogenous, introspective intelligent decision-making method and system, which differs from the passive optimization mode in existing technologies that rely on external rule constraints or manual annotation feedback. By evaluating the motivation purity of the value anchoring unit in the micro-loop layer, verifying the target consistency of the direction calibration unit, and conducting multi-dimensional collision detection of the boundary constraint unit, the system proactively completes multi-dimensional pre-review before the decision scheme is generated. Furthermore, the data classification and feedback in the large loop layer drives each unit to complete micro-loop iteration, enabling the system to possess endogenous self-verification and self-evolution capabilities. This fundamentally solves the problems of difficulty in locating accumulated decision deviations, difficulties in cross-subject collaborative optimization, and highly lagging risk management in existing technologies. Summary of the Invention

[0008] To address the shortcomings of existing intelligent decision-making systems, such as the lack of independent closed-loop feedback and adaptive iteration capabilities among functional modules, the lack of unified scheduling and data feedback mechanisms between different levels, insufficient cross-entity data collaboration and rule complementarity, and the lack of pre-processing linkage between risk identification and solution generation, this invention proposes an intelligent decision-making method based on a five-layer nested closed loop. The method includes: Acquire decision request data, and perform element identification and normalization processing on the decision request data to obtain decision element data; The decision element data is input into the micro-circulation layer, and the value anchoring unit, direction calibration unit, boundary constraint unit, and method matching unit in the micro-circulation layer respectively form the unit processing results; wherein, the value anchoring unit generates a value anchoring processing result containing a motivation purity score based on the decision element data; The unit processing results are input into the large loop layer, which summarizes and judges the processing results of each unit according to the preset scheduling relationship, and returns the full process operation data to the corresponding functional unit according to the unit classification, driving each unit to complete the micro-loop iteration and obtain the system processing results; The system processing result is input into the ecological collaboration layer, which performs cross-subject collaborative verification on the system processing result based on the desensitized interactive data to obtain the collaborative processing result. The collaborative processing results are input into the adaptive optimization layer, which performs vulnerability warnings and updates on the processing rules, calibrates the evaluation parameters, and dynamically expands the case knowledge base based on historical operation data and feedback data to obtain the optimized processing results. The optimization results are input into the collaborative symbiosis layer, which then adaptively generates and outputs a decision scheme based on human-machine calibration data, risk identification results, time-series simulation results, and scenario adaptation parameters.

[0009] Furthermore, The decision request data includes decision application text, decision type, supporting material data, and historical decision data. The element identification and normalization process extracts decision object, decision objective, constraints, and historical correlation information based on the decision application text and supporting material data, and forms the decision element data.

[0010] Furthermore, The value anchoring unit generates value anchoring processing results based on the matching relationship between the decision element data and preset policy rule data, organizational target data, and historical decision data. These results include directional consistency evaluation results, motivation evaluation results, rule hit results, and motivation purity scores. The motivation evaluation results are generated based on the probability distribution of public interest orientation and private interest drive in the decision text. The rule hit results are generated based on collision detection between the decision elements and the preset boundary rule base.

[0011] Furthermore, The direction calibration unit iteratively generates multiple candidate execution paths within the compliance framework based on the decision element data, resource constraint data, time constraint data, and historical case data. It then performs a comprehensive evaluation of each candidate execution path based on policy fit, resource consumption, time adaptability, risk level, and historical success rate, generating a direction calibration processing result.

[0012] Furthermore, The boundary constraint unit decomposes the recommended execution path into a sequence of operation nodes based on the direction calibration processing result, and generates a boundary constraint processing result based on the matching relationship between the operation node data and the preset boundary rules; wherein, the preset boundary rules include mission red line rules, discipline bottom line rules, and ethical high-voltage line rules; the boundary constraint processing result includes the rule hit status, risk level, and correction suggestions of each operation node.

[0013] Furthermore, The method matching unit extracts decision features based on the decision element data, and based on the mapping relationship between the decision features and the methods and tools in historical successful cases, combined with the boundary constraint processing results, excludes methods and tools that do not meet the constraints, and generates method matching processing results; wherein, the method matching processing results include recommended combinations of methods and tools and usage guidelines for each tool.

[0014] Furthermore, The large loop layer summarizes the processing results of the units according to the processing order of value anchoring unit, direction calibration unit, boundary constraint unit and method matching unit, and generates a continue processing result, warning result or blocking result according to the processing status of each unit's processing result; when any unit outputs a blocking result, subsequent unit calls are stopped; the large loop layer also returns the entire process operation data to the corresponding functional unit by unit classification, and performs attribution analysis and deviation analysis on each unit based on the returned data.

[0015] Furthermore, The ecological collaboration layer receives de-identified interactive data corresponding to different subjects, and performs benchmarking processing on rule data, case data, or method tool data based on the de-identified interactive data to generate the collaborative processing result.

[0016] Furthermore, The adaptive optimization layer determines the deviation between the unit processing result and the actual feedback result based on the historical operation data and feedback data, and updates the processing rules, evaluation parameters or case data according to the deviation; wherein, the adaptive optimization layer also generates rule redundancy prompts or rule vulnerability warnings based on the rule triggering frequency and changes in decision effect; the adaptive optimization layer also automatically triggers parameter fine-tuning within a preset parameter range based on the effect scores of multiple consecutive decisions.

[0017] The method or system supports both decision-making and evaluation modes. In decision-making mode, the proposed decision scheme is pre-examined and the optimal solution is output. In evaluation mode, the executed or proposed decision scheme is post-examined and an evaluation report is output. Both modes share the same underlying five-layer nested closed-loop architecture and four-dimensional review engine.

[0018] The value anchoring unit, direction calibration unit, boundary constraint unit, and method matching unit constitute a four-dimensional review engine; the large loop layer constitutes a unified scheduling engine; the independent closed-loop operation of each unit in the micro loop layer and the data backflow of the large loop layer together constitute a two-way feedback mechanism; the five-layer nested closed-loop architecture as a whole constitutes a four-dimensional law intelligent decision-making system.

[0019] This invention proposes an intelligent decision-making system based on a five-layer nested closed loop, comprising: The data acquisition and processing module is used to acquire decision request data and perform element identification and normalization processing on the decision request data to obtain decision element data. The micro-circulation layer processing module is used to input the decision element data into the micro-circulation layer, and the value anchoring unit, direction calibration unit, boundary constraint unit and method matching unit in the micro-circulation layer respectively form the unit processing results; The large loop layer processing module is used to input the unit processing results into the large loop layer, and the large loop layer summarizes, judges and backflows the processing results of each unit according to a preset scheduling relationship to obtain the system processing result; The ecological collaboration layer processing module is used to input the system processing results into the ecological collaboration layer, and the ecological collaboration layer performs cross-subject collaborative verification on the system processing results based on the desensitized interactive data to obtain the collaborative processing results. The adaptive optimization layer processing module is used to input the collaborative processing results into the adaptive optimization layer, which updates the processing rules, evaluation parameters or case data based on historical operation data and feedback data to obtain the optimization processing results. The collaborative symbiotic layer processing module is used to input the optimization processing results into the collaborative symbiotic layer, which then generates and outputs a decision scheme based on human-machine calibration data, risk identification results, and time series extrapolation results.

[0020] This invention proposes an endogenous, introspective intelligent decision-making method and system, which differs from the passive optimization mode in existing technologies that relies on external rule constraints or manual annotation feedback. The endogenous, introspective intelligent decision-making method and system proactively completes multi-dimensional pre-review before generating a decision scheme through motivation purity assessment of the value anchoring unit in the micro-loop layer, target consistency verification of the direction calibration unit, and multi-dimensional collision detection of the boundary constraint unit. Furthermore, it drives each unit to complete micro-loop iterations through data classification and backflow in the large loop layer, enabling the system to possess endogenous self-verification and self-evolution capabilities. Specifically, the motivation purity assessment generates motivation evaluation results based on the probability distribution of public interest orientation and private interest drive in the decision text; the multi-dimensional collision detection includes three-line collision detection of mission red line rules, discipline bottom line rules, and ethical high-voltage line rules; and the data classification and backflow classifies and returns the entire process operation data to the corresponding functional units according to unit classification and drives parameter iterative updates.

[0021] The beneficial effects of this invention are: To address the aforementioned problems, this invention proposes an intelligent decision-making method and system based on a five-layer nested closed loop, which offers the following improvements: The micro-loop layer sets up value anchoring units, direction calibration units, boundary constraint units, and method matching units, and enables each unit to form an independent operating closed loop around target setting, state calibration, boundary self-checking, logic execution, effect feedback, and self-iteration. This ensures that each type of decision processing task is no longer just a one-time call node in a linear process, but can complete the verification, output, and feedback update within its own unit after receiving input data, thereby reducing the problem that the operation deviation of a single functional unit needs to rely entirely on manual maintenance.

[0022] The value anchoring unit transforms decision-making application texts, historical behavioral data, policy and regulatory texts, and organizational goal texts into computable semantic vectors, historical features, and rule-matching objects. It then generates consistency scores, risk markers, and blocking results through semantic similarity, classification probabilities, and rule engine matching results. This allows unstructured text information and historical data to be uniformly converted into quantifiable review criteria, thereby reducing the instability in identification caused by relying solely on manual reading or fixed keyword judgments. Specifically, the value anchoring unit uses a pre-trained language model to perform semantic encoding and similarity calculation on decision and policy texts, employs a multi-label text classification model to assess the probability of public-interest-oriented and private-interest-driven decision motivations, and uses a rule engine to perform three-line collision detection on decision elements.

[0023] The direction calibration unit generates candidate paths based on the decision objectives, resource constraints, time constraints, and historical case data processed through value anchoring. It then evaluates the candidate paths based on policy fit, resource requirements, time efficiency, risk coefficient, and historical success rate. This ensures that the path output no longer relies solely on a single historical case, but rather forms a rankable path result under the combined effect of constraints and historical execution effects. This improves the matching degree between path selection and current input conditions in complex decision-making scenarios.

[0024] The boundary constraint unit breaks down the recommended path into a sequence of operation steps, and performs collision detection between the operation type, operation object, operation scope, and involved resources and the mission red line rule, discipline bottom line rule, and ethical high-voltage line rule, respectively. This moves risk identification from the final result review stage to the path construction stage, thereby discovering specific risky operation steps before the solution is output, and providing clear data location and rule basis for subsequent path correction or elimination.

[0025] The method matching unit transforms decision complexity, number of involved parties, time span, degree of conflict, degree of resource competition, and time limit requirements into decision feature vectors. It then matches these vectors based on the method tool library, historical decision features, and method usage effects. This allows the recommendation of tool combinations to no longer rely solely on fixed templates or human experience selection, but to combine current decision features and existing execution effects to form processing method combinations that are suitable for the current scenario.

[0026] The large loop layer sets up a unified scheduling engine and performs serial scheduling according to the processing order of value anchoring unit, direction calibration unit, boundary constraint unit and method matching unit. This enables the output of each bottom unit to serve as the input or judgment basis for subsequent units. After a single round of scheduling is completed, the running data is classified and flowed back to the corresponding unit, thereby solving the problem in the existing single-layer process where modules only pass results and do not flow back deviation and effect data.

[0027] The main loop layer performs attribution analysis, deviation analysis, and effect assessment on the data of the entire process operation, and then applies the analysis results back to the parameters, thresholds, and rules in the micro loop layer. This establishes a correspondence between the final execution effect and the output of the front-end module, making it easier to locate the specific modules and evaluation indicators that affect the decision results. This reduces the problem of accumulated deviations after long-term system operation, making it impossible to determine the adjustment targets.

[0028] The ecosystem collaboration layer uses a compliant and anonymized data exchange channel to exchange rule summaries, model parameters, method effects, and historical processing data between different entities or systems. This enables cross-entity collaboration to complete rule difference comparison, experience complementarity, and standard alignment without directly exposing complete original business data, thus balancing data security requirements and multi-entity collaborative optimization needs.

[0029] The ecological collaboration layer employs processing methods such as model parameter aggregation, rule base difference comparison, and method effect recommendation to transform local operational experiences in different organizations into reusable model parameters, rule supplement suggestions, or method recommendation results. This reduces information silos formed when each entity maintains the system independently and improves the consistency of rule boundaries and method selection in cross-entity scenarios.

[0030] The adaptive optimization layer connects to the full-level runtime database, historical case database, and scenario business database, and continuously captures data such as decision results, module scores, actual effects, and feedback sources. This makes system optimization no longer dependent on a single manual review, but uses continuous runtime data as the data foundation for parameter calibration, rule updates, vulnerability identification, and method library expansion, thereby improving the maintainability and scenario adaptability of the system in the long-term operation process.

[0031] The adaptive optimization layer establishes a performance monitoring queue based on the scores of the most recent decision effects, and triggers parameter fine-tuning when the scores are continuously lower than the preset baseline. This enables the system to automatically identify the performance decline trend based on the recent running effect, and adjust the weight, threshold or rule configuration within the limited parameter search range. This reduces the problem of output results deviating from the current scenario due to the continued use of old parameters after changes in the business environment.

[0032] The adaptive optimization layer detects rule triggering frequency, decline in decision effectiveness, and abnormal decision patterns. When a rule is triggered at a low frequency for a long period or when the effectiveness of a certain type of decision continues to decline but the rule is not triggered, it generates rule redundancy or rule vulnerability prompts. This transforms rule base maintenance from passive manual inspection to proactive discovery based on runtime data, thereby improving the rule system's coverage of new scenarios and abnormal patterns.

[0033] The collaborative symbiosis layer integrates a two-way calibration mechanism between humans and machines, providing the deviations, risks, and early warning results automatically identified by the system to human confirmation, and writing the human correction results back into the system's baseline parameters, rule thresholds, or model update data. This ensures that human intervention is not just a modification of a single result, but can become the basis for updating the subsequent automatic processing logic of the system, thereby improving the problem that human feedback in traditional human-computer interaction is difficult to be transformed into system capabilities.

[0034] The synergistic layer introduces long-term time series extrapolation and causal relationship modeling, taking historical decision sequences, socio-economic indicators, policy event codes, and cross-organizational feedback data as predictive inputs, and outputting trend results and confidence intervals for a future period. This enables the system to further utilize long-term time series dependencies and causal relationship information beyond short-term data processing, thereby enhancing its ability to analyze medium- and long-term trends and chain effects.

[0035] The collaborative symbiotic layer performs proactive risk identification simultaneously during the solution generation process. It calculates the risk probability based on compliance scores, target consistency scores, and risk assessment model outputs, and then generates warnings, suspensions, or freezes based on different thresholds. This allows the risk judgment results to directly participate in solution generation and process control, rather than being checked separately after solution generation, thereby improving the pre-risk control capability in the decision-making process.

[0036] The collaborative symbiotic layer dynamically combines functional logic based on scenario data, business requirements, constraints, historical cases, and method tool libraries to generate decision-making solutions corresponding to the current input conditions. This makes the solution output no longer limited to direct calls to fixed knowledge bases or preset templates, but can form combined solutions adapted to different scenarios based on the data processing results of multiple modules, thereby improving the system's adaptability to complex and differentiated business scenarios.

[0037] RESTful interfaces, JSON data format, gRPC module communication, and structured database tables collectively define the transmission methods of data input, processing results, scoring records, rule information, case information, and feedback information between different levels and modules. This gives the five-layer nested closed-loop architecture a clear data flow carrier and engineering implementation path, thereby avoiding remaining at the level of abstract decision rules or management processes.

[0038] The review progress page, review results page, decision feedback page, and historical decision analysis page correspond to the processing status, scoring results, risk markers, and feedback data of the backend modules, forming a continuous interactive link between user input, system processing, result display, and execution feedback. This facilitates the transformation of manual confirmation, actual execution effects, and subsequent improvement suggestions into data sources that the system can record, trace, and update.

[0039] Unlike existing intelligent decision-making systems that rely on passive optimization through external rule constraints or manual annotation feedback, this invention proactively performs multi-dimensional pre-review before generating a decision scheme by using the motivation purity assessment of the value anchoring unit, the target consistency verification of the direction calibration unit, and the three-line collision detection of the boundary constraint unit. Furthermore, data feedback from the large loop layer drives each unit to complete closed-loop iteration, enabling the system to possess inherent self-verification and self-evolution capabilities. This system simultaneously supports decision-making and evaluation modes. In decision-making mode, it performs pre-review of the proposed decision scheme and outputs the optimal solution. In evaluation mode, it performs post-review of executed or proposed decision schemes and outputs an evaluation report. Both modes share the same underlying five-layer nested closed-loop architecture and four-dimensional review engine. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0041] Figure 1 This is a schematic diagram of the overall process of the method of the present invention.

[0042] Figure 2 This is a schematic diagram of the micro-circulation operation structure of the method of the present invention.

[0043] Figure 3 This is a schematic diagram of the large loop layer scheduling process of the method of the present invention.

[0044] Figure 4 This is a schematic diagram of data interaction in the ecological collaboration layer of the method of the present invention.

[0045] Figure 5 This is a schematic diagram of the adaptive optimization layer logic of the method of the present invention.

[0046] Figure 6 This is a schematic diagram of the internal functional structure of the synergistic layer in the method of the present invention.

[0047] Figure 7 This is a schematic diagram of the human-machine bidirectional calibration workflow of the method of the present invention.

[0048] Figure 8 This is a schematic diagram illustrating the long-cycle time-series deduction principle of the method of the present invention.

[0049] Figure 9 This is a schematic diagram of the proactive risk identification workflow of the method of the present invention.

[0050] Figure 10This is a schematic diagram illustrating the adaptive scheme generation logic of the method of the present invention. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] Combination Figure 1 This invention proposes an intelligent decision-making method based on a five-layer nested closed loop, the method comprising: Acquire decision request data, and perform element identification and normalization processing on the decision request data to obtain decision element data; Furthermore, The decision request data includes decision application text, decision type, supporting material data, and historical decision data. The element identification and normalization process extracts decision object, decision objective, constraints, and historical correlation information based on the decision application text and supporting material data, and forms the decision element data.

[0053] The decision element data is input into the micro-circulation layer, and the value anchoring unit, direction calibration unit, boundary constraint unit and method matching unit in the micro-circulation layer respectively form the unit processing results; Furthermore, The value anchoring unit generates value anchoring processing results based on the matching relationship between the decision element data and preset policy rule data, organizational target data and historical decision data.

[0054] Furthermore, The direction calibration unit generates candidate execution paths based on the decision element data, resource constraint data, time constraint data, and historical case data, and performs matching and evaluation on the candidate execution paths to generate direction calibration processing results.

[0055] Furthermore, The boundary constraint unit generates operation node data based on the direction calibration processing result, and generates boundary constraint processing result based on the matching relationship between the operation node data and the preset boundary rules.

[0056] Furthermore, The method matching unit generates method matching processing results based on the matching relationship between the decision element data, boundary constraint processing results, and method tool data.

[0057] The unit processing results are input into the large loop layer, which summarizes, judges, and backflows the unit processing results according to the preset scheduling relationship to obtain the system processing result; Furthermore, The large loop layer summarizes the processing results of the units according to the processing order of the value anchoring unit, direction calibration unit, boundary constraint unit and method matching unit, and generates a continue processing result, warning result or blocking result according to the processing status of each unit's processing result.

[0058] The system processing result is input into the ecological collaboration layer, which performs cross-subject collaborative verification on the system processing result based on the desensitized interactive data to obtain the collaborative processing result. Furthermore, The ecological collaboration layer receives de-identified interactive data corresponding to different subjects, and performs benchmarking processing on rule data, case data, or method tool data based on the de-identified interactive data to generate the collaborative processing result.

[0059] The collaborative processing results are input into the adaptive optimization layer, which updates the processing rules, evaluation parameters, or case data based on historical operation data and feedback data to obtain the optimized processing results. Furthermore, The adaptive optimization layer determines the deviation between the unit processing result and the actual feedback result based on the historical operation data and feedback data, and updates the processing rules, evaluation parameters or case data according to the deviation.

[0060] The optimization results are input into the collaborative symbiosis layer, which generates and outputs a decision scheme based on human-machine calibration data, risk identification results, and time series extrapolation results.

[0061] This invention proposes an intelligent decision-making system based on a five-layer nested closed loop, comprising: The data acquisition and processing module is used to acquire decision request data and perform element identification and normalization processing on the decision request data to obtain decision element data. The micro-circulation layer processing module is used to input the decision element data into the micro-circulation layer, and the value anchoring unit, direction calibration unit, boundary constraint unit and method matching unit in the micro-circulation layer respectively form the unit processing results; The large loop layer processing module is used to input the unit processing results into the large loop layer, and the large loop layer summarizes, judges and backflows the processing results of each unit according to a preset scheduling relationship to obtain the system processing result; The ecological collaboration layer processing module is used to input the system processing results into the ecological collaboration layer, and the ecological collaboration layer performs cross-subject collaborative verification on the system processing results based on the desensitized interactive data to obtain the collaborative processing results. The adaptive optimization layer processing module is used to input the collaborative processing results into the adaptive optimization layer, which updates the processing rules, evaluation parameters or case data based on historical operation data and feedback data to obtain the optimization processing results. The collaborative symbiotic layer processing module is used to input the optimization processing results into the collaborative symbiotic layer, which then generates and outputs a decision scheme based on human-machine calibration data, risk identification results, and time series extrapolation results.

[0062] The present invention will be further described in detail below with reference to specific embodiments, but the scope of protection of the present invention is not limited to the following embodiments. Unless otherwise specified in the following embodiments, conventional conditions or existing test conditions were followed.

[0063] This invention provides an intelligent decision-making method and system based on a five-layer nested closed loop, comprising four hierarchical embodiments corresponding to different application scenarios: Embodiment 1 is a two-layer architecture, suitable for independent basic decision-making scenarios such as grassroots government service windows, single-department business, routine school administration, and grassroots service units; Embodiment 2 is a three-layer linkage architecture with an overlay of an ecological collaboration layer, suitable for multi-subject collaborative decision-making scenarios such as cross-departmental government linkage, regional education consortia, multi-unit joint supervision, and regional joint defense and dispatch; Embodiment 3 is a four-layer architecture with an overlay of an adaptive optimization layer, suitable for scenarios requiring global adaptive optimization such as routine industry supervision, group enterprise operation, comprehensive teaching quality control, and routine maintenance of national defense logistics; Embodiment 4 is a five-layer complete architecture, integrating all five layers of functions, suitable for complex comprehensive decision-making scenarios such as major government affairs assessment, fiscal and tax macro-control, urban comprehensive emergency command, high-end talent training planning, military command and dispatch, and national defense strategic assessment. Each embodiment has a complete data input, logical processing, and result output chain, and can be directly deployed in an engineering manner.

[0064] Example 1 Two-layer architecture This embodiment provides an intelligent decision-making method and system based on a five-layer nested closed loop, applicable to intelligent decision support systems. The method processes decision request data layer by layer through micro-loop layers and large loop layers, ensuring that the decision request data undergoes sequential unit-level processing and system-level scheduling.

[0065] S1. Receive decision request data, which includes decision application text, decision type, decision-maker identifier, organization identifier, supporting material data, and historical decision data; perform format verification, text extraction, entity recognition, and field normalization on the decision application text and supporting material data to obtain decision element data; write the decision element data, decision type, and historical decision data into the decision request record and use them as input data for the micro-loop layer.

[0066] S2. The value anchoring unit in the micro-loop layer receives decision element data, historical decision data, preset policy rule data, and organizational goal data. It converts the decision application text, policy rule data, and organizational goal data into semantic feature data, and converts the historical decision data into historical behavior feature data. Based on the semantic matching relationship between the decision application text, policy rule data, and organizational goal data, it obtains the direction consistency evaluation result. Based on the decision application text and historical behavior feature data, it obtains the motivation evaluation result. Based on the matching relationship between the decision element data and the preset rule base, it obtains the rule hit result. It combines the direction consistency evaluation result, motivation evaluation result, and rule hit result into the value anchoring processing result and sends the value anchoring processing result to the unified scheduling engine of the large loop layer.

[0067] S3. The unified scheduling engine receives the value anchoring processing result. When the value anchoring processing result is in a blocking state, it generates a blocking report and ends the automatic processing flow of the current decision request. When the value anchoring processing result is in a warning state, it sends the warning information and the value anchoring processing result to the manual confirmation end, and determines whether to continue processing or end processing based on the manual confirmation result. When the value anchoring processing result is in a pass state, it sends the decision target data, resource constraint data, time constraint data, and value anchoring processing result to the direction calibration unit.

[0068] S4. The direction calibration unit receives decision target data, resource constraint data, time constraint data, value anchoring processing results, and historical case data. Based on the decision target data, resource constraint data, and time constraint data, it forms the current decision scenario characteristics. It performs similarity matching between the current decision scenario characteristics and the scenario characteristics in the historical case data to filter out candidate reference cases. Based on the candidate reference cases, resource constraint data, and time constraint data, it generates multiple candidate execution paths and evaluates the policy fit, resource consumption, time adaptability, risk level, and historical execution effect of each candidate execution path. It sorts the candidate execution paths according to the comprehensive evaluation results to obtain a recommended execution path set and sends the recommended execution path set to the boundary constraint unit.

[0069] S5. The boundary constraint unit receives the recommended execution path set and the preset boundary rule base. It breaks down each recommended execution path into multiple operation nodes and extracts the operation type, operation object, operation scope, and resources involved for each operation node. It matches each operation node with the mission red line rule, discipline bottom line rule, and ethical high-voltage line rule in the preset boundary rule base to obtain the rule hit status and risk level of each operation node. When there are operation nodes that need to be excluded in the recommended execution path, the corresponding recommended execution path is marked as an unavailable path. When there are operation nodes that need to be corrected in the recommended execution path, a correction suggestion is generated and marked as a path to be confirmed. When there is no rule hit status in the recommended execution path, it is marked as an available path. The path boundary review results, correction suggestions, and available path data are sent to the method matching unit.

[0070] S6. The method matching unit receives available path data, path data to be confirmed, path boundary review results, decision element data, and method tool library data. Based on the decision element data, it extracts decision features such as the number of involved entities, processing time, degree of resource competition, degree of conflict, and complexity. It matches the decision features with the method application scenario tags in the method tool library and combines the method usage effects in similar scenarios in historical decisions to obtain a candidate method set. Based on resource constraints, time constraints, and path boundary review results, it eliminates unsuitable candidate methods to form a recommended method combination. It combines the recommended method combination with the corresponding available path data to form an initial decision scheme and sends it to the collaborative symbiosis layer.

[0071] S7. The collaborative symbiotic layer receives the initial decision-making scheme, value anchoring processing results, recommended execution path set, path boundary review results, and recommended method combination, and performs comprehensive verification of the initial decision-making scheme. During the comprehensive verification process, the collaborative symbiotic layer forms a risk level based on the compliance evaluation results, target consistency evaluation results, and risk assessment results, and determines the processing status of the initial decision-making scheme based on the risk level. When the risk level meets the output conditions, the decision-making scheme is output. When the risk level meets the warning conditions, a warning message is generated and sent to the manual confirmation end. When the risk level meets the freeze conditions, the current decision-making process is frozen and the reason for the freeze is output.

[0072] S8. After the decision scheme is output, the execution feedback data of the decision scheme is collected. The execution feedback data includes actual execution results, effect scores, abnormal events, manual evaluations, and business indicator change data. The execution feedback data is associated and stored with the value anchoring processing results, the recommended execution path set, the path boundary review results, and the recommended method combination to form full-process operation data. The full-process operation data is fed back to the large loop layer, which performs attribution analysis on the deviation between the output results of each unit and the actual execution results, and sends the attribution analysis results to the value anchoring unit, the direction calibration unit, the boundary constraint unit, and the method matching unit, respectively.

[0073] S9, the adaptive optimization layer receives full-process operation data, attribution analysis results, historical case data, and scenario business data. Based on the changes in the deviation between the results of multiple consecutive decision-making processes and the actual execution results, it determines the parameters, rules, or method data that need to be adjusted. When there is a continuous deviation between the value anchoring processing results and the actual execution results, it adjusts the evaluation weights or judgment thresholds in the value anchoring unit. When there is a continuous deviation between the direction calibration results and the actual execution results, it updates the historical case data or path evaluation rules. When there is a continuous deviation between the boundary constraint results and the actual risk events, it generates rule supplement suggestions or rule redundancy prompts. When there is a continuous deviation between the method matching results and the actual execution effect, it updates the applicable scenario tags or effect scores in the method tool library. The updated parameters, rules, cases, and method data are returned to the corresponding units for subsequent decision request processing.

[0074] When extracting semantic features from decision application texts, policy and rule data, and organizational goal data, the value anchoring unit can use a pre-trained text encoding model to convert texts from different sources into semantic vectors of a unified dimension, and form a directional consistency evaluation result based on the similarity between semantic vectors. For generating the motivation evaluation result, a text classification model can be used to identify the probability of different motivation categories in the decision application text, and a comprehensive evaluation result can be formed by combining the motivation classification results in the decision-maker's historical decision records. The pre-trained text encoding model is bert-base-chinese, employing a 12-layer Transformer structure with 768-dimensional hidden layers, 12 multi-head attention heads, approximately 110M parameters, a fine-tuning learning rate of 2e-5, a batch size of 16, 3 training epochs, an AdamW optimizer, and a maximum sequence length of 512. The text classification model is TextCNN, with 300-dimensional word vectors, convolutional kernel sizes of 3, 4, and 5 (128 kernels per size), max pooling, 128-dimensional fully connected layers, and a Softmax three-class output layer, outputting probabilities of public interest orientation, private interest drive, and mixed motives. The training data for the TextCNN model consists of no fewer than 1000 manually annotated decision texts, with annotation rules for public interest orientation, private interest drive, and mixed motives, and an annotation consistency Kappa coefficient of no less than 0.7.

[0075] The preset rule base in the value anchoring unit includes rule identifier, rule type, triggering condition, processing action, and severity fields. When the decision element data matches a blocking rule, the value anchoring unit outputs a blocking status. When the decision element data matches a warning rule, the value anchoring unit outputs a warning status. When the decision element data does not match a blocking rule or a warning rule, the value anchoring unit outputs a pass status.

[0076] When the direction calibration unit generates candidate execution paths, it can first select similar cases from historical case data based on the characteristics of the current decision-making scenario, and then generate multiple candidate execution paths within the scope of resource constraints and time constraints. Each candidate execution path forms a path evaluation record, which includes path ranking, path content, resource evaluation results, time evaluation results, risk evaluation results, and comprehensive evaluation results.

[0077] When the boundary constraint unit performs boundary review on the recommended execution path, it can decompose the recommended execution path into a sequence of operation nodes with sequential relationships, and establish a corresponding boundary review record for each operation node. The boundary review record includes the rule hit type, the content of the hit rule, the risk level, the correction suggestion, and the review time, so that subsequent manual confirmation or system correction can be located to the specific operation node.

[0078] The method tool library in the method matching unit is a structured dataset. Each method tool data corresponds to the storage of applicable scenario labels, resource requirements, time requirements, historical usage frequency, and historical effect scores. The method matching unit filters method tools based on the matching relationship between the current decision characteristics and the applicable scenario labels, and excludes method tools that do not meet the current constraints based on resource requirements and time requirements.

[0079] Example 2 Three-layer linkage architecture This embodiment adds an ecological collaboration layer to the two-layer architecture of embodiment 1 to form a three-layer linkage architecture, which is suitable for multi-entity collaborative decision-making scenarios such as cross-departmental government affairs linkage, regional education consortia, multi-unit joint supervision, and regional joint prevention and control.

[0080] The unified scheduling engine of the large loop layer sequentially schedules the four units in the micro loop layer and records the input data, output data, processing status and processing time of each unit. When any unit outputs a blocking status, the unified scheduling engine stops subsequent unit calls and generates a blocking report. When any unit outputs a warning status, the unified scheduling engine sends the corresponding warning information to the manual confirmation end and determines whether to continue calling subsequent units based on the manual confirmation result.

[0081] The ecological collaboration layer establishes a compliant and anonymized data exchange channel among multiple entities. Each entity uploads its local rule summary, model parameter summary, method effect data, and historical processing statistics to the ecological collaboration layer. The ecological collaboration layer performs difference comparison and aggregation processing on the data uploaded by different entities, generating rule supplement suggestions, model parameter update suggestions, or method recommendation results, and then distributes the processing results to the adaptive optimization layer of the corresponding entity. The anonymized interactive data is transmitted using the AES-256 encryption algorithm, and k-anonymization is used for fields involving personal privacy and trade secrets to ensure that the data is anonymized before leaving the local system. The difference comparison and aggregation processing uses the FedAvg algorithm in federated learning. Each entity maintains its model parameters locally and periodically uploads the parameters to the ecological collaboration server. The server performs a weighted average of the parameters from each entity before distributing them, achieving cross-entity value standard alignment.

[0082] Example 3 Four-layer architecture This embodiment adds an adaptive optimization layer to the three-layer architecture of embodiment 2, forming a four-layer architecture. It is suitable for scenarios that require global adaptive optimization, such as routine industry supervision, group enterprise operation, comprehensive teaching quality control, and routine operation and maintenance of national defense logistics.

[0083] An adaptive optimization layer establishes a performance monitoring queue to store the predicted scores, actual effect scores, and deviation values ​​corresponding to the most recent decision requests. When consecutive deviations in the performance monitoring queue meet preset trigger conditions, the adaptive optimization layer initiates parameter fine-tuning and updates the weights, thresholds, or rule configurations of relevant units within a preset adjustment range. The updated configuration is first written to the candidate configuration area, and after verification that it meets stability conditions, it replaces the current configuration. The parameter fine-tuning employs Bayesian optimization, with a search space of ±30% of the current parameter value, and the optimization objective is to maximize the average effect score of the most recent 50 decisions. The performance monitoring queue records the effect scores of the most recent 100 decisions, and automatically triggers the parameter fine-tuning process when the average score of 20 consecutive decisions falls below a preset baseline.

[0084] The adaptive optimization layer also employs the PPO algorithm from reinforcement learning for global optimization. The state space includes the processing scores of each unit, scheduling efficiency indicators, cross-organizational benchmarking results data, and the matching success rate of the historical case library. The action space includes adjusting the evaluation weights and judgment thresholds of the value anchoring unit, adjusting the path evaluation rules of the direction calibration unit, adjusting the rule base configuration of the boundary constraint unit, and adjusting the method recommendation weights of the method matching unit. The reward function is the product of the value purity score and the actual effect score. A policy update is triggered after each decision execution. The Adam optimizer is used with a learning rate of 0.0003, a discount factor of 0.99, a GAE parameter of 0.95, 2048 update steps per update, a batch size of 64, and 10 training rounds. The PPO algorithm policy network structure is Actor 256-128-action space dimension, Critic 256-128-1. The iteration stopping condition is that the reward improvement is less than 0.5% in 100 consecutive updates.

[0085] Example 4 Five-layer complete architecture This embodiment adds a collaborative symbiotic layer to the four-layer architecture of embodiment 3, forming a complete five-layer architecture. It is suitable for complex and comprehensive decision-making scenarios such as major political affairs assessment, fiscal and tax macro-control, urban comprehensive emergency command, high-end talent training planning, military command and dispatch, and national defense strategic assessment.

[0086] When the collaborative symbiotic layer performs long-term trend extrapolation, it receives historical decision sequences, business indicator time series, policy event codes, and cross-entity feedback data. It converts this data into time-series features and causal relationship features, and generates trend prediction results for the future target period based on these features. These trend prediction results serve as the basis for risk assessment during the comprehensive verification of the initial decision-making scheme. Specifically, the long-term time-series extrapolation employs a Transformer-based time-series prediction model. Input features include historical decision sequences, socio-economic indicator time series, and policy change event codes; the output is the predicted value and confidence interval for the future period. The causal relationship features are modeled using a graph neural network, where nodes represent decision elements and edges represent causal relationship types.

[0087] During the human-machine bidirectional calibration, the collaborative symbiosis layer sends the system-generated early warning information, freeze reasons, recommended execution paths, correction suggestions, and risk levels to the manual confirmation end. After the manual confirmation end returns the correction confirmation data, the collaborative symbiosis layer writes the correction confirmation data into the feedback data record and sends it to the adaptive optimization layer. The adaptive optimization layer updates the parameters, thresholds, rules, or method effectiveness scores of relevant units based on the correction confirmation data. The human-machine bidirectional calibration uses the Shapley value attribution analysis method to calculate the marginal contribution of each input feature to the decision-making effect, and uses the attribution results as the basis for manual correction and system parameter updates.

[0088] When the collaborative symbiotic layer proactively identifies risks, a dual-threshold triggering mechanism is adopted. The risk probability P is calculated using the formula P = 1 - (A × B × C), where A is the compliance score, B is the value consistency score, and C is the output of the risk assessment model. When P is greater than the soft threshold, an early warning is triggered; when P is greater than the hard threshold, the decision-making process is automatically frozen and awaits manual review. The soft threshold is 0.6, and the hard threshold is 0.85. The threshold values ​​are determined based on the statistical distribution of risk probabilities between normal and abnormal decisions in historical decision data.

[0089] Example 5 Evaluation Model This embodiment provides an intelligent evaluation method and system based on a five-layer nested closed loop, applicable to scenarios requiring post-event review and purity assessment of historical decisions, such as disciplinary inspection and supervision, organizational assessment, audit review, and third-party evaluation. This embodiment shares the same underlying five-layer nested closed-loop architecture and four-dimensional review engine as the decision-making models described in Embodiments 1 to 4. The difference lies in that the input data consists of historical decision records that have been executed or are to be executed, and the output results are a purity assessment report, a deviation diagnosis report, and a risk assessment conclusion.

[0090] In evaluation mode, the system receives historical decision data, including decision application text, decision type, decision-maker identifier, decision time, actual execution result, and effect score. This historical decision data is input into the micro-loop layer, where the value anchoring unit scores the motivation purity of the historical decisions, the direction calibration unit evaluates the goal consistency of the historical decisions, the boundary constraint unit performs rule hit backtracking detection of the historical decisions, and the method matching unit evaluates the method suitability of the historical decisions. The large-loop layer summarizes the evaluation results of each unit, generating a comprehensive evaluation report that includes a purity score curve, deviation distribution, rule collision records, and motivational tendency analysis. The evaluation report can serve as a quantitative reference for cadre assessment, organizational review, and audit diagnosis.

[0091] In this embodiment, modules can interact with each other via standardized interfaces. Decision request data is sent to the corresponding module in the form of a structured request body, which includes at least a request identifier, decision text, decision type, decision-maker information, and historical decision information. After processing, the module returns the processing result in a structured response body, which includes at least a request identifier, module name, processing status, scoring result, deviation result, compliance result, rule hit result, and detailed evaluation data. Inter-module communication can be accomplished using remote procedure calls. Data is serialized according to a unified data protocol before transmission, enabling modules to receive and return data with the same field meanings. The standardized interface uses a RESTful API with JSON data format; the remote procedure call uses gRPC, and the service is defined as service ValueAnchor { rpc Evaluate(DecisionRequest) returns(EvaluationResult);}; the unified data protocol uses Protocol Buffers, and the messageDecisionRequest is defined as messageDecisionRequest { string request_id=1; string decision_text=2; string decision_type=3; DecisionMaker decision_maker=4; repeated HistoricalDecisionhistorical_decisions=5;}.

[0092] In this embodiment, the system includes a decision request table, a value anchoring review record table, a direction calibration path record table, a boundary constraint review record table, a method matching record table, a feedback data table, a rule base table, and a case base table. The decision request table stores basic information and current processing status of decision requests, including fields for request identifier, decision text, decision type, decision-maker identifier, affiliated organization identifier, submission time, and processing status. The value anchoring review record table stores direction consistency evaluation results, motivation evaluation results, compliance judgment results, motivation purity score, and final judgment results, including fields for request identifier, review time, direction consistency score, public interest orientation probability, private interest driving probability, number of rule hits, rule hit list, motivation purity score, and comprehensive review result. The direction calibration path record table stores recommended path content, path ranking, various evaluation results, and bottleneck prediction results. The boundary constraint review record table stores path boundary review results, rule hit list, and correction suggestions. The method matching record table stores recommended method combinations and matching time. The feedback data table stores prediction score, actual effect score, deviation value, feedback source, and feedback time. The rule base table stores rule types, triggering conditions, processing actions, and effective status. The case database table stores historical decision-making scenarios, decision characteristics, chosen paths, actual effects, and experience data.

[0093] In this embodiment, when inter-module communication times out or processing fails, the system first retryes a preset number of times. If the retry still fails, the current decision request is marked as a system anomaly, and an alarm is sent to the administrator. When the model or rule engine fails to load, the system switches to a preset backup model or backup rule configuration. If the backup configuration is still unavailable, the system enters a safe processing mode and transfers the current decision request to manual processing. When the decision request data lacks necessary supporting materials, the system marks the data as incomplete in the review report and prompts the user to supplement the materials and resubmit. When the same decision-maker submits multiple decision requests, the system queues the decision requests according to priority and creates an audit record for priority adjustment actions.

[0094] In this embodiment, the system also includes identity authentication, hierarchical access control, operation auditing, and data encryption. After a user logs in, the system determines the types of decision requests they can initiate, the scope of review reports they can view, and the adjustable parameter permissions based on their identity. All submission, modification, approval, rejection, confirmation, and feedback operations are logged, and the logs record at least the operator, operation time, operation content, and operation source. Decision texts, review reports, feedback data, and historical case data are encrypted during transmission and storage to ensure data security during cross-module processing and cross-entity collaboration.

[0095] In this embodiment, the user terminal includes a decision initiation page, a review progress page, a review result page, a decision feedback page, and a historical decision analysis page. The decision initiation page receives the decision title, decision type, decision application text, and supporting materials. The review progress page displays the processing progress of value anchoring, direction calibration, boundary constraints, and method matching, and shows the corresponding review results after each module completes processing. The review result page displays the comprehensive judgment result, module scores, risk points, and recommended paths. The decision feedback page receives descriptions of actual effects, problem records, improvement suggestions, and effect scores. The historical decision analysis page retrieves historical decision records by time range, decision type, review result, and score range, and displays historical score trends, module score distribution, and statistics of common risk points. After receiving manual confirmation, feedback evaluation, or correction opinions, the user terminal writes the corresponding data into the feedback data record and sends it to the adaptive optimization layer for subsequent updates.

[0096] The above provides a detailed description of the intelligent decision-making method and system based on a five-layer nested closed loop proposed in this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. An intelligent decision-making method based on a five-layer nested closed loop, characterized in that, include: Acquire decision request data, and perform element identification and normalization processing on the decision request data to obtain decision element data; The decision element data is input into the micro-circulation layer, and the value anchoring unit, direction calibration unit, boundary constraint unit, and method matching unit in the micro-circulation layer respectively form the unit processing results; wherein, the value anchoring unit generates a value anchoring processing result containing a motivation purity score based on the decision element data; The unit processing results are input into the large loop layer, which summarizes and judges the processing results of each unit according to the preset scheduling relationship, and returns the full process operation data to the corresponding functional unit according to the unit classification, driving each unit to complete the micro-loop iteration and obtain the system processing results; The system processing result is input into the ecological collaboration layer, which performs cross-subject collaborative verification on the system processing result based on the desensitized interactive data to obtain the collaborative processing result. The collaborative processing results are input into the adaptive optimization layer, which performs vulnerability warnings and updates on the processing rules, calibrates the evaluation parameters, and dynamically expands the case knowledge base based on historical operation data and feedback data to obtain the optimized processing results. The optimization results are input into the collaborative symbiosis layer, which then adaptively generates and outputs a decision scheme based on human-machine calibration data, risk identification results, time-series simulation results, and scenario adaptation parameters.

2. The method according to claim 1, characterized in that, The decision request data includes decision application text, decision type, supporting material data, and historical decision data. The element identification and normalization process extracts decision object, decision objective, constraints, and historical correlation information based on the decision application text and supporting material data, and forms the decision element data.

3. The method according to claim 1, characterized in that, The value anchoring unit generates value anchoring processing results based on the matching relationship between the decision element data and preset policy rule data, organizational target data, and historical decision data. These results include directional consistency evaluation results, motivation evaluation results, rule hit results, and motivation purity scores. The motivation evaluation results are generated based on the probability distribution of public interest orientation and private interest drive in the decision text. The rule hit results are generated based on collision detection between the decision elements and the preset boundary rule base.

4. The method according to claim 1, characterized in that, The direction calibration unit iteratively generates multiple candidate execution paths within the compliance framework based on the decision element data, resource constraint data, time constraint data, and historical case data. It then performs a comprehensive evaluation of each candidate execution path based on policy fit, resource consumption, time adaptability, risk level, and historical success rate, generating a direction calibration processing result.

5. The method according to claim 1, characterized in that, The boundary constraint unit decomposes the recommended execution path into a sequence of operation nodes based on the direction calibration processing result, and generates a boundary constraint processing result based on the matching relationship between the operation node data and the preset boundary rules; wherein, the preset boundary rules include mission red line rules, discipline bottom line rules, and ethical high-voltage line rules; the boundary constraint processing result includes the rule hit status, risk level, and correction suggestions of each operation node.

6. The method according to claim 1, characterized in that, The method matching unit extracts decision features based on the decision element data, and based on the mapping relationship between the decision features and the methods and tools in historical successful cases, combined with the boundary constraint processing results, excludes methods and tools that do not meet the constraints, and generates method matching processing results; wherein, the method matching processing results include recommended combinations of methods and tools and usage guidelines for each tool.

7. The method according to claim 1, characterized in that, The large loop layer summarizes the processing results of the units according to the processing order of value anchoring unit, direction calibration unit, boundary constraint unit and method matching unit, and generates a continue processing result, warning result or blocking result according to the processing status of each unit's processing result; when any unit outputs a blocking result, subsequent unit calls are stopped; the large loop layer also returns the entire process operation data to the corresponding functional unit by unit classification, and performs attribution analysis and deviation analysis on each unit based on the returned data.

8. The method according to claim 1, characterized in that, The ecological collaboration layer receives de-identified interactive data corresponding to different subjects, and performs benchmarking processing on rule data, case data, or method tool data based on the de-identified interactive data to generate the collaborative processing result.

9. The method according to claim 1, characterized in that, The adaptive optimization layer determines the deviation between the unit processing result and the actual feedback result based on the historical operation data and feedback data, and updates the processing rules, evaluation parameters or case data according to the deviation; wherein, the adaptive optimization layer also generates rule redundancy prompts or rule vulnerability warnings based on the rule triggering frequency and changes in decision effect; the adaptive optimization layer also automatically triggers parameter fine-tuning within a preset parameter range based on the effect scores of multiple consecutive decisions.

10. An intelligent decision-making system based on a five-layer nested closed loop, characterized in that, include: The data acquisition and processing module is used to acquire decision request data and perform element identification and normalization processing on the decision request data to obtain decision element data. The micro-circulation layer processing module is used to input the decision element data into the micro-circulation layer, and the value anchoring unit, direction calibration unit, boundary constraint unit and method matching unit in the micro-circulation layer respectively form the unit processing results; wherein, the value anchoring unit generates a value anchoring processing result including the motivation purity score based on the decision element data; The large loop layer processing module is used to input the unit processing results into the large loop layer, which summarizes and judges the processing results of each unit according to a preset scheduling relationship, and returns the full process operation data to the corresponding functional unit by unit classification, driving each unit to complete the micro-loop iteration and obtain the system processing result; The ecological collaboration layer processing module is used to input the system processing results into the ecological collaboration layer, and the ecological collaboration layer performs cross-subject collaborative verification on the system processing results based on the desensitized interactive data to obtain the collaborative processing results. The adaptive optimization layer processing module is used to input the collaborative processing results into the adaptive optimization layer. The adaptive optimization layer performs vulnerability warning and updates on the processing rules, calibrates the evaluation parameters, and dynamically expands the case knowledge base based on historical operation data and feedback data to obtain the optimized processing results. The collaborative symbiotic layer processing module is used to input the optimization processing results into the collaborative symbiotic layer, which then adaptively generates and outputs a decision scheme based on human-machine calibration data, risk identification results, time-series simulation results, and scenario adaptation parameters.

Images