A deterministic conflict arbitration method based on multi-source score consistency detection and mathematical stable convergence

By constructing an independent evaluation domain and introducing an arbitration method based on consistency detection and mathematical stability convergence, the problems of decision chaos and security risks in multi-source information fusion are solved, achieving a unique and deterministic decision output and security assurance.

CN123092259APending Publication Date: 2026-11-13INNOKY (NINGBO) AUTOMATION CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611217342.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-12
Publication Date
2026-11-13

AI Technical Summary

Technical Problem

Existing multi-source information fusion schemes lack deterministic intervention and convergence mechanisms when faced with conflicting opinions among evaluation sources, leading to chaotic decision-making, an inability to output a unique and convincing conclusion, and a lack of hard-line blocking mechanisms for safety red lines, thus posing security risks.

Method used

Three independent evaluation domains (discrete feature and rule matching domain, topological association domain, and semantic probability domain) are constructed. A consistency detection and mathematically stable convergence arbitration method is introduced. The arbitration gateway is switched by consistency confidence and preset threshold. Combined with a veto listener, a decision-making process is ensured that the safety red line is not touched.

Benefits of technology

It achieves a unique and definite decision output in any conflict situation, ensuring the clarity and security of the physical meaning of the decision and eliminating the output of unsafe results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123092259A_ABST
    Figure CN123092259A_ABST
Patent Text Reader

Abstract

This invention discloses a deterministic conflict arbitration method based on multi-source scoring consistency detection and mathematically stable convergence. Addressing the shortcomings of existing multi-source evaluation fusion schemes, such as lack of conflict perception, divergent and uncertain decision results, and susceptibility to weighted penetration of safety red lines, this invention constructs three independent evaluation domains: discrete rule matching, topological association evaluation, and semantic probability matching. After basic quality pre-screening, the consistency confidence is calculated based on the overlap of high-scoring candidate sets in each evaluation domain, and a threshold-driven bi-state arbitration gateway is used. Simultaneously, a safety veto listener with the highest execution priority is configured, which can bypass all downstream computations and directly execute a veto. When a conflict is detected, the system automatically adapts to two convergence modes: targets that hit the mandatory selection rule are injected with a large-scale bias term and forcibly placed at the top; targets without rule constraints are selected through variance penalty to find the most stable solution across domains.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to a deterministic conflict arbitration method based on multi-source scoring consistency detection and mathematically stable convergence. Background Technology

[0002] In complex intelligent decision-making systems, to improve the comprehensiveness and robustness of decisions, multiple independent information sources or algorithmic models are typically introduced for joint evaluation. Common techniques include fusing the results of deep learning-based semantic understanding models, symbolic logic-based rule verification engines, and graph-based topological association analysis algorithms, aiming to achieve better decision-making performance than a single model.

[0003] However, existing multi-source information fusion schemes mainly suffer from the following significant technical defects:

[0004] First, traditional fusion algorithms lack the ability to perceive and resolve conflicts of opinion among evaluation sources.

[0005] Current mainstream solutions mostly employ fixed-weight linear weighting or simple voting mechanisms. When there is a significant discrepancy in the scores given by different evaluation domains, traditional weighted averaging methods cannot effectively identify this discrepancy, and the resulting composite score often becomes a compromise value that has lost its physical meaning. For example, a candidate target might score 95 points in the first evaluation domain (representing a high recommendation) but only 30 points in the second (representing a very low recommendation). A simple weighted average might yield a passing score of approximately 62 points. However, this score neither reflects the high approval of the first evaluation domain nor the strong opposition of the second; in reality, no evaluation source truly agrees with this result. The system thus falls into a predicament, unable to provide a unique and convincing conclusion.

[0006] Second, there is a lack of certain intervention and convergence mechanisms when facing conflicts.

[0007] When multiple evaluation sources offer conflicting opinions and no external prior knowledge is available for reference, existing systems often fall into a state of decision-making chaos. This means the final output can fluctuate randomly due to weight adjustments or different initial states. For example, in some implementations, when two candidate targets have similar scores, the system may randomly change the output order due to floating-point precision issues or the instability of the sorting algorithm. This uncertainty makes it impossible for the system to be stably deployed in industrial environments requiring absolute rigor and reproducibility.

[0008] Third, there is a lack of a rigid mechanism to block the safety red line.

[0009] In existing fusion frameworks, even if a target has crossed the safety red line, such as lacking critical safety certifications, it may still pass muster by receiving an extremely high score in another evaluation domain and being manipulated through a weighted average total score. For example, a device lacking explosion-proof certification might be assigned a high score of 95 in the semantic model simply because its text description contains the word "explosion-proof," and its total score after weighted fusion might still reach over 70, ultimately leading to an incorrect recommendation. Such situations directly result in system outputs that severely violate physical safety regulations, posing a fatal security penetration risk. Summary of the Invention

[0010] This invention provides a deterministic conflict arbitration method based on multi-source scoring consistency detection and mathematically stable convergence to address the problems existing in the prior art.

[0011] The technical solutions adopted in this invention are as follows:

[0012] A deterministic conflict arbitration method based on multi-source score consistency detection and mathematically stable convergence includes the following steps:

[0013] S1: For the set of candidate targets to be processed, each candidate target is scored in three independent evaluation domains that do not share intermediate states, generating corresponding first, second, and third score vectors. The first evaluation domain is a discrete feature and rule matching domain, used to match the discrete physical attributes of the candidate targets with a preset positive condition library and output a discrete rule score. The second evaluation domain is a topological association domain, used to calculate the topological association score based on the association path between candidate target nodes and demand center nodes in the knowledge graph. The third evaluation domain is a semantic probability domain, used to calculate the semantic score based on the semantic matching probability.

[0014] S2: Calculate the basic comprehensive quality score of each candidate target based on the scores of each evaluation domain, and eliminate candidate targets whose basic comprehensive quality score is lower than the preset minimum quality threshold;

[0015] S3: For the candidate targets that were not eliminated in S2, they are sorted in descending order of scores in the first evaluation domain, the second evaluation domain, and the third evaluation domain respectively. The candidate targets ranked at the top in each evaluation domain are extracted to form the corresponding local high-quality candidate sets. The overlap between the local high-quality candidate sets in each evaluation domain is calculated to obtain the consistency confidence. The consistency confidence is then compared with the preset conflict threshold.

[0016] S4: Throughout the entire arbitration process, a veto listener, which is independent of all evaluation domains and arbitration logic and has the highest execution priority, monitors in real time whether the candidate target touches the preset safety red line. If the candidate target touches the preset safety red line, the final composite score of the candidate target is directly assigned to the absolute minimum value, and all downstream fusion and arbitration calculations are bypassed.

[0017] S5: For candidate targets not marked by the rejection listener, the arbitration gateway is driven to switch states based on the comparison result of the consistency confidence score and the preset conflict threshold.

[0018] When the consistency confidence level is greater than or equal to the preset conflict threshold, the arbitration gateway is in a dormant pass-through state and uses the basic comprehensive quality score as the final composite score.

[0019] When the consistency confidence level is less than the preset conflict threshold, the arbitration gateway is activated and queries the rule mapping set:

[0020] If a candidate target hits a mandatory optimization rule in the rule mapping set, a preset large-scale bias term is injected into the candidate target, and the sum of the basic comprehensive quality score and the large-scale bias term is used as the final synthetic score.

[0021] If the candidate target does not match any rule in the rule mapping set, the variance of the candidate target's scores in multiple evaluation domains is calculated as the degree of divergence. A stable convergence function containing a variance penalty term is constructed, and the function value of the stable convergence function is used as the final composite score.

[0022] S6: Sort the final composite scores of all candidate targets in descending order and output a unique and deterministic sequence of decision results.

[0023] Furthermore, the first evaluation domain performs the following operations:

[0024] Extract the discrete physical attributes of candidate targets, match the discrete physical attributes with a preset positive condition library item by item, and accumulate the base score according to the degree of matching; if the candidate target contains preset cross-boundary attributes or irrelevant attribute features, trigger the orthogonal penalty term to deduct points, and output the discrete rule score.

[0025] Furthermore, the second evaluation domain calculates the topological association score as follows:

[0026] A knowledge graph is pre-constructed, where nodes represent entities and edges represent relationships between entities. Starting from the demand center node, the shortest path to the candidate target node is searched along the associated edges. For each node on the shortest path, the number of hops from the demand center node to that node is calculated. The attenuation weight is obtained by exponentiation with a preset attenuation coefficient as the base and the number of hops as the exponent. The attenuation weight is multiplied by the inherent weight of the corresponding node and then summed to obtain the topological association score. The higher the number of hops, the more exponentially the weight contribution of the corresponding node decreases.

[0027] Furthermore, the third evaluation domain performs the following operations:

[0028] The user-inputted requirement statement and the text description of the candidate target are combined into a text pair, which are then input into a large language model to calculate the semantic matching probability. The semantic matching probability is then mapped to a preset scoring range to obtain the semantic score.

[0029] Furthermore, the local high-quality candidate set is composed of the top K candidate targets in each evaluation domain, and the consistency confidence is the minimum number of elements in the pairwise intersection of the local high-quality candidate sets in each evaluation domain.

[0030] Furthermore, the basic comprehensive quality score is obtained by multiplying the scores of each evaluation domain by the corresponding configurable weight coefficients and then summing them up, and the sum of each configurable weight coefficient is equal to 1.

[0031] Furthermore, the preset safety red lines include the lack of legal security certification, exceeding the physical limit range, or touching preset business prohibition conditions; the absolute minimum value is negative infinity or a preset constant much smaller than the normal scoring range.

[0032] Furthermore, the mandatory selection rules in the rule mapping set include industry mandatory access rules, statutory certification priority rules, or physical security compliance rules; the numerical scale of the large-scale bias term is greater than the maximum value of the absolute value of the original score of each evaluation domain.

[0033] Furthermore, the stability convergence function is the difference between the product of the basic comprehensive quality score and the penalty coefficient and the degree of divergence; wherein, the degree of divergence is the variance of the candidate target's scores in multiple evaluation domains, and the penalty coefficient is a preset positive value.

[0034] Furthermore, each evaluation domain runs in parallel in an independent memory space or computation space, and no temporary variables or intermediate calculation results are shared between the evaluation domains.

[0035] The present invention has the following beneficial effects:

[0036] (1) This invention constructs at least three independent evaluation domains that do not share intermediate states, including a discrete feature and rule matching domain, a topological association domain, and a semantic probability domain, and introduces a cross-domain consistency detection mechanism. By calculating the degree of overlap between the local high-quality candidate sets in each evaluation domain, the consistency confidence is obtained, realizing an explicit quantitative perception of the degree of disagreement in decision-making opinions, and providing a mathematical basis for subsequent accurate arbitration.

[0037] (2) This invention introduces a threshold-driven arbitration gateway after consistency detection. When the consistency confidence is greater than or equal to the preset conflict threshold, the arbitration gateway is in a dormant pass-through state, directly outputting the basic comprehensive quality score to ensure processing efficiency in conflict-free scenarios. When the consistency confidence is less than the preset conflict threshold, the arbitration gateway is activated and automatically switches to rule arbitration mode or mathematical convergence mode based on whether the candidate target hits the mandatory selection rule in the rule mapping set. In rule arbitration mode, the system injects a preset large-scale bias term into the candidate target, suppressing all disagreements through strong rule-based decision-making. In mathematical convergence mode, the system calculates the variance of the candidate target's scores in multiple evaluation domains as the degree of disagreement, constructs a stability convergence function containing a variance penalty term, and selects the target with the most stable cross-domain performance through disagreement penalty. This ensures that a unique and definite decision result can be output in any conflict situation.

[0038] (3) This invention deploys a veto listener that is independent of all evaluation domains and arbitration logic and has the highest execution priority throughout the arbitration process. Once a candidate target is detected to have touched the preset safety red line, the veto listener directly assigns the final composite score of the candidate target to the absolute minimum value and bypasses all downstream fusion and arbitration calculations. This mechanism fundamentally eliminates the possibility of unsafe results being output, providing a hard safety net for the system. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the overall module architecture and logical flow principle of the system of the present invention.

[0040] Figure 2 This is a flowchart of the core algorithm for multi-source consistency detection and state machine arbitration of the present invention.

[0041] Figure 3 This is a schematic diagram illustrating the construction of three independent evaluation domains and their feature extraction in this invention. Detailed Implementation

[0042] The invention will now be further described with reference to the accompanying drawings.

[0043] This embodiment uses the intelligent selection of explosion-proof gas detectors in an industrial setting as the application scenario. The candidate targets are different models of portable gas detection devices, and the user requirement is an explosion-proof portable gas detector suitable for natural gas leak scenarios. For example... Figure 1 As shown, the overall architecture of the system of the present invention is divided into four core levels from bottom to top, and is equipped with a top-level security watchdog that runs through the entire process, forming a complete closed-loop link from data input to decision output.

[0044] The top-level security watchdog is an independently operating veto listener, independent of all evaluation domains and arbitration logic. It runs continuously in parallel throughout the decision-making process and has the highest execution priority. The veto listener monitors the attribute characteristics of all candidate targets in real time. Preset security red lines include polarity veto conditions such as lack of legal explosion-proof certification, exceeding the physical range limit, and failure to meet industry-mandated access standards. Once any candidate target is detected to have violated a security red line, regardless of whether the arbitration gateway is in a consistent or conflicting state, or whether any weighted summation or bias injection calculations are in progress, all downstream fusion and calculation processes will be directly bypassed. The system will forcibly set the final composite score of that target to an absolute minimum.

[0045] .

[0046] In engineering implementation, a constant value much smaller than the normal scoring range can also be assigned, such as -1e9. This operation forcibly removes the target from the output queue, ensuring it is always at the end of the recommendation list or is removed altogether. This veto power has the highest, insurmountable priority, fundamentally eliminating the possibility of unsafe results being output.

[0047] The first layer is a multi-source independent evaluation layer, containing three parallel and independent evaluation domains: a discrete feature and rule matching domain, a topological association domain, and a semantic probability domain. These three evaluation domains are deployed in independent computation processes and memory spaces, and do not share any temporary variables or intermediate calculation results during operation, ensuring the mathematical independence of the three evaluation outputs. Figure 3 As shown, the three evaluation domains score the candidate targets from different dimensions, generating the corresponding first score vector, second score vector and third score vector.

[0048] The discrete feature and rule matching domain corresponds to the first evaluation domain. Its input consists of the discrete physical attributes of the candidate target, including objective parameters such as equipment type, explosion-proof rating, certification status, detection principle, measurement range, and operating temperature. The system pre-builds a positive condition library, which stores various attribute standards and corresponding scores that match user requirements. During the evaluation process, the system compares the discrete physical attributes of the candidate target with the positive condition library item by item. For each matching positive condition, the corresponding base score is accumulated. If the candidate target contains preset cross-boundary or irrelevant attribute features, an orthogonal penalty term is triggered, deducting the corresponding penalty score from the accumulated base score. Finally, the discrete rule score is output. .

[0049] Taking the natural gas detector selection scenario in this embodiment as an example, the explosion-proof rating of Exia, the detection principle of electrochemistry, and the certification including ISO standards are all positive conditions, and the corresponding scores will be accumulated after matching; if the candidate target is a fixed installation device, it belongs to the cross-border attribute that does not meet the portable requirements, triggering orthogonal penalty and deducting points.

[0050] The topological association domain corresponds to the second evaluation domain, which calculates the association tightness based on a pre-constructed industry knowledge graph. In this knowledge graph, each node represents an entity, including equipment model, detection principle, application scenario, technical standards, etc., and the edges between nodes represent objectively existing associations between entities, such as electrochemical detectors being suitable for natural gas detection and Exia levels meeting explosion-proof requirements. During the evaluation process, the system first maps user needs to demand center nodes in the knowledge graph, and then, starting from the demand center nodes, performs a breadth-first search along the association edges in the graph to calculate the shortest path to the candidate target node.

[0051] For each node on the shortest path Calculate the number of hops from the demand center node to this node. and with a preset attenuation coefficient of The power is used as the decay weight, along with the inherent weight corresponding to that node. After multiplication, the calculation results for all nodes on the path are summed to obtain the topological association score. The calculation formula is:

[0052] ,

[0053] Among them, attenuation coefficient The value is a preset positive value and less than 1; in this embodiment, a value of 0.5 is preferred. (Number of jumps) The larger the value, the more exponentially the contribution weight of the corresponding node decreases. This evaluation domain measures the matching degree between candidate targets and user needs through the structured association relationships of the knowledge graph. The closer the candidate target is to the demand center node and the shorter the association path, the higher the topological association score. Continuing with the industrial gas detector example above, if a detector node is directly connected to the demand center node for natural gas leak detection, i.e., the hop count d=1, it means that the device is clearly defined as a dedicated device for natural gas leak detection in the industry knowledge system, and its topological contribution is higher. If it requires multiple intermediate nodes to establish an association, i.e., the hop count d=3, then the contribution of each hop is halved, and the final contribution is reduced to one-eighth of the original value.

[0054] The semantic probability domain corresponds to the third evaluation domain, which assesses text matching based on the semantic understanding capability of the large language model. The system pre-trains the large language model using a large-scale industry corpus, enabling it to recognize semantic equivalence relations in professional scenarios. During the evaluation process, the system combines the user's input natural language requirement statement with the text description of the candidate target into a text pair, inputs it into the trained large language model, and outputs the semantic matching probability of the two. The semantic matching probability is then mapped to a preset scoring range of 0-100 to obtain a semantic score. ,Right now This evaluation domain can identify semantically equivalent expressions such as portable and handheld, and natural gas leak and combustible gas detection, thus compensating for the insufficient coverage of rule matching and topological association under fuzzy demand expressions.

[0055] The second layer is the basic quality pre-screening layer, which receives the score vectors output from the three evaluation domains, calculates the basic comprehensive quality score for each candidate target, and eliminates low-quality targets. (Basic comprehensive quality score) The calculation uses a linear weighting method, multiplying the score of each evaluation domain by its corresponding configurable weight coefficient and then summing the results. The formula is as follows:

[0056] ,

[0057] in, For the first The weight coefficients corresponding to each evaluation domain satisfy the following conditions: All weighting coefficients can be dynamically configured according to business scenarios. For example, in security-critical scenarios, the weight of the rule matching domain can be increased, while in creative content recommendation scenarios, the weight of the semantic probability domain can be increased. The system presets a minimum quality threshold. If the candidate target's basic comprehensive quality score If a target is deemed low-quality and invalid, it is directly eliminated and will not proceed to the next stage of processing. Only candidate targets with scores higher than or equal to the threshold pass the pre-screening and enter the next stage of conflict detection and arbitration. In this embodiment, the minimum quality threshold is... The preferred setting is 60 points.

[0058] The third layer is the conflict perception and arbitration gateway layer, which is the core processing layer of this invention. It includes two parts: a cross-domain consistency detection module and a threshold-driven arbitration gateway. Figure 2 As shown, the system first performs a consistency check, and then drives the arbitration gateway to switch working modes based on the check results.

[0059] The cross-domain consistency detection module is used to quantitatively assess the degree of disagreement across evaluation domains. For all candidate targets that have passed the pre-screening, the system sorts them in descending order of score across the three evaluation domains, extracting the top K candidate targets in each domain to form the corresponding locally superior candidate set. The set of discrete rule-matching domains is denoted as... The set of topological association domains is The set of semantic probability domains is Then, the number of elements in the intersection of each pair of the three sets is calculated, and the minimum value is taken as the consistency confidence score. The calculation formula is:

[0060] ,

[0061] Consistency confidence The value ranges from 0 to K. A larger value indicates a higher degree of overlap in the generally accepted excellent objectives across evaluation domains, and a smaller decision-making disagreement; a smaller value indicates a greater difference in opinions across evaluation domains, and a higher degree of conflict. A preset conflict threshold is set. This threshold is used to determine whether the system has entered a conflict state. The threshold value can be configured according to the business scenario's tolerance for disagreements. For security-critical scenarios, a higher threshold can be set to increase sensitivity to disagreements, while for low-risk, recommended scenarios, a lower threshold can be set to ensure processing efficiency. In this embodiment, K is set to 5, representing the conflict threshold. The preferred setting is 2.

[0062] The arbitration gateway is driven by a comparison between consistency confidence and a conflict threshold, and globally switches between consistent and conflict states. The arbitration gateway only routes based on the comparison between consistency confidence and a preset conflict threshold when the target has not been marked by a rejection listener.

[0063] When consistency confidence At this point, the system is in a state of consensus, with a high degree of agreement across evaluation domains. The arbitration gateway is in a dormant state (direct-access mode), and the system directly uses a conventional linear weighting mechanism to synthesize the total score, taking the candidate target's basic comprehensive quality score as the final composite score. No additional intervention is required to ensure processing efficiency in conflict-free scenarios.

[0064] When consistency confidence When the current decision-making process is deemed to have entered a multi-source conflict state, with significant disagreements among the opinions in different evaluation domains, the arbitration gateway is activated and queries the pre-defined rule mapping set R for each candidate target. The rule mapping set R stores predefined mandatory business selection rules, including industry mandatory access rules, statutory certification priority rules, and physical security compliance rules. Security red line exclusion conditions are independently monitored and vetoed by the veto listener and are not included in the arbitration query scope of the rule mapping set R. If the physical attributes of candidate target x, such as industry, certification, or range, match one of the mandatory selection rules in the rule mapping set R, then the process enters rule arbitration mode; if no rule is matched, the process automatically downgrades to mathematical convergence mode.

[0065] In the rule-based arbitration model, the system assigns a large-scale bias term to the objective. The final composite score is the sum of the basic comprehensive quality score and the large-scale bias term.

[0066] ,

[0067] This bias term must satisfy:

[0068] ,

[0069] This means that the value of the bias term is much larger than the maximum absolute value of the original scores of each evaluation domain. In engineering implementation, the large-scale bias term can be set to be greater than the sum of the full scores of all evaluation domains, ensuring that the final score of the candidate target that hits the forced selection rule is necessarily higher than that of all candidate targets that do not hit the rule. By suppressing all evaluation discrepancies through rule priority, forced mathematical convergence is achieved.

[0070] In the mathematical convergence model, the system calculates the variance of the candidate target's scores in the three evaluation domains as the degree of divergence. A stable convergence function, including a variance penalty term, is constructed, and its calculated value is used as the final composite score. The formula for calculating the stable convergence function is as follows:

[0071] ,

[0072] in, The penalty coefficient is a preset value, which is positive; in this embodiment, a value of 0.5 is preferred. (Discrepancy degree) The variance of the scores across the three evaluation domains is calculated using the following formula:

[0073] ,

[0074] In the formula , where is the arithmetic mean of the scores of the candidate objective across the three evaluation domains. This convergence mechanism suppresses candidate objectives with large fluctuations in cross-domain scores through a variance penalty term. The greater the cross-domain divergence and the more unstable the scores of the objective, the heavier the penalty is imposed. Finally, the candidate objective with the most stable cross-domain performance is selected as the convergent solution.

[0075] The fourth layer is the deterministic decision output layer. It receives the final composite scores of all candidate targets, sorts them in descending order of scores, and outputs a unique and deterministic sequence of decision results. Since all calculations in the entire decision-making chain are deterministic mathematical operations, state transitions are triggered only by numerical comparisons. There is no random sampling, probability sampling, or branching logic dependent on the initial state. Therefore, the output results under the same input conditions are completely unique and reproducible. The output decision results can be persisted to a database for auditing and traceability, or pushed to the front-end page for display via a streaming interface.

[0076] To more clearly illustrate the forced convergence effect of this invention under conflict conditions, a specific calculation is performed using three candidate targets as an example. Assume that the weight coefficients of the three evaluation domains are equal, i.e. Penalty coefficient Large-scale bias term The maximum score for a single evaluation domain is 100 points.

[0077] The three-way scoring of candidate target A is as follows: , , Its basic comprehensive quality score ; Calculate the score variance The result is obtained by calculating the stability convergence function. The candidate target showed a high degree of consistency in cross-domain scores with extremely low divergence. Even after penalties, the score remained high, ranking first in stability in unregulated scenarios.

[0078] The three-way scoring of candidate target B is as follows: , , Its basic comprehensive quality score ; Calculate the score variance The result is obtained by calculating the stability convergence function. The candidate target only scored highly in a single evaluation domain, but showed significant divergence across domains. After variance penalty, the score dropped sharply, effectively filtering it out and completely eliminating the possibility of a single evaluation domain biasing the final decision.

[0079] The three-way scoring of candidate target C is as follows: , , Its basic comprehensive quality score If the candidate target meets the forced selection rule, such as having legal explosion-proof certification, the final score will be determined after injecting a large-scale bias term. With an absolute advantage, they secured first place, effectively making a strong decision on the rules and suppressing all disagreements.

[0080] When using the traditional weighted average method, the scores of the three candidate targets are 93.7, 61.0, and 87.7, respectively, ranking candidate A first, candidate C second, and candidate B third. This method fails to identify the serious cross-domain conflict of candidate B and also fails to reflect the rule priority of candidate C, resulting in security risks and decision bias in the output. However, the solution of this invention ensures efficiency through direct output when there is no conflict, selects the most stable solution through variance penalty when there is conflict but no rule, and makes a strong decision through bias injection when there is conflict but a rule. Regardless of the scenario, it can output a unique, definite, and physically meaningful decision result, fundamentally solving the triple dilemma of traditional methods where weighted average loses physical meaning in conflict states, false high scores are used to deceive, and rule constraints are diluted.

[0081] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements without departing from the principle of the present invention, and these improvements should also be considered within the scope of protection of the present invention.

Claims

1. A deterministic conflict arbitration method based on multi-source score consistency detection and mathematically stable convergence, characterized in that: Includes the following steps: S1: For the set of candidate targets to be processed, each candidate target is scored in three independent evaluation domains that do not share intermediate states, generating corresponding first, second, and third score vectors. The first evaluation domain is a discrete feature and rule matching domain, used to match the discrete physical attributes of the candidate targets with a preset positive condition library and output a discrete rule score. The second evaluation domain is a topological association domain, used to calculate the topological association score based on the association path between candidate target nodes and demand center nodes in the knowledge graph. The third evaluation domain is a semantic probability domain, used to calculate the semantic score based on the semantic matching probability. S2: Calculate the basic comprehensive quality score of each candidate target based on the scores of each evaluation domain, and eliminate candidate targets whose basic comprehensive quality score is lower than the preset minimum quality threshold; S3: For all candidate targets that have passed the pre-screening, sort them in descending order of score in the first evaluation domain, the second evaluation domain and the third evaluation domain respectively, extract the candidate targets ranked first in each evaluation domain to form the corresponding local high-quality candidate set, calculate the degree of overlap between the local high-quality candidate sets in each evaluation domain, obtain the consistency confidence, and compare the consistency confidence with the preset conflict threshold. S4: Throughout the entire arbitration process, a veto listener, which is independent of all evaluation domains and arbitration logic and has the highest execution priority, monitors in real time whether the candidate target touches the preset safety red line. If the candidate target touches the preset safety red line, the final composite score of the candidate target is directly assigned to the absolute minimum value, and all downstream fusion and arbitration calculations are bypassed. S5: For candidate targets not marked by the rejection listener, the arbitration gateway is driven to switch states based on the comparison result of the consistency confidence score and the preset conflict threshold. When the consistency confidence level is greater than or equal to the preset conflict threshold, the arbitration gateway is in a dormant pass-through state and uses the basic comprehensive quality score as the final composite score. When the consistency confidence level is less than the preset conflict threshold, the arbitration gateway is activated and queries the rule mapping set: If a candidate target hits a mandatory optimization rule in the rule mapping set, a preset large-scale bias term is injected into the candidate target, and the sum of the basic comprehensive quality score and the large-scale bias term is used as the final synthetic score. If the candidate target does not match any rule in the rule mapping set, the variance of the candidate target's scores in multiple evaluation domains is calculated as the degree of divergence. A stable convergence function containing a variance penalty term is constructed, and the function value of the stable convergence function is used as the final composite score. S6: Sort the final composite scores of all candidate targets in descending order and output a unique and deterministic sequence of decision results.

2. The deterministic conflict arbitration method based on multi-source scoring consistency detection and mathematically stable convergence as described in claim 1, characterized in that: The first evaluation domain performs the following operations: Extract the discrete physical attributes of candidate targets, match the discrete physical attributes with a preset positive condition library item by item, and accumulate the base score according to the degree of matching; if the candidate target contains preset cross-boundary attributes or irrelevant attribute features, trigger the orthogonal penalty term to deduct points, and output the discrete rule score.

3. The deterministic conflict arbitration method based on multi-source scoring consistency detection and mathematically stable convergence as described in claim 1, characterized in that: The second evaluation domain calculates the topological association score as follows: A knowledge graph is pre-constructed, where nodes represent entities and edges represent relationships between entities. Starting from the demand center node, the shortest path to the candidate target node is searched along the associated edges. For each node on the shortest path, the number of hops from the demand center node to that node is calculated. The attenuation weight is obtained by exponentiation with a preset attenuation coefficient as the base and the number of hops as the exponent. The attenuation weight is multiplied by the inherent weight of the corresponding node and then summed to obtain the topological association score. The higher the number of hops, the more exponentially the weight contribution of the corresponding node decreases.

4. The deterministic conflict arbitration method based on multi-source scoring consistency detection and mathematically stable convergence as described in claim 1, characterized in that: The third evaluation domain performs the following operations: The user-inputted requirement statement and the text description of the candidate target are combined into a text pair, which are then input into a large language model to calculate the semantic matching probability. The semantic matching probability is then mapped to a preset scoring range to obtain the semantic score.

5. The deterministic conflict arbitration method based on multi-source scoring consistency detection and mathematically stable convergence as described in claim 1, characterized in that: The local high-quality candidate set consists of the top K candidate targets in each evaluation domain, and the consistency confidence is the minimum number of elements in the pairwise intersection of the local high-quality candidate sets in each evaluation domain.

6. The deterministic conflict arbitration method based on multi-source scoring consistency detection and mathematically stable convergence as described in claim 1, characterized in that: The basic comprehensive quality score is obtained by multiplying the scores of each evaluation domain by the corresponding configurable weight coefficients and then summing them up, and the sum of each configurable weight coefficient is equal to 1.

7. The deterministic conflict arbitration method based on multi-source scoring consistency detection and mathematically stable convergence as described in claim 1, characterized in that: The preset safety red lines include the lack of legal security certification, exceeding the physical limit range, or touching preset business prohibition conditions; the absolute minimum value is negative infinity or a preset constant much smaller than the normal scoring range.

8. The deterministic conflict arbitration method based on multi-source scoring consistency detection and mathematically stable convergence as described in claim 1, characterized in that: The mandatory selection rules in the rule mapping set include industry mandatory access rules, statutory certification priority rules, or physical security compliance rules; the numerical scale of the large-scale bias term is greater than the maximum absolute value of the original score of each evaluation domain.

9. The deterministic conflict arbitration method based on multi-source scoring consistency detection and mathematically stable convergence as described in claim 1, characterized in that: The stability convergence function is the difference between the product of the basic comprehensive quality score and the penalty coefficient and the degree of divergence; wherein, the degree of divergence is the variance of the candidate target's scores in multiple evaluation domains, and the penalty coefficient is a preset positive value.

10. The deterministic conflict arbitration method based on multi-source scoring consistency detection and mathematically stable convergence as described in claim 1, characterized in that: Each evaluation domain runs in parallel in an independent memory space or computation space, and no temporary variables or intermediate calculation results are shared between the evaluation domains.