Decision deviation degree real-time identification method, device, equipment, medium and system based on interactive negative feedback

Abstract

The invention provides a decision deviation degree real-time identification method, device and system based on interactive negative feedback, equipment and a medium. Obtaining a target decision sequence, comparing the corresponding real-time decision feature flow with a preset reference decision criterion, and identifying decision deviation points and associated deviation attribute features; obtaining an attribution option set, pushing the attribution option set to an external interaction entity, and receiving a feedback attribution selection signal; task context features are obtained, dimension mapping processing is carried out on the attribution selection signals and the task context features, feature deviation types are recognized, and feature deviation nodes in the feature probability graph model are updated; and calling the updated feature probability graph model to extract a probability state value to execute distribution quantization, and outputting a quantization deviation feature vector. According to the method, by introducing interactive feedback and probability graph modeling, real-time and objective recognition of the decision deviation degree and quantitative measurement of the deviation intensity are achieved, and the technical bottleneck that recognition is lagged and a deep motivation cannot be quantitatively recognized in a traditional scheme is effectively solved.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and cognitive computing technology, and more specifically, to a method, apparatus, device, medium, and system for real-time identification of decision deviation based on interactive negative feedback. Background Technology

[0002] With the rapid development of fintech and personalized education systems, intelligent decision support systems are being increasingly widely used in areas such as personal investment and financial management, and professional skills training. These systems need to analyze users' decision-making behavior to identify their inherent cognitive patterns and potential biases, thereby providing personalized guidance and suggestions.

[0003] Existing cognitive analytics solutions typically employ statistical analysis methods based on passive behavioral trajectories such as clickstreams and dwell time. This approach first collects user operation logs on the interactive interface to extract simple behavioral metrics; then, it uses a pre-defined classification algorithm to categorize the metrics and define users' static preference tags; finally, it matches fixed prompts based on these tags.

[0004] However, this passive analysis approach has significant technical limitations. Because mere behavioral trajectories cannot reveal the underlying motivations behind decisions, a definite causal relationship is lacking between the identified results and the actual cognitive biases, making it difficult to quantitatively model the intensity of the bias. Furthermore, the lack of a real-time feedback correction mechanism prevents dynamic adjustment of the cognitive model based on the current task context, resulting in delayed and less objective identification of user decision deviations, failing to meet the need for accurate correction in complex decision-making scenarios. Summary of the Invention

[0005] To address the aforementioned technical problems, this application provides a method, apparatus, device, medium, and system for real-time identification of decision deviation based on interactive negative feedback, thereby at least alleviating the aforementioned technical problems.

[0006] A real-time decision deviation identification method based on interactive negative feedback includes: Step 1, acquiring a target decision sequence, comparing its corresponding real-time decision feature stream with a preset reference decision criterion to identify decision deviation points and associated deviation attribute features in the target decision sequence; Step 2, acquiring a set of attribution options associated with the decision deviation points, triggering an interactive interface to push the attribution option set to an external interactive entity, and receiving attribution selection signals fed back by the external interactive entity; Step 3, acquiring task context features associated with the decision deviation points, performing dimensional mapping processing on the attribution selection signals and the task context features to identify the feature deviation type for the decision deviation points, and updating the feature deviation nodes in a preset feature probability graph model using the feature deviation type; Step 4, retrieving the updated feature probability graph model, extracting the probability state values ​​of each feature deviation node, and performing distribution quantization processing to produce a quantized deviation feature vector for the target decision sequence.

[0007] Optionally, the reference decision criteria are obtained by modeling the execution characteristics of expert decision logic within the task domain, which includes a standard state sequence representing the ideal decision path and associated logical trigger thresholds.

[0008] Optionally, the task context features include external environmental noise components when executing the target decision sequence, task urgency parameters, and historical decision deviation records corresponding to the target decision sequence.

[0009] Optionally, updating the feature deviation nodes in the preset feature probability graph model using the feature deviation type specifically includes: retrieving the prior probability distribution of each feature deviation node in the feature probability graph model; using the feature deviation type as observation evidence, performing posterior probability mapping on the prior probability distribution through a Bayesian update algorithm to generate the updated deviation intensity value of each feature deviation node as the probability state value.

[0010] Optionally, before step 2, the method further includes: calculating the magnitude of the numerical deviation of the real-time decision feature stream relative to the reference decision criterion; and, in response to the magnitude of the numerical deviation exceeding a preset interaction trigger threshold, executing the action of triggering the interaction interface to push the attribution option set to an external interaction entity.

[0011] Optionally, the attribution option set includes multiple attribution label components preset based on the cognitive feature library; pushing the attribution option set specifically includes: matching a target number of attribution label components from the cognitive feature library based on the deviation attribute features associated with the decision deviation point; performing structured encapsulation on the matched attribution label components and sending them to the external interactive entity via the interaction interface.

[0012] Optionally, after generating the quantized deviation feature vector for the target decision sequence, the method further includes: injecting the quantized deviation feature vector into a preset decision correction operator to drive the target decision sequence to perform state transition in a direction that conforms to the reference decision criterion.

[0013] A real-time decision deviation identification device based on interactive negative feedback, comprising: The deviation identification module is used to acquire the target decision sequence, compare its corresponding real-time decision feature stream with the preset reference decision criteria, so as to identify the decision deviation points in the target decision sequence and the deviation attribute features associated therewith. The interactive feedback module is used to obtain the set of attribution options associated with the decision deviation point, trigger the interactive interface to push the set of attribution options to the external interactive entity, and receive the feedback attribution selection signal. The model update module is used to obtain the task context features associated with the decision deviation point, perform dimensional mapping between the attribution selection signal and the task context features to identify the feature deviation type for the decision deviation point, and update the feature deviation node in the preset feature probability graph model using the feature deviation type. The quantization identification module is used to retrieve the updated feature probability graph model, extract the probability state values ​​of each feature deviation node, and perform distribution quantization processing to produce a quantized deviation feature vector for the target decision sequence.

[0014] An electronic device includes a memory and a processor; the memory is used to store a computer program; the processor is used to execute the program to implement the steps of the method as described in any one of the embodiments of this application.

[0015] A computer-readable storage medium having a computer program stored thereon, characterized in that: when the program is executed by a processor, it implements the steps of the method as described in any one of the embodiments of this application.

[0016] A real-time decision deviation identification system based on interactive negative feedback, comprising: The decision monitoring terminal is used to collect real-time decision feature streams generated by the target decision sequence. The cloud-based interaction engine is used to compare the real-time decision feature stream with the reference decision criteria to identify decision deviation points and their associated deviation attribute features, and to push a set of attribution options to the associated external interaction entities. The cognitive modeling center is used to receive the attribution selection signal fed back by the external interactive entity, and update the feature probabilistic graphical model in combination with task context features to produce a quantified deviation feature vector. The feature probability graph model is composed of multiple topological nodes representing the feature deviation nodes, and each topological node performs a state update based on the feature deviation type.

[0017] This application's real-time decision deviation identification scheme based on interactive negative feedback addresses the technical shortcomings of traditional passive behavior analysis schemes, such as strong subjectivity in identification results and inability to reveal deep-seated motivations for decision-making. It identifies decision deviation points by comparing real-time decision feature streams with reference decision criteria and introduces attribution selection signals based on an interactive interface, thus resolving the problem of uncertain causal relationships between identification results and cognitive motivations. Compared to traditional schemes that rely solely on passive trajectory statistics, this application actively obtains attribution feedback from external entities through interaction and performs dimensional mapping in conjunction with task context features. This ensures that the determination of decision deviation is based on clear behavioral attributions, improving the objectivity and rationality of decision deviation point identification.

[0018] This invention updates the feature deviation nodes in the probabilistic graphical model based on attribution selection signals and generates quantized deviation feature vectors, overcoming the technical shortcomings of traditional schemes that are static and unable to measure the intensity of deviations. By extracting the probability state values ​​of deviation nodes and performing distributed quantization processing, this application can dynamically capture the changing trends of the decision-making subject's cognitive state. The generated quantized deviation feature vectors provide a high-precision quantitative basis for subsequent decision correction. Compared with the static labels of traditional schemes, this approach is more sensitive and accurate in perceiving the degree of decision deviation in complex and ever-changing task environments.

[0019] Finally, by injecting the quantized deviation feature vector into the correction operator to drive state transition, a closed-loop feedback mechanism of "identification-attribution-modeling-correction" is formed. In high-risk decision-making scenarios (such as financial credit or medical auxiliary diagnosis), this scheme can specifically guide the decision sequence back to standard criteria, not only optimizing the system's response time, but also improving the credibility and technical robustness of the artificial intelligence system in complex interaction processes through a structured logical proof path. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a real-time decision deviation identification method based on interactive negative feedback, according to an embodiment of this application. Figure 2 This is a block diagram of a real-time decision deviation identification device based on interactive negative feedback, according to an embodiment of this application. Figure 3 This is a schematic diagram of an electronic device structure according to an embodiment of this application; Figure 4 This is a schematic diagram of a real-time decision deviation identification system based on interactive negative feedback, as an embodiment of this application. Detailed Implementation

[0021] like Figure 1 The illustration shows a real-time decision deviation identification method based on interactive negative feedback according to an embodiment of this application, comprising: Step 1, acquiring a target decision sequence, comparing its corresponding real-time decision feature stream with a preset reference decision criterion to identify decision deviation points in the target decision sequence and their associated deviation attribute features; Step 2, acquiring a set of attribution options associated with the decision deviation points, triggering an interactive interface to push the attribution option set to an external interactive entity, and receiving attribution selection signals fed back by the external interactive entity; Step 3, acquiring task context features associated with the decision deviation points, performing dimensional mapping processing on the attribution selection signals and the task context features to identify the feature deviation type for the decision deviation points, and updating the feature deviation nodes in a preset feature probability graph model using the feature deviation type; Step 4, retrieving the updated feature probability graph model, extracting the probability state values ​​of each feature deviation node and performing distribution quantization processing to produce a quantized deviation feature vector for the target decision sequence.

[0022] Optionally, the reference decision criteria are obtained by modeling the execution characteristics of expert decision logic within the task domain, which includes a standard state sequence representing the ideal decision path and associated logical trigger thresholds.

[0023] Preferably, the feature modeling implementation process for the reference decision criteria is as follows: For the high-dimensional expert decision dataset in the relevant task domain, a semantic parsing operator is invoked to perform causal attribute extraction. This action processes the historical decision trajectories of experts in complex system control, multi-source data fusion, and risk response strategies. By performing discretization annotation of semantic units, the continuous operation flow is deconstructed into discriminative feature vectors with atomic logical meaning. The processing result obtained from the previous step is defined as an expert logical feature cluster. Its significance lies in providing deterministic and quantifiable raw inputs for subsequent construction of judgment benchmarks, solving the technical deficiency of traditional standards that are difficult to perform objective comparisons at the physical layer due to their strong subjectivity.

[0024] Preferably, in the specific technical implementation of generating the standard state sequence representing the ideal decision-making path in the step, expert logical feature clusters are injected into the temporal state mapping module. This module establishes directed acyclic relationships between decision nodes by performing topological sorting on key logical nodes throughout the decision-making lifecycle. Specifically, in complex data processing scenarios, the initial feature capture state is mapped to the sequence start position, the multidimensional logical verification state is mapped to the intermediate jump position, and finally points to the final control signal issuance state. By calculating the transition probability distribution between each logical node, an execution trajectory with dynamic temporal attributes is generated. The processing result obtained from the above processing actions is defined as a standard state sequence, the purpose of which is to establish the optimal execution path under physical environment constraints, providing deterministic spatiotemporal reference coordinates for real-time monitoring.

[0025] Preferably, the expert logic feature cluster is a set of high-dimensional feature vector matrices that integrates discrete decision primitives extracted from the expert's decision history logs. Each feature vector component contains physical variable dimensions representing the system state, logical decision operators, and associated causal influence weights. The existence of this feature cluster provides a structured semantic input carrier for subsequent state mapping, solving the technical bottleneck of the unstructured nature of the original decision records and their difficulty in directly participating in temporal evolution calculations. By materializing the expert's thought logic into computable feature vectors, a physical path for transferring from empirical knowledge to digital logic is established.

[0026] Preferably, during the process of injecting expert logic feature clusters into the temporal state mapping module, a graph construction operator within the logic-driven module is executed to map each decision primitive in the feature cluster to a vertex in a directed graph. The operator automatically identifies the logical predecessor and successor relationships between vertices based on the causal influence weights contained in the primitives, thereby establishing directed edges in memory representing the flow of decision logic. This processing action deals with a high-dimensional feature vector matrix, and the processing result is defined as a decision logic topology graph. The construction of this topology graph lays the physical structural foundation for subsequent topological sorting to eliminate logical loops and establish a deterministic execution sequence.

[0027] Preferably, in the specific technical implementation of topological sorting for the decision logic topology graph, the temporal state mapping module calls a sorting algorithm to linearize the logical nodes in the graph. The technical essence of this action lies in using the in-degree and out-degree relationships between nodes to filter out all legal execution sequences that do not have cyclic dependencies. The sorting process ensures that each logical decision point has a definite preceding input and subsequent output in the time domain. The processing result obtained from the above actions is defined as an acyclic logical sequence. This sequence eliminates logical conflicts and infinite loops in the decision-making process, and physically ensures that the subsequently generated state sequence has strict causal unidirectionality.

[0028] Preferably, based on the acyclic logical sequence, a driving dimension mapping operator assigns roles to each logical node. The specific technical processing actions are as follows: logical nodes with zero in-degree are identified and mapped as the sequence start position, responsible for capturing and loading initial features; nodes with non-zero in-degree and non-zero out-degree are mapped as intermediate jump positions, performing multi-dimensional logical verification and intermediate variable passing; and nodes with zero out-degree are mapped as the final control signal issuing position. This action processes the sorted logical nodes, and the processing result is defined as a hierarchical, role-based node set. Through this role-based mapping, a physical transformation from an abstract logical chain to a concrete execution timing path is achieved.

[0029] Preferably, the core technology for generating standard state sequences lies in the fact that the temporal state mapping module uses Markov chain logic to calculate the transition probability distribution of adjacent nodes in the hierarchical role-based node set. Specifically, the technical process involves assigning a probability weight representing the determinism of state transitions to each directed edge based on the frequency and success rate of node transitions in historical expert data. The processing result obtained from the above actions is defined as a dynamic temporal execution trajectory. This trajectory is not merely a static node connection, but a statistically significant behavioral evolution model that reflects the physical possibility of the ideal decision-making behavior evolving to the next stage under different input characteristics.

[0030] Preferably, the final output standard state sequence physically represents a highly deterministic behavioral reference trajectory. The technical processing essentially involves optimizing the execution path of the dynamic temporal execution trajectory to lock onto the execution path that minimizes the cumulative deviation across all dimensions under physical constraints. The result of this processing is the standard state sequence. As a deterministic spatiotemporal reference coordinate in the real-time monitoring process, it enables the execution program to quickly identify behavioral deviations by calculating the geometric distance between the real-time operational feature flow and this reference coordinate. This process represents a technological leap from experience-based guidance to automated logic auditing, providing a stable zero-point benchmark for subsequent quantitative modeling.

[0031] Preferably, the aforementioned causal deconstruction and temporal mapping logic for expert logical feature clusters realizes the physical evolution from static features to dynamic trajectories. By transforming the expert's discrimination primitives into standard state sequences with topological structure and probabilistic attributes, this application solves the technical defects of traditional judgment criteria lacking logical depth and temporal robustness. Each nominal technical feature is closely linked to the underlying execution actions of topological sorting and transition probability calculation, ensuring the logical closed loop and technical stability of the overall solution at the physical implementation level. This in-depth refinement based on the instruction execution perspective ensures that the standard state sequence can serve as a high-confidence benchmark level, supporting subsequent accurate quantification and corrective intervention of decision deviation.

[0032] Preferably, in the specific technical implementation of the associated logical trigger threshold, the tolerance model for each node in the standard state sequence is performed using a deviation variance quantization operator. The execution logic retrieves the behavioral dispersion index of the expert group within the relevant domain when executing instructions, and calculates the upper and lower limits of the confidence interval representing the reasonable range of operational fluctuations. The confidence interval obtained in the previous step serves as the physical benchmark for determining the threshold, and is converted into a numerical threshold that can be automatically recognized by the program logic through a nonlinear proportional transformation. The processing result obtained from the above actions is defined as the logical trigger threshold. The necessity of this threshold lies in the fact that it provides an adaptive dynamic checkpoint for determining whether real-time decision-making behavior has deviated substantially, ensuring that the recognition action can be compatible with reasonable physical environment fluctuations.

[0033] Preferably, the deviation variance quantization operator is a statistical boundary calculation module whose core function is to realize the numerical mapping from expert behavior samples to physical execution thresholds. In terms of technical processing, this operator first reads the execution trajectory components representing the logical step size from the expert logic feature cluster, and then calls the built-in discreteness analysis function to perform preliminary extraction of the probability distribution features of the sample set. The processing result obtained in the previous step is defined as behavioral discrete benchmark data, which provides physically consistent raw input for subsequent quantification of reasonable fluctuation boundaries, solving the problem of logical rigidity caused by the lack of sample support in traditional schemes.

[0034] Preferably, in the specific technical implementation of quantization processing for discrete benchmark data of behavior, the deviation variance quantization operator drives the internal second-order moment calculation unit to calculate the expected value and variance of the execution parameters (such as instruction trigger delay or state variable deviation) corresponding to each logical node. By performing statistical modeling on the execution consistency of the expert group under similar physical conditions, a central component representing the central tendency of behavior and a deviation distribution component representing the random fluctuation of behavior are produced. The processing result obtained from the above processing is defined as the characteristic fluctuation moment matrix, the significance of which lies in transforming the complex expert behavior attributes into numerical tensors that can be processed by algebraic logic, thus establishing a data carrier for subsequent execution error decomposition.

[0035] Preferably, the technical essence of the bias variance decomposition of the characteristic fluctuation moment matrix by the bias variance quantization operator is to use a preset orthogonal decomposition algorithm to separate the characteristic fluctuation moment matrix into a bias weight term representing the systematic judgment tendency and a variance noise term representing random environmental disturbances. This action decouples the inherent biases in the expert logic from the random fluctuations caused by environmental noise, producing behaviorally robust boundary features with denoising properties. The processing result obtained from the previous processing action is defined as the robust boundary component set, which physically ensures that the subsequently generated threshold can both retain the core judgment logic of the expert and avoid over-responding to non-essential minor disturbances.

[0036] Preferably, in the technical implementation of constructing the upper and lower limits of the confidence interval using a robust boundary component set, the logic calls the distribution function fitting operator to perform envelope calculation on the probability density distribution of each logic node in the multidimensional state space. By introducing a preset reliability probability factor (e.g., set to a coverage probability of 95% or 99%), the operator determines the corresponding confidence boundary points on the probability density distribution curve. The processing result obtained from the above actions is defined as the physical range of behavioral tolerance. This range defines the compliance fluctuation range of the expert logic at each decision node at the physical level, solving the technical deficiency of traditional fixed thresholds that cannot balance the flexibility and rigor of logic.

[0037] Preferably, in the specific technical implementation of performing a nonlinear proportional transformation on the physical interval of behavioral tolerance to produce the logic trigger threshold, the execution unit drives the nonlinear mapping module to perform follow-up compression or expansion of the width of the confidence interval based on the current task urgency parameter and external environmental noise components. The technical processing action is to map the physical interval of behavioral tolerance to the interrupt trigger bit field of the execution unit, and calculate the final numerical comparison quantity in combination with a preset sensitivity gain coefficient. The processing result obtained in the previous step is defined as the logic trigger threshold. The technical essence of this transformation process lies in materializing the statistically significant distribution boundary into a physical switch in real-time control logic, realizing the leap from macroscopic behavioral patterns to microscopic judgment signals.

[0038] Preferably, in one scenario, the deviation variance quantization operator achieves adaptive adjustment of the judgment threshold under complex working condition monitoring tasks. When external environmental noise components indicate high-frequency disturbances in the current signal link, the operator automatically increases the weight component of the variance noise term, thereby physically shifting the generated logic trigger threshold outward and expanding the tolerance range for real-time behavioral features. The technical effect of this design is that it not only establishes a dynamic barrier with adaptive capabilities, but also ensures that the recognition action can be compatible with reasonable physical environment fluctuations through deep utilization of the variance structure within the expert logic. This detailed refinement based on the instruction execution perspective eliminates logical breakpoints in technical implementation and provides a physical trigger benchmark for subsequently producing high-confidence quantized deviation feature vectors.

[0039] Preferably, in the process of structurally integrating the standard state sequence and logical trigger thresholds to generate reference decision criteria, matrix encoding logic is used to align the dimensions of the two. Technically, a judgment benchmark matrix is ​​constructed, with the time step size as the row dimension and the logical trigger thresholds of each dimension as the column dimension. The processing unit performs coordinate mapping to load environmental constraints such as current system resource load, link bandwidth, and latency into the judgment benchmark matrix in real time, thereby retrieving the judgment boundaries adapted to the current physical conditions. The processing result obtained from the above actions is the generated reference decision criterion. Its significance lies in establishing a dynamic comparison framework based on high-dimensional expert prior information, physically ensuring that the solution has an objective judgment origin in execution tasks of varying complexity.

[0040] Preferably, in one scenario, the reference decision criteria, when applied to the verification of complex control commands, provide refined judgment logic for identifying abnormal jump behaviors. When the generated real-time decision feature stream is at a critical logic jump point, the processing logic retrieves the corresponding logic trigger threshold from the reference decision criteria in real time and compares the instantaneous amplitude of the current signal with the robustness weight of historical experts at that node. If the real-time feature deviates from the evolution path defined by the standard state sequence and the amplitude exceeds the threshold constraint, it is immediately identified as a decision deviation point and associated deviation attribute features are generated. This criterion generation method based on expert behavior modeling can profoundly reveal the nonlinear dependency between system state feedback and human operation commands, providing highly reliable evidence for subsequent probabilistic modeling.

[0041] Preferably, the modeling process of expert decision-making logic in the reference decision-making criteria also involves the parameterized mapping of expert feedback correction characteristics. The designed reflective weight calculation operator is used to quantitatively extract the remedial trajectory of the expert after a judgment error, and this trajectory is injected into the standard state sequence as a secondary correction factor. By reshaping the verification mechanism in the expert logic, an enhanced logic node with self-calibration characteristics is produced. The processing result obtained from the above actions is defined as a multidimensional enhanced decision-making criterion. Its value lies not only in providing positive execution guidance, but also in physically simulating the prudent evaluation process of experts facing uncertain signals, greatly expanding the robustness of the solution in extreme abnormal data environments.

[0042] Preferably, the specific process of quantifying and extracting the remedial trajectory of an expert after a judgment error by the reflective weight calculation operator is as follows: First, the abnormal state recovery sequence from the expert's historical decision flow is extracted from non-volatile memory. This sequence contains complete time-series samples from the triggering of the logical judgment error signal to the return of the execution state to the robust interval. The execution logic identifies a series of remedial actions performed by the expert at the physical execution level and their movement trajectory in the feature space by comparing the changes in feature gradients before and after the anomaly. The processing results obtained from the above processing are defined as a set of corrective logic primitives, which provide micro-level data support for the subsequent quantification of the reflective characteristics in the expert's thinking.

[0043] Preferably, the trajectory dynamics analysis unit, driven by the internal design of the reflexive weight calculation operator, performs spatiotemporal cost analysis on the set of corrective logic primitives produced by the preceding steps. In terms of technical processing, the operator calculates the response time constant when the expert corrects decision biases and the convergence rate of the corrective action in the feature space. The technical essence of this action lies in measuring the strength and robustness of the expert logic's correction when dealing with uncertain signals. The processing result obtained from the above actions is defined as the reflexive response feature vector. This vector materializes the self-correcting dynamic characteristics of expert thinking, solving the technical bottleneck that static models cannot simulate the "trial and error correction" process in human thinking.

[0044] Preferably, in the technical implementation process of transforming the reflective response feature vector into a second-order correction factor, the execution unit invokes the designed sensitivity mapping operator. This operator maps complex dynamic features into scalar weights that can participate in the correction of the state transition matrix by performing projection in a preset reflective intensity coordinate system. In the processing action, this weight is coupled with the stability variance of the original standard decision path for calculation. The processing result obtained from the above processing action is the second-order correction factor. The existence of this factor enables the execution program not only to obtain positive execution guidance, but also to perceive under what logical conditions the decision path needs to be supplemented with additional verification levels.

[0045] Preferably, in the specific implementation of injecting the secondary correction factor into the standard state sequence, the execution logic retrieves the target logic node in the standard sequence and performs register-level parameter reloading. The secondary correction factor is injected into the logic threshold discriminant of the node, and the original trigger probability is adjusted with a secondary bias. This action processes the originally fixed ideal execution path. The processing result obtained from the above actions is defined as an enhanced logic node. At the underlying physical structure, this node evolves from the original unidirectional trigger logic to a self-interpretive bidirectional verification logic, giving it the physical attributes of self-state evaluation and dynamic threshold correction during instruction execution.

[0046] Preferably, the multidimensional enhanced decision criterion is generated by performing the aforementioned feature reshaping actions on each key node in the standard state sequence. The topology optimization operator is driven to perform weight fusion and logical alignment between the self-calibrating path composed of enhanced logical nodes and the original standard state sequence. The essence of this technical processing lies in using a reflection factor to strengthen the logical layer of the ideal path execution. The processing result obtained from the above actions is the final generated multidimensional enhanced decision criterion. Its value lies in constructing a physical-level "prudent evaluation logic," enabling the execution program to perform deep logical auditing through the verification mechanism of enhanced nodes when facing extreme abnormal data or low signal-to-noise ratio environments.

[0047] Preferably, in one scenario, the application of the reflexive weight calculation operator achieves a physical-level improvement in decision robustness. When the feature signal input to the controlled object deviates significantly from the expected trajectory, the self-calibration mechanism within the enhanced logic node is triggered. Using the injected secondary correction factor, the node physically enhances the rigor of the logic comparison and dynamically adjusts the logic trigger threshold to filter out non-steady-state noise signals that may lead to misjudgments. The technical essence of this design lies in establishing a set of adaptively corrective logical redundancy at the execution level by simulating the expert's reflexive process. This deeply refined implementation logic eliminates the technical black box in cognitive modeling, ensuring that the generated criteria can serve as a highly reliable physical benchmark, supporting the long-term steady-state operation of complex systems.

[0048] Optionally, the task context features include external environmental noise components when executing the target decision sequence, task urgency parameters, and historical decision deviation records corresponding to the target decision sequence.

[0049] Preferably, the external environmental noise component characterizes the instantaneous dispersion and signal jitter rate of the input data stream in the decision execution chain. In complex logic processing scenarios, this component is obtained by real-time monitoring of the variance fluctuation of the input signal and the frequency of influx of non-correlated information. Its existence is to provide a background interference benchmark for subsequent decision logic, enabling the execution program to identify which deviations are due to occasional jumps caused by the instability of the data transmission channel. The technical processing action performs smoothing and denoising on the high-frequency sampled data to remove interference levels belonging to the environmental level. The existence of this component allows the execution logic to identify and eliminate non-essential decision deviations caused by external random signal interference, thereby focusing more on capturing the inherent deviations of the logic core.

[0050] Preferably, the task urgency parameter is used to measure the execution pressure of the target decision sequence under time-restricted resource conditions, and it is determined by calculating the ratio of the remaining available time period of the task to the preset logical discrimination complexity. The fundamental principle lies in using the coupled analysis of real-time clock counts and task stack depth to produce a quantitative weight characterizing time pressure. The existence of this parameter reveals the physical constraints of processing load on the consistency of decision logic, providing a dimensional reference for determining whether the system state is in an unsteady state due to processing frequency pressure. By introducing this parameter, the decision drift caused by time pressure can be compensated during linear mapping, effectively improving the logical coherence of feature deviation type recognition.

[0051] Preferably, the historical decision deviation record originates from an addressing retrieval of past feature deviation vectors stored in a non-volatile storage area. The technical processing action is a design-driven associative query operator that retrieves a deviation distribution matrix with a time index from the feature storage space based on an identifier. This feature includes the average deviation magnitude of the controlled object after quantification over multiple processing cycles and the activity frequency of deviation nodes. Its value lies in providing a longitudinal timeline reference benchmark for current behavioral analysis, enabling the recognition process to extend from single behavioral samples to long-term pattern recognition. In this way, the temporal continuity and data robustness of cognitive deviation quantification modeling are physically established.

[0052] Preferably, the association query operator is a feature retrieval unit with hardware addressing logic and data reconstruction capabilities. Its core task is to realize the semantic link from instantaneous decision features to long-term historical trajectories at the physical level. The technical processing action first manifests as performing hash parsing or register-level address mapping on the input controlled object identifier to lock the associated storage start bit in the physical space of the non-volatile storage area. The object of this action is the controlled object identifier from the current instruction cycle, and the processing action is to convert it into the corresponding physical storage offset. The processing result obtained from the above actions is defined as a feature addressing pointer, which establishes the physical addressing coordinates for subsequent high-bandwidth data reading.

[0053] Preferably, in the specific technical implementation of driving the correlation query operator to perform data retrieval, the execution unit extracts the deviation distribution matrix from the non-volatile storage area according to the feature addressing pointer. The deviation distribution matrix is ​​physically represented as a multi-dimensional tensor structure, where the row dimension corresponds to multiple processing cycles in history, and the column dimension corresponds to each feature deviation node in the feature probability graph model. The technical processing action involves performing a bit-by-bit read of the storage array, loading the historical deviation components stored in discrete storage cells into the current operation cache. The processing result obtained from this action is defined as the loaded deviation matrix stream, whose value lies in providing a complete underlying data sample set for subsequent time-axis correlation analysis.

[0054] Preferably, in the further processing actions performed by the correlation query operator on the loaded deviation matrix stream, the built-in statistical aggregation function is invoked to perform cross-period calculations on each component of the matrix. Specifically, the operator performs root mean square operation or arithmetic mean calculation on the deviation components corresponding to each row of the matrix to produce the average deviation magnitude representing long-term characteristics; simultaneously, it performs cumulative statistics on the active state bits of deviation nodes in the matrix to produce the activity frequency of deviation nodes representing cognitive inertia. The processing result obtained from the above actions is defined as a long-period deviation statistical feature vector. The technical essence of this vector lies in compressing massive amounts of historical raw data into highly representative feature statistics, eliminating the pressure of redundant historical data on real-time computing performance.

[0055] Preferably, in the technical implementation of establishing a longitudinal time axis reference benchmark using the generated long-period deviation statistical feature vectors, the execution logic performs alignment processing on these statistical features and the quantized features of the current processing cycle within the same feature space. The fundamental principle is to establish a feature envelope that evolves over time in a multi-dimensional semantic coordinate system, enabling the recognition process to extend from single behavioral samples to long-period pattern recognition. The processing actions are manifested as follows: using the extracted activity frequency as the prior bias in the current Bayesian update process, and using the average deviation amplitude as the physical benchmark to measure the severity of the current behavioral deviation. The processing result obtained from the above actions is defined as the temporal alignment reference benchmark, which physically establishes the temporal continuity of cognitive bias quantification modeling.

[0056] Preferably, the association query operator strengthens the robustness of real-time decision deviation judgment by outputting the long-period deviation statistical feature vector to the feature probability graphical model. In terms of technical processing logic, confidence correction is performed on the currently captured feature deviation type using historical activity frequencies. If the currently identified deviation has an extremely high activity frequency in the historical record, the operator physically increases the probability update weight of that node. This weight compensation based on historical patterns effectively reduces the probability of misjudgment caused by instantaneous environmental noise components. The processing result obtained from the above actions is defined as the historical weighted decision deviation model. This process ensures that the model output can truly reflect the long-period logical characteristics of the controlled object, rather than short-term fluctuation noise.

[0057] Preferably, in one scenario, the application of the association query operator enables "pattern memory" at the physical execution level. When a controlled object exhibits non-steady-state operations in a complex control task, the operator instantly extracts the deviation evolution trend of the object over thousands of processing cycles through fast addressing of non-volatile storage regions. By injecting the current real-time decision feature stream into a manifold structure constructed from historical feature deviation vectors, it is possible to identify whether the deviation belongs to an inertial pattern with fixed causes. The technical essence of this design lies in introducing a deep temporal dimension constraint to the real-time identification logic through the physical association of historical data, improving the generation accuracy of quantified deviation feature vectors, and providing high-reliability historical data proof for subsequent accurate triggering of correction operators.

[0058] Preferably, in the specific technical implementation of acquiring the task context features, the execution unit drives multiple feature scanning threads to synchronously capture the environmental state bits, time window parameters, and historical behavior feature bit fields of the task to be processed. The real-time acquired signal fluctuation amplitude, time ratio parameters, and historical deviation data obtained from memory addressing are subjected to time-series alignment processing and loaded into the feature normalization register group. Technical processing actions include performing scaling transformations on environmental interference values ​​and historical deviation vectors of different dimensions, mapping them to a unified context vector space. The processing result obtained from the above actions is defined as a multi-dimensional task background feature vector. This vector serves as the core input object for subsequent dimensional mapping processing, ensuring that the recognition process can fully integrate the current instantaneous physical environment constraints and long-term logical evolution features.

[0059] Preferably, the essence of performing dimensional mapping processing between the attribution selection signal and the task context features lies in using objective environmental constraints to perform logical consistency verification on the input signal. The technical process is as follows: the attribution selection signal characterizing the feedback motivation is used as the principal axis feature, and the multi-dimensional task background feature vector is used as the constraint component, injected into the designed nonlinear discriminant operator. If the environmental noise component indicates a high interference level, the discriminant operator will automatically reduce the weight of the rational component in the attribution signal and redirect it to the background-driven feature deviation type. The technical essence of this design is to eliminate data distortion that may be caused by a single signal source by introducing external physical constraints, achieving a physical-level leap from capturing superficial features to tracing the essential motivation.

[0060] Preferably, when using the task context features to assist in updating the feature deviation nodes in the feature probabilistic graphical model, the prior evolution rate of the probabilistic network is dynamically adjusted according to the intensity of the current task background. The technical processing action involves dynamically controlling the step size of weight adjustments for associated nodes in the network based on the level of the task urgency parameter. If historical records indicate that the controlled object has long-term logical deviations, and the current environmental noise is low, the activity level of the feature deviation node during the probability update process will be physically increased. This processing action ensures that the evolution trajectory of the probabilistic graphical model is highly aligned with the true logical characteristics of the controlled object, eliminating model shifts caused by missing context, and providing robust physical logic support for producing quantified deviation feature vectors.

[0061] Optionally, updating the feature deviation nodes in the preset feature probability graph model using the feature deviation type specifically includes: retrieving the prior probability distribution of each feature deviation node in the feature probability graph model; using the feature deviation type as observation evidence, performing posterior probability mapping on the prior probability distribution through a Bayesian update algorithm to generate the updated deviation intensity value of each feature deviation node as the probability state value.

[0062] Preferably, in the specific implementation of updating the feature deviation nodes in the preset feature probability graph model using the feature deviation type, the execution unit drives the instruction addressing logic to retrieve the data structure of the target feature probability graph model from the persistent bit field of non-volatile memory. The feature probability graph model, in terms of physical properties, consists of a set of topological matrices representing causal relationships and associated node state attributes. Its existence is to provide a non-linear modeling framework for the evolution of decision deviations. The technical processing action loads the node connection weights and topological parameters stored in the discrete address space into the current computational cache by performing memory mapping. The processing result obtained from the above processing action is defined as the activated cognitive feature topological model, which establishes the physical-level logical topological foundation for subsequent probabilistic deduction.

[0063] Preferably, in the technical processing of retrieving the prior probability distribution of each feature deviation node in the feature probability graph model, the execution logic traverses each node storage unit in the activated cognitive feature topology model. The feature deviation node corresponds to a latent variable bit in the model in terms of data structure, and is used to carry weight values ​​representing the probability of deviation occurring. The technical processing action is to extract the probability distribution parameter set at the start of the current calculation cycle from the associated state register based on the node index value. The processing result obtained in the previous step is the prior probability distribution, which reflects the initial state level accumulated based on history, enabling the model update action to achieve incremental evolution on the existing quantization foundation, rather than disordered overlay.

[0064] Preferably, in the specific technical implementation of using the feature deviation type as observational evidence, the logic module performs signal vectorization transformation on the feature deviation type identified in the preceding steps. In terms of processing, by retrieving a preset deviation type mapping table, the discrete classification identifier is transcoded into observation component vectors with physical dimensions. This processing object is the identification conclusion belonging to external input; by performing semantic transcription of the bitstream, it is transformed into a physical excitation source that can participate in tensor operations. The processing result obtained from the above processing is defined as the observational evidence. The existence of this evidence allows the instantaneous negative feedback signal to physically act on the long-term modeling framework, establishing a mapping path between external observation signals and the internal state of the model.

[0065] Preferably, the underlying principle of performing posterior probability mapping on the prior probability distribution using the Bayesian update algorithm drives a pre-defined probability inference operator to perform a recursive distribution reshaping action. Specifically, the technical processing involves using the observed evidence as a constraint term for the conditional probability, combining it with the node coupling weights stored in the prior probability distribution, and performing an integral accumulation operation and normalization of the joint probability distribution. This action, at the physical execution level, manifests as a dynamic offset correction of the node state vector, and the resulting processing is defined as the posterior distribution mapping flow. This mapping flow reveals the physical drift of the cognitive bias state in each dimension, given that the current decision deviates from the facts.

[0066] Preferably, in the technical processing of generating updated deviation intensity values ​​for each of the aforementioned feature deviation nodes, the execution unit performs amplitude quantization and feature merging on the posterior distribution mapping stream. In terms of technical processing actions, the expected value of the probability distribution curve is numerically extracted and proportionally calculated with a preset deviation scale benchmark to determine the quantized activation intensity of each node under the current task slice. The processing result obtained from the above actions is the deviation intensity value, which serves as the latest probability state value of the feature deviation node. The significance of this component lies in providing a physically deterministic scalar measure for abstract logical deviations, solving the technical bottleneck that cognitive state variables are difficult to directly invoke by the underlying control logic.

[0067] Preferably, after generating the deviation intensity value, a logical attribute write-back instruction is executed to synchronize the generated probability state value to the memory write buffer of the corresponding feature deviation node. Using physical address offset logic, atomic-level write updates are performed on the node attribute bits in the activated cognitive feature topology model. This action achieves a physical migration from instantaneous calculation results to persistent model states, and the resulting processing result is defined as the evolved quantized deviation graph model. The technical essence of this design lies in materializing the mathematical output of probability deduction into state changes of storage units, physically enabling the cognitive model to adaptively evolve with the behavioral fluctuations of the controlled object.

[0068] Preferably, in one scenario, the aforementioned model update process driven by observational evidence achieves physical filtering of decision noise. When the identified feature deviation type is an occasional, low-probability jump, the robustness of the prior distribution will perform damping compensation on the posterior mapping through the weight adjustment mechanism in the Bayesian update logic, thereby keeping the generated deviation intensity value within a stable range and reducing the false alarm rate. Conversely, if similar deviation evidence appears consecutively, the excitation level of the associated node will be physically increased. This closed-loop update logic eliminates the defect of parameter fixation in traditional schemes, enabling real-time alignment of the true deviation pattern of the controlled object at the physical layer, providing solid logical support for the subsequent generation of high-confidence quantified deviation feature vectors.

[0069] Optionally, before step 2, the method further includes: calculating the magnitude of the numerical deviation of the real-time decision feature stream relative to the reference decision criterion; and, in response to the magnitude of the numerical deviation exceeding a preset interaction trigger threshold, executing the action of triggering the interaction interface to push the attribution option set to an external interaction entity.

[0070] Preferably, before executing the interactive negative feedback loop, a real-time decision feature stream is first acquired. This feature stream, at the data level, is a high-dimensional state vector captured in real-time by the underlying perception module within the current instruction cycle. It includes the timing offset component triggered by the instruction, the amplitude deviation component, and the logical path branch identifier. The corresponding reference decision criterion is an expert logic benchmark feature cluster pre-stored in a non-volatile memory register. The existence of this criterion provides a zero-point reference with deterministic physical meaning for subsequent deviation detection. In terms of technical processing, a logic-driven feature alignment unit is executed to match and align the real-time decision feature stream with the reference decision criterion in terms of spatiotemporal dimensions, thereby extracting the deviation from the original signal to be solved and establishing a stable comparison benchmark for subsequent quantitative analysis.

[0071] Preferably, in the technical essence of calculating the magnitude of the numerical deviation, the execution unit calls the built-in Euclidean distance operator or Manhattan distance calculation module to perform summation and square root processing on the aligned feature components. This action processes the residual vector of the real-time decision feature stream relative to the reference decision criterion. By performing a dimensionality reduction mapping from a high-dimensional vector to a low-dimensional scalar, the complex, multi-dimensional decision deviation is materialized into a level magnitude that can be directly determined by logic gates, thus producing the magnitude of the numerical deviation. This feature solves the recognition ambiguity problem caused by multi-dimensional attribute conflicts, enabling a unified quantitative measurement of the severity of decision deviations at the physical level, ensuring the objectivity of the judgment results.

[0072] Preferably, in the specific technical implementation for calculating the magnitude of the numerical deviation, the execution unit first obtains the dimensional projection of the real-time decision feature stream and the reference decision criterion at the same timestamp. The technical processing action is a subtractive clustering unit that drives the design, performing element-wise difference operations on the corresponding attribute components in the two sets of high-dimensional tensors. This action processes each dimension feature of the high-dimensional state vector (e.g., displacement deviation components, torque fluctuation components, etc. in industrial control commands), and the processing result is defined as a residual vector. This vector physically represents the original offset of the controlled object relative to the zero-position reference in multi-dimensional space, providing a complete differential data source for subsequent distance metric calculation.

[0073] Preferably, the technical essence of the solution process, which involves calling the built-in Euclidean distance operator, drives the internal square accumulation circuit of the operator. Specifically, the technical process is as follows: element-wise squaring is performed on each dimension component of the residual vector to produce the deviation energy components for each dimension; subsequently, parallel summation across all dimensions is performed on the deviation energy components using an addition tree structure (e.g., accumulating the square values ​​of each dimension), and finally, the square root unit is called to perform a square root operation on the accumulated sum. The result obtained in the previous step is the numerical deviation magnitude. The technical essence of this operator lies in establishing a linear geometric mapping of multidimensional deviations in space through norm calculation, thus solving the problem of dimensional inconsistency in heterogeneous properties.

[0074] Preferably, in the technical essence of calling the Manhattan distance calculation module to execute the alternative solution, the absolute value accumulation unit within the logic-driven module is executed. The technical processing action is as follows: extract the absolute value components of each dimension component in the residual vector and perform a linear accumulation operation based on the grid topology. The object of this action is also the multidimensional deviation difference, but its physical meaning lies in calculating the cumulative deviation steps of the feature flow on the logic step axis. The processing result obtained from the above actions is also mapped to the numerical deviation amplitude. The significance of this module is to provide a norm metric with high robustness for operating conditions with large nonlinear jitter (such as state transitions in high-noise signal interference environments), ensuring the sensitivity of the decision logic to extreme value deviations.

[0075] Preferably, the essence of the technique of performing dimensionality reduction mapping from a high-dimensional vector to a low-dimensional scalar using the above operators lies in achieving semantic-level weight fusion. The execution logic uses a preset weight matrix to perform gain compensation on different dimensions of the residual vector, so that the contribution weight of key decision attributes (such as system safety boundary constraints) to the final magnitude is at a high level (e.g., setting the weight ratio to 70%). This mapping process physically compresses complex, potentially conflicting, multi-dimensional decision features into a single numerical component. The processing result obtained by the above process achieves full normalization of deviation information, solves the recognition ambiguity caused by multi-dimensional attribute conflicts, and enables complex system deviations to be materialized into quantitative indicators that can be directly processed by subsequent logic gates.

[0076] Preferably, the generated numerical deviation amplitude is represented at the physical execution level as a set of level signals or binary numerical codes characterizing the deviation intensity, which are directly injected into the subsequent comparator circuit. Since the amplitude after dimensionality reduction possesses a monotonically increasing physical characteristic, it can achieve rapid interception of complex decision-making risks by setting a single interactive trigger threshold. The technical essence of this process lies in transforming behavioral pattern features that cannot be directly compared into physical quantities that can be atomically determined by logic gates (e.g., performing logical comparisons of values ​​greater than or equal to). The processing result obtained in the previous step not only eliminates the identification noise of multidimensional attributes but also ensures the objectivity and uniqueness of the judgment result at the physical level, providing a solid physical triggering basis for subsequent accurate triggering of attribution feedback.

[0077] Preferably, the preset interaction trigger threshold serves as a logical decision threshold with adaptive adjustment characteristics. Its value is dynamically loaded to the reference terminal of the comparison circuit by the load evaluator based on the availability of current processing resources and the fault tolerance window depth of the execution task. This threshold is necessary to perform preliminary screening of massive execution behaviors at the physical level, filtering out non-essential minor deviations caused by signal link jitter, random sensor drift, or internal quantization noise. By establishing this logical barrier, it is physically ensured that subsequent high-overhead interactive feedback loops are activated only when the decision deviation exceeds the set confidence interval boundary, effectively avoiding invalid redundancy in the instruction stream and improving overall recognition efficiency.

[0078] Preferably, in response to the determination action that the numerical deviation amplitude exceeds a preset interactive trigger threshold, the numerical comparison operator is driven to perform a differential comparison of the level. When the signal level of the detected numerical deviation amplitude is higher than the reference level of the preset interactive trigger threshold, the comparison operator generates a logic high-level trigger pulse. This pulse, as the sole enable signal for activating the interactive scheduling module, is sent to the interrupt input of the interactive scheduling controller. The technical essence of this action is to realize the physical jump from silent background acquisition to active front-end intervention in the decision monitoring path using logic threshold comparison. The generated trigger pulse signal not only characterizes the objective occurrence of severe deviation behavior, but also physically locks the error logic coordinates within the current processing cycle.

[0079] Preferably, in the specific technical implementation of executing the trigger interaction interface to push the attribution option set to the external interaction entity, after receiving a high-level trigger pulse, the interaction scheduling controller immediately starts the memory indirect addressing logic to retrieve a set of predefined attribution description identifiers associated with the current deviation attribute feature from the attribution feature library. In terms of technical processing, the designed standardized protocol encapsulation operator is invoked to concatenate the discrete description identifiers, the context feature bits of the current task, and the feedback control code into bit fields, generating an interaction control message with self-interpretive characteristics. The processing result obtained from this action is defined as the attribution payload message, which physically establishes a unified communication format for subsequent negative feedback signal acquisition, ensuring data consistency between heterogeneous interfaces.

[0080] Preferably, in the technical implementation of pushing the attribution option set to the external interactive entity, the physical layer transmission interface module maps the encapsulated attribution payload message to the data transmission buffer according to the selected communication channel bandwidth. This action processes the structured message generated in the preceding steps, and the processing action involves driving the physical layer protocol stack to modulate the message into an electromagnetic signal or bit stream suitable for remote transmission, and then distributing it to the receiving port of the associated external interactive terminal. The processing result obtained in the previous step is defined as the reached interactive command stream, which is explicitly rendered in the display buffer of the external interactive entity. This action achieves a complete closed loop from the logical level of "identification bias" to the physical level of "guided feedback," providing an indispensable physical triggering basis for subsequently updating the feature probability graphical model by capturing subjective attribution selection signals.

[0081] Optionally, the attribution option set includes multiple attribution label components preset based on the cognitive feature library; pushing the attribution option set specifically includes: matching a target number of attribution label components from the cognitive feature library based on the deviation attribute features associated with the decision deviation point; performing structured encapsulation on the matched attribution label components and sending them to the external interactive entity via the interaction interface.

[0082] Preferably, in the specific implementation of pushing the attribution option set, the addressing logic is first driven to activate the cognitive feature library from the preset address space of the non-volatile memory. The cognitive feature library, in its physical data structure, is a multi-dimensional correlation matrix composed of offset attribute components and causal logic identifiers. Its significance lies in providing an indexable causal explanation benchmark for underlying execution deviations. In terms of technical processing, the library pre-stores a large number of feature vectors representing instruction execution anomalies (such as instruction timing errors, logical branch jump errors, or state variable out-of-bounds errors), and assigns a corresponding attribution logic code to each feature vector. Through the establishment of this storage structure, the originally isolated execution errors are physically mapped and locked to the preset causal relationship chain, establishing a data foundation for the subsequent leap from physical deviations to logical recognition.

[0083] Preferably, in the technical implementation of matching based on the deviation attribute features associated with the decision deviation point, the processing unit injects the acquired deviation attribute features into the designed feature comparison operator. The deviation attribute features include the instruction offset magnitude, data jitter frequency, and associated register state mirror captured within the current execution cycle. The technical processing action manifests as: driving the feature comparison operator to perform cosine similarity calculation or Euclidean distance quantization on the deviation attribute features and each benchmark offset feature stored in the cognitive feature library. The technical essence of this action lies in using high-dimensional space similarity discrimination to quickly lock several attribution logic codes with a high degree of correlation with the current execution anomaly logic in the storage bit field, thereby realizing the physical layer retrieval of the causes of execution deviation.

[0084] Preferably, in the specific processing flow of the attribution label component that matches the target number from the cognitive feature library, a dynamic truncation logic based on association weights is introduced. The attribution label component, as the smallest data unit carrying attribution logic information, contains a category field representing the cause type and an association weight coefficient. In terms of technical processing, based on the similarity calculation results from the previous steps, all retrieved candidate labels are sorted in descending order, and the total number of labels allowed to be sent, i.e., the target number, is dynamically determined according to the transmission channel bandwidth of the current instruction stream and the instruction stack depth of the receiving processor. The processing result obtained in the previous step is the filtered ordered label sequence. The value of this design lies in the fact that, through proactive intervention in the data load, redundant data is physically avoided from blocking the real-time communication link, ensuring a high level of response efficiency in the feedback link.

[0085] Preferably, in the structured encapsulation processing of the matched attribution tag components, a standardized message construction operator is invoked. The fundamental principle is to allocate a contiguous address segment in a memory buffer and, according to a preset communication protocol bit field format, concatenate the binary encoding of the attribution tag component, the timestamp index of the current task, and the corresponding logical verification checksum at the bit level. The processing result obtained from the above actions is defined as the attribution payload message to be distributed. Through this action, discrete logical identifiers are transformed into a structured bit stream with self-interpretive characteristics, realizing the physical transformation of logical layer data into transport layer payload, and providing a standardized data encapsulation format for subsequent stable cross-interface transmission.

[0086] Preferably, in the technical implementation of sending data to the external interactive entity via the interaction interface, the physical layer control logic maps the aforementioned generated attribution payload message to the data transmission queue. The interaction interface, at the hardware level, is a set of logic ports driven by the protocol stack, responsible for managing the physical movement of data between different clock domains or address spaces; the external interactive entity corresponds to the downstream execution node or display terminal that receives and parses the message. In terms of technical processing, asynchronous communication pulses are used to push the message from the transmission queue to the physical transmission medium. This action achieves a complete closed loop of the cognitive negative feedback signal on the physical link, ensuring that the generated attribution option set can reach the external receiver with low latency, establishing a physical trigger path for subsequent capture of feedback signals to update the feature probability graphical model.

[0087] Preferably, in one scenario, the aforementioned matching and encapsulation actions for the attribution option set play a crucial role in feature alignment during the auditing of complex control logic. When the monitoring logic identifies an unexpected deviation (i.e., a decision deviation point) in the current execution sequence at a certain logical jump point, it does not directly report an error and halt the system. Instead, based on the deviation attribute features, it matches the attribution label components that reflect the drift of the internal decision logic from the cognitive feature library. In this way, at the physical execution level, "what error occurred" is bound to "what might have caused the error" in real time. This feature similarity-driven attribution matching and structured encapsulation logic not only enhances the objectivity of the identification process but also improves the technical robustness and logical consistency of the solution when dealing with complex decision conflicts through deep utilization of the underlying storage structure.

[0088] Optionally, after generating the quantized deviation feature vector for the target decision sequence, the method further includes: injecting the quantized deviation feature vector into a preset decision correction operator to drive the target decision sequence to perform state transition in a direction that conforms to the reference decision criterion.

[0089] Preferably, after generating the quantized deviation feature vector for the target decision sequence, the execution unit initiates a semantic decoupling action for the quantized deviation feature vector. This feature vector contains high-dimensional numerical components characterizing the deviation probability distribution, and its significance lies in providing parameter inputs with quantized weights for subsequent execution interventions. The technical processing action involves: by parsing the deviation magnitudes of each dimension in the quantized deviation feature vector, converting the abstract probability statistics into offset increments that can be recognized by the underlying operators. This action transforms the original statistical features into a data stream with driving attributes by extracting bits and converting their format within the memory address space. The processing result obtained from the previous action is defined as the feature stream to be injected for correction, which physically establishes the energy source for subsequent correction logic.

[0090] Preferably, the quantized deviation feature vector is injected into a preset decision correction operator. The execution logic drives the internal bus to move the feature stream to be injected into the dedicated instruction register file of the decision correction operator. The decision correction operator is a logic correction module pre-embedded in the control unit, its purpose being to generate a reverse balancing gain based on the known deviation intensity. The technical processing action involves physically mapping the register addresses to load the deviation values ​​of each dimension into the weight factor bit field inside the operator. The technical essence of this action lies in realizing the physical transition of data from the "monitoring space" to the "execution space." The resulting processing result is defined as the correction operator activation instruction, which provides a physical-level logical trigger source for subsequently generating specific correction actions.

[0091] Preferably, within the decision correction operator, the execution unit drives a pre-set intervention strategy mapping algorithm to perform a nonlinear weighted transformation on the correction operator activation command. The essence of this algorithm lies in utilizing a pre-set association mapping table in memory to match corresponding logical correction components for different combinations of deviations. If the deviation intensity value of a certain dimension is in a high range, the operator automatically increases the gain of the corresponding correction factor. The technical processing action involves performing vector rotation and scaling in a multi-dimensional control vector space to calculate a logical bias that can offset the original deviation direction. The processing result obtained from the above actions is defined as the correction intervention vector, which serves as the physical guide quantity driving the target decision sequence back to the standard path, thus solving the technical deficiency of the lack of specificity in correction actions in traditional schemes.

[0092] Preferably, in the specific technical implementation of driving the target decision sequence to perform state transition in a direction conforming to the reference decision criterion, the execution logic injects the generated correction intervention vector into the state transition probability matrix corresponding to the target decision sequence. In terms of technical processing, by modifying the jump weights between logical nodes in the sequence, the subsequent execution path is physically offset. The technical essence of this action is to use externally injected intervention signals to suppress the already deviated logical jump tendency, thereby inducing the decision trajectory to converge within subsequent timestamps. The result produced by the previous processing step is defined as the offset state transition matrix, whose significance lies in forcibly guiding the decision-making process back to a compliant execution track through real-time correction of physical parameters.

[0093] Preferably, the final manifestation of the state transition is the realignment of the execution state in physical space. The technical processing action is as follows: A synchronization monitoring operator is invoked to capture real-time features after executing the offset state transition matrix, and a real-time difference operation is performed between these features and the ideal trajectory defined by the reference decision criterion. If the difference result is within a preset tolerance range, the state transition is determined to have reached the expected level. The processing result obtained from this action is defined as a normalized decision sequence. This sequence maintains a high degree of consistency with the reference decision criterion in terms of logical rigor and temporal coherence. Through this dynamic adjustment of transition weights, logical drift in time-varying environments can be flexibly addressed, ensuring that the produced decision sequence is physically aligned with the preset reference benchmark.

[0094] Preferably, in one scenario, the injection and correction process of the quantized deviation feature vector achieves physical coupling between the execution link and the cognitive modeling conclusion. Without interrupting the current task flow, the execution logic utilizes physical-level register write-back and parameter overwrite techniques to dynamically intervene in the decision evolution trend. This complete and detailed logic, from quantized feature extraction to intervention vector calculation and then to physical state transition, eliminates uncertainties in the actual implementation process. Each noun-based technical feature is described in detail, demonstrating how it participates in the operator's calculation and actions, ensuring the rigor of the overall logic of the solution. This detailed processing based on the instruction execution perspective greatly enhances the adaptive correction capability and technical robustness of the hybrid intelligent system when handling extreme deviation tasks.

[0095] like Figure 2 As shown, this is an embodiment of a real-time decision deviation identification device based on interactive negative feedback, which includes: The deviation identification module is used to acquire the target decision sequence, compare its corresponding real-time decision feature stream with the preset reference decision criteria, so as to identify the decision deviation points in the target decision sequence and the deviation attribute features associated therewith. The interactive feedback module is used to obtain the set of attribution options associated with the decision deviation point, trigger the interactive interface to push the set of attribution options to the external interactive entity, and receive the feedback attribution selection signal. The model update module is used to obtain the task context features associated with the decision deviation point, perform dimensional mapping between the attribution selection signal and the task context features to identify the feature deviation type for the decision deviation point, and update the feature deviation node in the preset feature probability graph model using the feature deviation type. The quantization identification module is used to retrieve the updated feature probability graph model, extract the probability state values ​​of each feature deviation node, and perform distribution quantization processing to produce a quantized deviation feature vector for the target decision sequence.

[0096] like Figure 3 As shown, an electronic device according to an embodiment of this application includes a memory and a processor; the memory is used to store a computer program; the processor is used to execute the program to implement the steps of the method as described in any one of the embodiments of this application.

[0097] This application also provides a computer-readable storage medium storing a computer program thereon, characterized in that: when the program is executed by a processor, it implements the steps of the method as described in any one of the embodiments of this application.

[0098] like Figure 4 As shown, this embodiment of the present application discloses a real-time decision deviation identification system based on interactive negative feedback, which includes: The decision monitoring terminal is used to collect real-time decision feature streams generated by the target decision sequence. The cloud-based interaction engine is used to compare the real-time decision feature stream with the reference decision criteria to identify decision deviation points and their associated deviation attribute features, and to push a set of attribution options to the associated external interaction entities. The cognitive modeling center is used to receive the attribution selection signal fed back by the external interactive entity, and update the feature probabilistic graphical model in combination with task context features to produce a quantified deviation feature vector. The feature probability graph model is composed of multiple topological nodes representing the feature deviation nodes, and each topological node performs a state update based on the feature deviation type.

[0099] Preferably, for Figure 4The hardware implementation of the system comprises a physical layer architecture consisting of distributed heterogeneous hardware units working together via a high-speed system bus and network transmission links. The decision monitoring terminal, at the hardware level, is an embedded data acquisition terminal integrating a high-speed signal sampler and a local data preprocessor. In terms of processing, the high-speed signal sampler captures the execution status signals of the controlled device through a physical sensing interface, and the local data preprocessor performs analog-to-digital conversion and feature dimensionality reduction. The processing result obtained from the previous processing is defined as the real-time decision feature stream. This hardware terminal, through a configured physical layer communication module, transmits the real-time decision feature stream to the cloud hardware layer in real time, establishing the physical entry point for the system to perceive the target execution sequence in real time.

[0100] Preferably, the cloud-based interaction engine is deployed at the hardware level in a high-performance logic processing cluster. Its core hardware includes a multi-core parallel processor, high-speed random access memory (RAM), and a dedicated network distribution acceleration card. The high-speed RAM has a dedicated decision criterion cache bit field for persistently storing the reference decision criteria. In terms of technical processing, the multi-core parallel processor performs a bit-level comparison between the received real-time decision feature stream and the reference decision criteria through an instruction pipeline. When the logic comparison unit determines that the value deviates beyond a threshold level, it triggers the dedicated network distribution acceleration card to retrieve the associated attribution option data packet from the cognitive feature storage array and encapsulates it into an attribution payload message, which is then sent to the associated external display terminal. This hardware cluster achieves a physical-level transition from feature recognition to attribution guidance, establishing a physical trigger source for subsequent acquisition of subjective feedback signals.

[0101] Preferably, the cognitive modeling center is composed of a probabilistic inference accelerator and a high-speed topological storage array at the hardware level. The high-speed topological storage array is used to physically store the feature probabilistic graphical model, which is mapped in the storage bit domain as a causal topological matrix composed of multiple physical storage units representing the feature deviation nodes. In terms of technical processing, the probabilistic inference accelerator receives the attribution selection signal returned from the external communication interface through a dedicated instruction set, and performs Bayesian evolutionary operations in conjunction with the task context features extracted from the environmental variable acquisition bit domain. The aforementioned hardware entity, by performing high-speed direct memory access operations, writes the probability update values ​​of each node generated by the operation back to the high-speed topological storage array, thereby producing the quantized deviation feature vector representing the current deviation state.

[0102] Preferably, the topology nodes in the feature probabilistic graphical model correspond at the hardware level to independent register groups in the high-speed topology storage array, with each register group carrying the probability component values ​​of the corresponding cognitive bias dimension. In terms of technical processing, after the feature deviation type is identified and injected into the cognitive modeling center, the probabilistic inference accelerator drives the address encoder to lock the physical address of the target register group, performing atomic-level numerical weighting correction. This state update method based on hardware addressing logic ensures that each topology node can achieve millisecond-level state transitions upon the injection of external observation signals. The updated register value stream obtained in the previous step constitutes the physical payload of the quantized deviation feature vector, providing directly addressable control parameters for the downstream correction execution unit.

[0103] Preferably, the aforementioned hardware entities interact and support each other functionally through a high-bandwidth, low-latency data transmission bus. The signal sampling frequency provided by the decision monitoring terminal determines the real-time benchmark for the comparison performed by the cloud interaction engine; the attribution payload messages generated by the cloud interaction engine provide the physical prerequisite for the cognitive modeling center to obtain valid evidence signals; and the evolutionary updates performed by the cognitive modeling center on the high-speed topology storage array through the probabilistic inference accelerator provide the entire system with a quantitative judgment basis with dynamic evolutionary characteristics. Each hardware module is tightly coupled through a physical layer handshake protocol and interrupt triggering mechanism, solving the computational lag problem caused by memory walls when traditional software architectures process high-dimensional probabilistic graphical models, and physically realizing closed-loop quantitative modeling of decision deviation.

[0104] Preferably, for Figure 4 The physical implementation of the system solution involves an external interactive entity that corresponds to an interactive display terminal with an independent controller. This terminal establishes a bidirectional encrypted link with the cloud-based interactive engine through a physical layer security protocol stack. It is responsible for parsing the attribution payload message into a visual pixel array and capturing the subject's feedback touch actions to generate the attribution selection signal. Driven synchronously by a physical clock source, each hardware entity forms a geographically distributed execution logic: the monitoring end performs physical data acquisition, the cloud cluster performs logical judgment, the modeling center performs probabilistic evolution, and the interactive terminal performs signal acquisition and presentation. This hardware-level technical coordination and collaborative output ensures that the system can accurately quantify deviation states in physical environments with time-varying characteristics.

[0105] The above Figures 2-4 For an exemplary description, please refer to the above. Figure 1 .

Claims

1. A real-time decision deviation identification method based on interactive negative feedback, characterized in that, Includes the following steps: Step 1: Obtain the target decision sequence, compare its corresponding real-time decision feature stream with the preset reference decision criteria, and identify the decision deviation points and associated deviation attribute features in the target decision sequence. Step 2: Obtain the set of attribution options associated with the decision deviation point, trigger the interaction interface to push the set of attribution options to the external interaction entity, and receive the attribution selection signal fed back by the external interaction entity; Step 3: Obtain the task context features associated with the decision deviation point, perform dimensional mapping processing on the attribution selection signal and the task context features to identify the feature deviation type for the decision deviation point, and use the feature deviation type to update the feature deviation node in the preset feature probability graph model; Step 4: Retrieve the updated feature probability graph model, extract the probability state values ​​of each feature deviation node, and perform distribution quantization processing to produce a quantized deviation feature vector for the target decision sequence.

2. The real-time decision deviation identification method based on interactive negative feedback according to claim 1, characterized in that, The reference decision criteria are obtained by modeling the execution characteristics of expert decision-making logic within the task domain. They include a standard state sequence representing the ideal decision-making path and associated logical trigger thresholds.

3. The real-time decision deviation identification method based on interactive negative feedback according to claim 1, characterized in that, The task context features include the external environmental noise component when executing the target decision sequence, the task urgency parameter, and the historical decision deviation records corresponding to the target decision sequence.

4. The real-time decision deviation identification method based on interactive negative feedback according to claim 1, characterized in that, The step of updating the feature deviation nodes in the preset feature probability graph model using the feature deviation type specifically includes: Retrieve the prior probability distribution of each of the feature deviation nodes in the feature probability graph model; Using the feature deviation type as observational evidence, a posterior probability mapping is performed on the prior probability distribution using a Bayesian update algorithm to generate updated deviation intensity values ​​for each feature deviation node, which serve as the probability state values.

5. The real-time decision deviation identification method based on interactive negative feedback according to claim 1, characterized in that, Before step 2, the following is also included: Calculate the magnitude of the numerical deviation of the real-time decision feature stream relative to the reference decision criterion; In response to the numerical deviation exceeding a preset interaction trigger threshold, the action of triggering the interaction interface to push the attribution option set to the external interaction entity is executed.

6. The real-time decision deviation identification method based on interactive negative feedback according to claim 1, characterized in that, The attribution option set includes multiple attribution label components based on a cognitive feature library; The set of attribution options pushed specifically includes: Based on the deviation attribute features associated with the decision deviation points, the attribution label components of the target number are matched from the cognitive feature library; The matched attribution label components are structurally encapsulated and sent to the external interaction entity via the interaction interface.

7. A real-time decision deviation identification device based on interactive negative feedback, characterized in that, include: The deviation identification module is used to acquire the target decision sequence, compare its corresponding real-time decision feature stream with the preset reference decision criteria, so as to identify the decision deviation points in the target decision sequence and the deviation attribute features associated therewith. The interactive feedback module is used to obtain the set of attribution options associated with the decision deviation point, trigger the interactive interface to push the set of attribution options to the external interactive entity, and receive the feedback attribution selection signal. The model update module is used to obtain the task context features associated with the decision deviation point, perform dimensional mapping between the attribution selection signal and the task context features to identify the feature deviation type for the decision deviation point, and update the feature deviation node in the preset feature probability graph model using the feature deviation type. The quantization identification module is used to retrieve the updated feature probability graph model, extract the probability state values ​​of each feature deviation node, and perform distribution quantization processing to produce a quantized deviation feature vector for the target decision sequence.

8. An electronic device, characterized in that, It includes a memory and a processor; the memory is used to store a computer program; the processor is used to execute the program to implement the steps of the method as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, it implements the steps of the method as described in any one of claims 1-6.

10. A real-time decision deviation identification system based on interactive negative feedback, characterized in that, include: The decision monitoring terminal is used to collect real-time decision feature streams generated by the target decision sequence. The cloud-based interaction engine is used to compare the real-time decision feature stream with the reference decision criteria to identify decision deviation points and their associated deviation attribute features, and to push a set of attribution options to the associated external interaction entities. The cognitive modeling center is used to receive the attribution selection signal fed back by the external interactive entity, and update the feature probabilistic graphical model in combination with task context features to produce a quantified deviation feature vector. The feature probability graph model is composed of multiple topological nodes representing the feature deviation nodes, and each topological node performs a state update based on the feature deviation type.

Images