Decision logic encapsulation method, transaction verification method and system for target object

Abstract

The application provides a decision logic packaging method, a transaction verification method and a system for a target object. The method collects multi-dimensional interaction behavior sequence data of the target object in the process of executing a standardized decision task, fine-tunes a decision neural network model based on data distribution to generate a target adaptation parameter set, extracts quantized metadata to configure a standardized application programming interface, and finally packages a decision logic container that complies with a set software package specification. The application converts dynamic decision characteristics into a standardized container that can be independently distributed, thereby alleviating the technical difficulties of complex decision logic being difficult to be objectively quantified, packaged and safely migrated across platforms, achieving efficient packaging and isolated operation of decision-making capabilities, and improving the adaptability and transaction credibility of decision logic in a heterogeneous environment.

Description

Technical Field

[0001] This application relates to the interdisciplinary fields of artificial intelligence, cognitive computing, and software architecture. Specifically, it relates to a method for encapsulating decision logic for a target object, a transaction verification method, and a system. Background Technology

[0002] As the scale of digital asset transactions continues to expand, data, algorithms, and computing power have become core production factors, and the market demand for logical models with defined decision-making characteristics is growing. Digitally extracting and encapsulating the judgment logic and behavioral preferences within a given scenario is a key foundation for achieving personalized collaboration and the flow of decision-making capabilities in intelligent decision-making systems.

[0003] In existing decision logic processing solutions, logic delivery is typically achieved using hard-coded expert systems based on preset rules. This approach involves technical personnel first developing a static decision tree by analyzing business processes; then writing fixed code modules based on the decision tree; and finally compiling and linking to generate an executable file for a specific operating system.

[0004] However, this traditional encapsulation approach has significant technical limitations. Since advanced decision-making logic is often implicit in non-linear sequences of interactive behaviors, hard-coding based on static rules struggles to capture the deep-seated features, weights, and rhythmic characteristics of the decision-making process, resulting in weak fitting capabilities of the encapsulated model in complex scenarios. Furthermore, existing model distribution methods lack environmental isolation mechanisms and standardized performance metrics, leading to severe compatibility issues when migrating decision logic across systems. They also struggle to perform pre-transaction security verification while protecting intellectual property rights, failing to meet the need for standardized circulation of decision logic as a digital asset. Summary of the Invention

[0005] To alleviate the aforementioned technical problems, this application provides a method for encapsulating decision logic for a target object, a transaction verification method, and a system, so as to at least alleviate the aforementioned technical problems.

[0006] A method for encapsulating decision logic for a target object includes: Step 1, collecting multi-dimensional interaction behavior sequence data of the target object during the execution of a standardized decision task; Step 2, fine-tuning a decision neural network model based on the multi-dimensional interaction behavior sequence data to generate a target adaptation parameter set that fits the distribution of the multi-dimensional interaction behavior sequence data; Step 3, extracting quantitative metadata describing the decision performance of the target adaptation parameter set under different logical dimensions, and configuring a standardized application programming interface that conforms to predefined communication specifications; Step 4, encapsulating the target adaptation parameter set, the quantitative metadata, and the standardized application programming interface together into a decision logic container that conforms to a set software package specification to achieve standardized encapsulation of the decision logic of the target object.

[0007] Optionally, the multi-dimensional interactive behavior sequence data includes at least two of the following: selection result data, operation trajectory data, response time data, and decision confidence data under a preset decision-making scenario.

[0008] Optionally, the decision neural network model adopts a Transformer architecture or a long short-term memory network architecture.

[0009] Optionally, the decision logic container conforms to the OpenContainerInitiative specification or the WebAssembly module specification.

[0010] Optionally, the decision logic container includes a metadata layer, an interface layer, and a core model layer; the metadata layer defines the attribute description information of the target adaptation parameter set; the interface layer configures the communication protocol for external systems to access the core model layer; and the core model layer encapsulates the algorithm weights for performing decision calculations.

[0011] Optionally, the metadata layer includes a unique hash identifier generated based on the binary content of the decision logic container, a creator identity identifier processed by asymmetric encryption, and a quantized capability vector representing the decision dimension of the target adaptation parameter set.

[0012] Optionally, the standardized application programming interface includes an initialization interface, a prediction and decision interface, and a logic interpretation interface; the initialization interface is used to load runtime environment configuration information; the prediction and decision interface is used to receive a structured decision scenario description, and based on the target adaptation parameter set, outputs a standardized decision result object containing a list of action options, a preference probability distribution, and a decision confidence level.

[0013] Optionally, the standardized decision result object further includes a summary of key decision factors; the summary of key decision factors is generated by performing a reverse attribution analysis on the feature contribution of the core model layer when performing predictive decision through the logical interpretation interface.

[0014] Optionally, before fine-tuning the decision neural network model, the method further includes: performing temporal feature aggregation processing on the multi-dimensional interactive behavior sequence data to extract the operational rhythm features and feature attention weights of the target object when making decisions; mapping the operational rhythm features and feature attention weights into a structured decision feature matrix; and using the structured decision feature matrix as a fine-tuning constraint term for the decision neural network model.

[0015] Optionally, the fine-tuning of the decision neural network model includes: fixing the shallow network weights representing the extraction of general logic in the decision neural network model, and performing gradient updates on the deep network weights representing the mapping of decision preferences based on the multi-dimensional interaction behavior sequence data, so as to generate the target adaptation parameter set.

[0016] A transaction verification method for a decision logic container as described above, applied to a distributed transaction platform, includes: receiving the encapsulated decision logic container and recording its hash identification code in a blockchain ledger; in response to a buyer's verification request, allocating a trusted hardware execution environment as a sandbox verification environment for the decision logic container; loading the decision logic container into the sandbox verification environment and inputting a test decision scenario to obtain the corresponding standardized decision result object; feeding back the standardized decision result object to the buyer, and preventing the sandbox verification environment from externally reading and writing access to the target adaptation parameter set during the verification process.

[0017] Optionally, it also includes: a smart contract for deploying and executing transaction logic, wherein the smart contract is pre-set with the asset identifier of the decision logic container, the royalty allocation ratio, and the access token required to decrypt the core model layer; in response to the smart contract meeting the execution conditions, the access token is distributed to the buyer to authorize the buyer to obtain full access to the decision logic container.

[0018] A decision logic encapsulation system for a target object includes: a behavior access terminal for collecting multi-dimensional interactive behavior sequence data; a parameter mapping server for generating a target-adaptive parameter set; and a container encapsulation module for generating a decision logic container.

[0019] This application presents a decision logic encapsulation scheme for a target object. Addressing the shortcomings of traditional encapsulation schemes, such as inaccurate capture of behavioral features and weak fitting capabilities, this scheme alleviates the logical distortion caused by reliance on static hard-coded rules by collecting multi-dimensional interaction behavior sequence data of the target object and fine-tuning the neural network model. Compared to traditional schemes, this application generates a target adaptation parameter set in step 2, which can accurately fit the nonlinear distribution in the interaction data, capture deep decision preferences, and enable the generated model parameters to accurately reproduce the judgment characteristics of the target object.

[0020] Based on the target adaptation parameter set generated through fine-tuning, quantitative metadata is extracted and standardized application programming interfaces are configured. Finally, the resulting container is encapsulated into a decision logic container that conforms to the set specifications. This alleviates the technical bottleneck of high environmental coupling and difficulty in cross-platform migration in traditional delivery models. Compared to the traditional method of compiling to generate fixed executable files, this application, through containerization encapsulation technology, not only achieves standardized representation of decision logic but also, through the design of the interface layer and metadata layer, enables the logic assets to have high compatibility and isolation in heterogeneous environments, ensuring the robustness of the decision logic during its flow.

[0021] Finally, by introducing a distributed transaction verification method and utilizing a trusted hardware execution environment and sandbox mechanism, the problem of lack of trust foundation in the transaction of novel logical assets is alleviated. Compared with the traditional crude distribution model of directly copying code, the verification mechanism of this application can provide the buyer with the real decision result object without disclosing the core model parameters. Combined with blockchain and smart contracts, it improves the security of asset delivery and the efficiency of rights confirmation, and provides reliable technical protection for the assetization operation of complex decision logic. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating a method for encapsulating decision logic for a target object according to an embodiment of this application; Figure 2 This is a flowchart illustrating a transaction verification method for a decision logic container according to an embodiment of this application; Figure 3 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application; Figure 4 This is a structural block diagram of a decision logic encapsulation system for a target object according to an embodiment of this application. Detailed Implementation

[0023] like Figure 1The illustration shows a method for encapsulating decision logic for a target object according to an embodiment of this application, comprising: Step 1, collecting multi-dimensional interaction behavior sequence data of the target object during the execution of a standardized decision task; Step 2, fine-tuning a decision neural network model based on the multi-dimensional interaction behavior sequence data to generate a target adaptation parameter set that fits the distribution of the multi-dimensional interaction behavior sequence data; Step 3, extracting quantitative metadata describing the decision performance of the target adaptation parameter set under different logical dimensions, and configuring a standardized application programming interface that conforms to a predefined communication specification; Step 4, encapsulating the target adaptation parameter set, the quantitative metadata, and the standardized application programming interface together into a decision logic container that conforms to a set software package specification, thereby achieving standardized encapsulation of the decision logic of the target object.

[0024] Optionally, the multi-dimensional interactive behavior sequence data includes at least two of the following: selection result data, operation trajectory data, response time data, and decision confidence data under a preset decision-making scenario.

[0025] Preferably, the specific implementation process of step 1 is as follows: In the runtime environment where the target object performs the standardized decision-making task, the original interaction event stream triggered by the target object is monitored in real time. The original interaction event stream is a sequence of atomic operations (e.g., coordinate offset events, key trigger events, and cursor hover events) captured by the execution terminal and bearing high-precision timestamps. The captured original interaction event stream is mapped to a preset decision state space, and the logical transition relationship between each atomic operation is identified using state transition parsing logic to generate a structured behavior sequence object representing the entire decision-making process of the target object. By converting the originally discrete physical operation events into the structured behavior sequence object with logical topological attributes, a logically coherent data source is provided for subsequent extraction of multi-dimensional behavioral features.

[0026] Preferably, in the specific technical implementation of step 1: selection result data and operation trajectory data are extracted from the structured behavior sequence object. Specifically, the termination node of the structured behavior sequence object is parsed to obtain the final decision component submitted by the target object, thereby generating the selection result data. Simultaneously, the coordinate change trajectory contained in the structured behavior sequence object is extracted, and the coordinate change trajectory is projected onto the logical layout coordinate system of the task interface using a spatial topology mapping operator to generate trajectory topology features representing path search preferences, thereby obtaining the operation trajectory data. The operation trajectory data not only records the physical path of the operation but also contains the logical hesitation features in the decision-making process. By extracting the static results and dynamic trajectories, a preliminary dimensionality stripping of the decision-making behavior is achieved, and the generated selection result data and operation trajectory data are passed to the subsequent feature alignment process.

[0027] Preferably, in one scenario, step 1 is specifically implemented as follows: Response duration data and decision confidence data are extracted based on the structured behavior sequence object, and aggregated with the previously extracted features to generate the multi-dimensional interactive behavior sequence data. Specifically, for each decision task unit in the structured behavior sequence object, the time difference between the task information display node and the first valid operation node is calculated, generating a node response delay vector (e.g., a value within the range of 300 milliseconds to 1500 milliseconds), which is then aggregated to obtain the response duration data. Simultaneously, by analyzing the frequency of interaction modifications and the hovering duration of non-selected targets in the structured behavior sequence object, a confidence quantification component (e.g., a value distributed in the range of 0.1 to 0.95) is generated using a designed confidence evaluation function, thereby determining the decision confidence data. The decision confidence data characterizes the degree of certainty of the target object when performing a judgment. Finally, at least two of the selection result data, operation trajectory data, response duration data, and decision confidence data are fused along a time axis to generate high-fidelity multi-dimensional interaction behavior sequence data. The generated multi-dimensional interaction behavior sequence data is output as a high-dimensional feature tensor, serving as the input entity for fine-tuning the decision neural network model in subsequent steps.

[0028] Preferably, the refined extraction of multi-dimensional behavioral data described above alleviates the technical shortcomings of traditional solutions that focus only on a single decision outcome while ignoring the intermediate decision-making process. By introducing implicit feature components such as response time, operation trajectory, and confidence level, the multi-dimensional interactive behavior sequence data constructed in this application can more accurately reflect the deep reasoning preferences of the target object (e.g., whether it is a decisive decision or a hesitant, weighing-the-points decision). This feature extraction from the physical interaction level to the logical decision-making level provides a deterministic technical guarantee for the subsequent generation of a target adaptation parameter set with higher fitting degree and stronger logical consistency. Since the multi-dimensional interactive behavior sequence data is generated in a standardized task environment that excludes external interference, the nonlinear correlation between its feature dimensions can be more effectively captured by subsequent neural network models, thereby fundamentally improving the quality of the digital output of individual decision logic encapsulation.

[0029] Optionally, the decision neural network model adopts a Transformer architecture or a long short-term memory network architecture.

[0030] Preferably, in the specific implementation of the decision neural network model using the Transformer architecture, the decision neural network model first acquires the collected multi-dimensional interaction behavior sequence data through a built-in feature embedding layer. The feature embedding layer utilizes a designed dimension alignment transformation algorithm to map the originally discrete physical operation events (e.g., selection result data and operation trajectory data) to a high-dimensional continuous vector space with a preset dimension (e.g., 256 or 512 dimensions) to generate initial behavior embedding features. Subsequently, the decision neural network model calls the position encoding component to perform temporal position marking on the initial behavior embedding features, ensuring that subsequent processing can identify the logical order of decision actions. Through this transformation of physical interaction into mathematical representation, a preliminary semantic representation of the decision logic is achieved.

[0031] Preferably, in the specific technical implementation of the Transformer architecture: the decision neural network model includes multiple stacked self-attention computation layers. Each self-attention computation layer performs a correlation scan on the initial behavioral embedding features through its internal weight matrix. Specifically, the self-attention computation layer calculates the mutual influence weights between behavioral components at different time points, thereby generating an attention weight distribution matrix. In this process, the row and column intersection values ​​of the attention weight distribution matrix characterize the degree of dependence of the current decision action on the historical interaction context. Subsequently, the model utilizes a multi-head extraction mechanism to capture nonlinear correlation patterns in parallel from multiple semantic perspectives (e.g., one perspective focuses on the accuracy of the decision result, and another perspective focuses on the rhythm of the response duration), and combines this with a feedforward propagation layer to perform deep nonlinear mapping. Through this cascading cooperation between layers, the model can extract hidden decision patterns from massive amounts of sequence data.

[0032] Preferably, the specific implementation process of the decision neural network model using a long short-term memory network architecture is as follows: The decision neural network model performs streaming processing on multi-dimensional interactive behavior sequence data through an internally integrated cluster of temporal gating units. The temporal gating unit cluster includes a forget gate structure layer responsible for information filtering, an input gate structure layer responsible for state supplementation, and an output gate structure layer responsible for result output. Specifically, during execution, the forget gate structure layer reads the implicit state trajectory tensor of the previous decision time and uses the sigmoid activation function to calculate the retention probability of each feature component (the value is between 0.01 and 0.99) to discard random interference unrelated to the core decision pattern. Through this dynamic information filtering action, the stable characteristics in the decision logic are extracted.

[0033] Preferably, in the specific technical implementation of the Long Short-Term Memory (LSTM) network architecture: the input gate structure layer acquires the newly added behavioral feature components at the current interaction moment, and, in conjunction with the cell state update component, injects decision preference information with long-term impact (such as the target object's persistent hesitation characteristics when faced with high-risk options) into the global logical storage. Subsequently, the output gate structure layer performs linear transformation and hyperbolic tangent (Tanh) normalization processing based on the current global logical storage state, thereby producing a temporal implicit vector representing the cognitive state at the current decision moment. Through the continuous transmission and feedback of historical deviation information, the above structure layers achieve a more accurate fit to decision habits in the long temporal dimension, and finally, the fully connected output layer transforms this dynamically accumulated implicit knowledge into the target adaptation parameter set.

[0034] Preferably, the aforementioned deep design and hierarchical coordination logic of the model architecture alleviates the problem of encapsulated logic distortion caused by the lack of temporal correlation modeling in traditional technical solutions. By utilizing the global correlation capture capability of Transformer or the long-range memory filtering mechanism of LSTM, this application can transform the originally fuzzy and dynamic target object decision logic into a target adaptation parameter set with definite mathematical weights. This fine-tuning process based on deep learning hierarchical architecture not only improves the model's fitting depth to individual decision traits, but also ensures that the generated decision logic container has high logical interpretability and operational robustness through standardized parameter mapping. Compared with simple linear fitting, the model structure adopted in this application can mine deep causal reasoning chains from multi-dimensional behavioral data, thereby technically supporting the high-quality digital encapsulation of decision-making capability assets.

[0035] Optionally, the decision logic container conforms to the OpenContainerInitiative specification or the WebAssembly module specification.

[0036] Preferably, the specific implementation process of generating the decision logic container in step 4 is as follows: Using the designed image layer construction component, the target adaptation parameter set, quantization metadata, and standardized application programming interface (API) generated in the preceding steps are obtained. The image layer construction component first performs logical layer division according to the selected software package specification. When the selected specification conforms to the Open Container Initiative (OCI), the image layer construction component maps the target adaptation parameter set to a read-only data layer image and encapsulates the standardized API and its associated inference engine logic into an executable binary bootstrap layer. Subsequently, the image layer construction component writes the quantization metadata as an image configuration file into a manifest file. By performing this multi-level topology reorganization process, discrete model parameters and interface logic are integrated into a decision logic container with a file system abstraction layer. This process ensures that the generated assets have binary-level compatibility in different container orchestration environments, providing a physical foundation for the secure and isolated operation of subsequent decision logic execution.

[0037] Preferably, in the specific technical implementation of step 4: for a selected scenario conforming to the WebAssembly (Wasm) specification, the designed binary instruction stream generation operator is used to perform logic compilation processing. Specifically, the binary instruction stream generation operator performs linearization and serialization processing on the high-dimensional tensor data in the target adaptation parameter set, transforming it into a constant memory segment with a defined physical offset. Simultaneously, the binary instruction stream generation operator parses the logic of the standardized application programming interface into a portable assembly instruction set and introduces a capability-based security mechanism to perform permission binding on the quantized metadata. By performing atomic linking of the serialized parameters, the compiled instruction set, and the encrypted metadata, a decision logic container with a single-file structure is generated. The resulting decision logic container achieves complete decoupling of the decomposition logic from the host environment at the instruction level.

[0038] Preferably, in one scenario, step 4 is specifically implemented as follows: Runtime configuration injection is performed on the generated decision logic container using the designed environment dependency isolation mechanism. The environment dependency isolation mechanism, based on the performance boundaries defined in the quantified metadata, pre-sets an independent logic stack space and resource quota constraints for the decision logic container. In this process, by attaching a predefined communication contract to the decision logic container, external callers can only perform data interaction through the standardized application programming interface, thereby technically preventing unauthorized reading of the target adaptation parameter set. This encapsulation logic based on standardized image specifications alleviates the technical bottlenecks of intellectual property leakage and runtime environment conflicts that are easily generated during the cross-system distribution of complex decision characteristics, ultimately completing the digital standardized encapsulation of the target object's decision logic container.

[0039] Preferably, the aforementioned refined matching process for the mirror specification enhances the liquidity and security of the decision logic as a digital asset. By performing standardized encapsulation of the fitted model parameters and standardized interaction interfaces, the decision logic container produced in this application not only possesses "plug-and-play" engineering attributes, but also ensures, at the physical computing level, that decision logic from different sources does not interfere with each other when running on the same physical device through the native sandbox isolation capabilities provided by OCI or Wasm. This technological leap from model parameters to standardized software images provides a highly reliable and adaptable logical carrier for hybrid artificial intelligence systems to perform decision tasks for different data distributions, enhancing dynamic evolution performance in heterogeneous computing environments.

[0040] Optionally, the decision logic container includes a metadata layer, an interface layer, and a core model layer; the metadata layer defines the attribute description information of the target adaptation parameter set; the interface layer configures the communication protocol for external systems to access the core model layer; and the core model layer encapsulates the algorithm weights for performing decision calculations.

[0041] Preferably, in the internal architecture design of the decision logic container, a standardized encapsulation of decision assets is achieved by constructing a three-layer decoupled logical architecture. Specifically, a structured metadata mapping module is used to perform multi-dimensional attribute extraction processing on the target adaptation parameter set. The structured metadata mapping module scans the topology of the target adaptation parameter set, extracts the hash sequence representing the model version, the corresponding creator digital signature, and the quantified performance components describing the model in the risk control or performance output dimensions, thereby defining the attribute description information of the target adaptation parameter set and writing it into the metadata layer of the decision logic container. This metadata layer transforms unstructured parameter features into a set of attributes with physical self-description, providing a definite logical basis for subsequent retrieval and rights confirmation in a heterogeneous computing environment.

[0042] Preferably, in the specific technical implementation of the decision logic container, a predefined communication specification description document is obtained by calling a designed standard contract configuration operator. The standard contract configuration operator parses the access requirements of external callers (such as intelligent financial risk control systems or task allocation engines) and configures the interface layer inside the decision logic container accordingly. In this process, the standard contract configuration operator maps the corresponding communication protocol stack (e.g., a protocol flow based on a remote procedure call protocol or a representational state transfer architecture) to the decision logic container and generates a corresponding standardized application programming interface. By configuring the communication protocol for external systems to access the core model layer at the interface layer, physical isolation between internal high-dimensional tensor operations and external business logic is achieved, ensuring protocol consistency and data interaction security during cross-system calls of the decision logic.

[0043] Preferably, in one scenario, when constructing the core computational logic of the decision logic container, the target adaptation parameter set produced after the aforementioned fine-tuning is obtained by activating the designed weight tensor solidification component. The weight tensor solidification component maps the neuron connection weights (e.g., floating-point numerical components in the range of -1.0 to +1.0) contained in the target adaptation parameter set to constant segments in a binary image, thereby constructing a core model layer within the decision logic container. During this process, the core model layer synchronously loads the operator execution logic corresponding to the decision neural network model architecture (such as the Transformer architecture) to encapsulate the algorithm weights for performing decision computation. This encapsulation method, which atomically binds algorithm weights to execution operators, allows the core model layer to run independently of the original training environment.

[0044] Preferably, the cascaded encapsulation logic for the metadata layer, interface layer, and core model layer alleviates the technical bottlenecks of high environmental dependence and weak intellectual property protection in the delivery process of traditional decision-making models. Since the metadata layer provides a deep description of the target adaptation parameter set, the interface layer standardizes the external interaction path, and the core model layer achieves a physical closed loop of logic, the decision logic container produced in this application not only possesses high portability but also prevents unauthorized read / write access to core parameters through inter-layer decoupling mechanisms. This three-layer deconstruction encapsulation scheme provides a standardized implementation scheme for transforming discrete decision characteristics based on behavioral data fitting into digital assets with industrial circulation value, improving the deployment efficiency and security of artificial intelligence decision-making components in complex task scenarios.

[0045] Optionally, the metadata layer includes a unique hash identifier generated based on the binary content of the decision logic container, a creator identity identifier processed by asymmetric encryption, and a quantized capability vector representing the decision dimension of the target adaptation parameter set.

[0046] Preferably, the specific implementation process of recording the unique hash identification code in the metadata layer is as follows: Obtain the full binary byte stream corresponding to the encapsulated decision logic container, where the full binary byte stream covers the core model layer, interface layer, and initial configuration parameters. Perform unidirectional feature mapping processing on the full binary byte stream using a designed digest feature extraction operator to produce a bit string representing the physical uniqueness of the decision logic container, thereby determining the unique hash identification code. In the processing logic, the digest feature extraction operator employs a hash calculation logic with collision resistance properties, ensuring that any minute binary change within the decision logic container will cause a state shift in the generated unique hash identification code. By writing the generated unique hash identification code into the asset description area of ​​the metadata layer, the physical fingerprint anchoring of the digital decision logic asset is achieved, providing a deterministic verification basis for subsequent content integrity verification in a distributed transaction environment.

[0047] Preferably, the specific technical implementation of recording the creator's identity in the metadata layer involves: obtaining the initial signature feature associated with the current decision logic container, wherein the initial signature feature includes the creator's digital certificate serial number and the timestamp component of the encapsulation time. A digital signature operation is performed on the initial signature feature using the private key in the designed asymmetric encryption algorithm to generate an identity authentication message with physical anti-counterfeiting attributes. Subsequently, the identity authentication message is binary encoded to generate the creator's identity and embedded in the metadata layer. During this process, the creator's identity is atomically bound to the aforementioned unique hash identification code. Because this identifier is generated using asymmetric encryption technology, external verification entities can only perform decryption verification on the creator's identity using the corresponding public key, thereby ensuring, in essence, the immutability and judicial traceability of the ownership of decision logic assets, effectively mitigating the risk of false intellectual property injection during the assetization of cognitive abilities.

[0048] Preferably, in a specific implementation of generating a quantized capability vector in a given scenario: The designed parameter dimension decomposition operator is used to obtain the target adaptation parameter set encapsulated in the core model layer. The parameter dimension decomposition operator performs sensitivity scanning on the target adaptation parameter set across multiple preset logical dimensions. Specifically, for application scenarios involving financial investment or risk game theory, the parameter dimension decomposition operator calculates the weight distribution components of the target adaptation parameter set on the risk preference dimension, decision response speed dimension, and logical consistency dimension, respectively. Subsequently, a feature standardization algorithm is used to map the weight distribution components of each dimension to a preset interval (e.g., score values ​​between zero and one hundred), and they are arranged in topological order to generate a quantized capability vector representing the decision dimension of the target adaptation parameter set.

[0049] Preferably, in the above detailed implementation of the internal components of the metadata layer: the generated quantified capability vector is then written into the capability declaration area of ​​the metadata layer. When performing decision asset transaction matching, the distributed trading platform first reads the quantified capability vector in the metadata layer and performs spatial vector cosine similarity calculation between it and the target decision characteristics of the demand side, thereby generating a decision logic fit score. In response to the decision logic fit score meeting a preset recommendation threshold, the physical storage address of the decision logic container is locked using the unique hash identification code. This multi-dimensional vector representation-based technique allows implicit decision characteristics to be transformed into explicit technical indicators that can be retrieved by the algorithm, achieving high-quality automated classification and supply-demand matching of complex decision logic assets without exposing the underlying core algorithm weights.

[0050] Preferably, the aforementioned unique hash identifier, creator identity identifier, and quantified capability vector are logically coupled within the metadata layer through a verification link mechanism. Specifically, the creator identity identifier is used to perform a complete digital encapsulation of the quantified capability vector, ensuring a physical technical binding between the capability level described by the metadata layer and the creator's credit system. This hierarchical data encapsulation logic not only enhances the credibility of the decision logic container in the distributed trading platform but also provides a benchmark performance reference for subsequent sandbox verification tasks conducted in a trusted hardware execution environment through the introduction of quantitative indicators. Compared to traditional general software documentation, this application, through the refined construction of the metadata layer, enables the generated decision logic assets to possess extremely high descriptive accuracy and technical certainty, thereby supporting the reliable transformation of decision capabilities into standardized digital assets.

[0051] Optionally, the standardized application programming interface includes an initialization interface, a prediction and decision interface, and a logic interpretation interface; the initialization interface is used to load runtime environment configuration information; the prediction and decision interface is used to receive a structured decision scenario description, and based on the target adaptation parameter set, outputs a standardized decision result object containing a list of action options, a preference probability distribution, and a decision confidence level.

[0052] Preferably, the specific implementation process of configuring the standardized application programming interface (API) and its internal interaction logic in step 3 is as follows: Using the designed interface contract mapping module, the standardized API is constructed at the logical boundary of the decision logic container to achieve isolation and interaction between external call instructions and internal model states. The standardized API is decomposed into a complementary initialization interface, a prediction decision interface, and a logic interpretation interface, and follows a predefined underlying communication protocol. Through this multi-dimensional interface segmentation, the environment decoupling and call standardization of the decision logic during physical runtime are achieved. The resulting standardized API serves as the only legitimate channel for external systems to enter the decision logic container, ensuring the security and black-box operation of the internal target adaptation parameter set.

[0053] Preferably, in the specific technical implementation of step 3: the initialization interface is used to load runtime environment configuration information to complete the warm-up of the decision logic container. Specifically, the initialization interface receives runtime environment configuration information from the external main system, including memory quota components, thread priority components, and log collection levels. After receiving the above information, the initialization interface calls the runtime environment establishment operator inside the container, and performs address alignment and resource anchoring on the sandbox environment (e.g., WebAssembly virtual memory space) according to the loaded runtime environment configuration information, thereby generating a ready-to-execute environment. This processing action realizes the dynamic definition of the physical parameters at the execution time of the decision logic container, ensuring that subsequent inference tasks can run smoothly based on preset hardware constraints.

[0054] Preferably, in one scenario, the prediction and decision interface is used to perform logical reasoning for a specific task. Specifically, the prediction and decision interface receives an externally input structured decision scenario description (e.g., standardized message data containing financial market indicators, asset volatility sequences, or risk factor characteristics). The prediction and decision interface then calls an internal high-dimensional feature transformation component to perform feature space projection on the structured decision scenario description, transforming it into an input feature tensor with a preset dimension. Next, the prediction and decision interface drives computational logic to perform hierarchical weighted mapping and activation operations on the input feature tensor in conjunction with the target adaptation parameter set, producing raw score components representing the probability distribution of the decision outcome. Through this interface-triggered deep reasoning process, the static scenario description is mapped into a dynamic expression of decision preferences.

[0055] Preferably, after completing the calculation, the prediction and decision interface performs structured encapsulation of the output results. Specifically, the prediction and decision interface obtains the original score components and calls the output format standardization module to perform normalization mapping to generate a standardized decision result object containing an action option list (e.g., a discrete set of options representing increasing holdings, decreasing holdings, or remaining on the sidelines), a preference probability distribution (e.g., the mapping scores associated with each option, with values ​​distributed between 0.01 and 0.99), and a decision confidence score (a confidence interval component representing the certainty of the model's judgment). The generated standardized decision result object is output in a protocol format conforming to predefined communication specifications, thereby achieving lossless transmission of decision outputs between heterogeneous systems. This technical means of forcibly constraining the output format through the interface alleviates the application integration difficulties caused by the uncertainty of complex model outputs and produces decision output components with high engineering friendliness.

[0056] Preferably, the logical interpretation interface synchronously performs source analysis on decision motivation. Specifically, the logical interpretation interface obtains the neuron activation path trajectories generated by the predictive decision interface during the inference process and calls the built-in feature-sensitive quantification operator. The feature-sensitive quantification operator performs contribution reduction on each feature dimension in the structured decision scenario description, generating a feature contribution distribution map representing the decision cause. The logical interpretation interface generates textual decision interpretation metadata based on this distribution map and incorporates it as an additional component into the standardized decision result object, thereby improving the transparency of the decision process. The above interfaces, through the cascading cooperation of "environment initialization - scenario reception - weight mapping - result standardization - attribution interpretation", achieve an efficient, secure, and transparent external presentation of the decision logic of the target object, providing a complete technical interface foundation for subsequent asset invocation and performance verification in the distributed trading platform.

[0057] Optionally, the standardized decision result object further includes a summary of key decision factors; the summary of key decision factors is generated by performing a reverse attribution analysis on the feature contribution of the core model layer when performing predictive decision through the logical interpretation interface.

[0058] Preferably, the specific implementation process of the logical interpretation interface in generating the summary of key decision factors is as follows: Using a designed logical tracing operator, the activation state data of neurons in the core model layer during the execution of the prediction decision task is obtained. The logical tracing operator performs tensor capture on the output features of each structural layer within the neural network (e.g., the self-attention mechanism layer in the Transformer architecture or the temporal gating unit in the Long Short-Term Memory network architecture) to extract the neuron activation traces representing the decision path. Subsequently, the logical interpretation interface uses these neuron activation traces to perform gradient backtracking operations on each dimension of the feature components in the structured decision scenario description (e.g., numerical features representing market fluctuations or textual semantic features representing policy guidance) to calculate the sensitivity components of each dimension of the feature components to the final decision probability distribution. By performing quantization mapping on the sensitivity components, a feature contribution value reflecting the strength of the causal relationship between the input features and the decision result is generated. Since the feature contribution value directly anchors the original features that influence the judgment result, the logical interpretation interface further uses a preset text mapping logic to semantically combine the top-ranked feature dimensions and their corresponding logical meanings, thereby generating the summary of key decision factors. This technique, which transforms the implicit neural network weight transformation process into explicit logical basis, enables the generated summary of key decision factors to intuitively reveal the technical support behind the model's decision-making, thereby improving the interpretability of the decision logic container.

[0059] Preferably, in the specific technical implementation of the summary of key decision factors: a designed reverse attribution analysis module is used to perform in-depth analysis of the feature contribution. Specifically, the reverse attribution analysis module adopts attribution logic based on integral gradients. It constructs a linear interpolation path between a preset baseline feature matrix and the current structured decision scenario description, and calculates the cumulative gradient value generated by the core model layer on the linear interpolation path to obtain the feature contribution. In this processing action, the generated feature contribution is presented in the form of a sensitivity matrix of the feature dimension, where the higher the element value of the matrix, the stronger the dominant role of that feature in the current decision. Subsequently, the reverse attribution analysis module inputs the feature contribution to a semantic alignment unit. The semantic alignment unit, by retrieving a preset feature semantic lookup table, converts the abstract feature index into readable logical phrases (e.g., "market liquidity depletion" or "golden cross signal of moving average system") to obtain the summary of key decision factors. The generated summary of key decision factors is dynamically injected into the standardized decision result object, realizing the atomic binding of decision outputs and decision basis, and alleviating the technical bottleneck that traditional decision models, as "black boxes," cannot meet compliance verification and trust assessment.

[0060] Preferably, in one scenario, when the logical interpretation interface processes the financial market prediction decision-making task, it obtains the original decision score produced by the core model layer and simultaneously retrieves the macroeconomic indicator feature stream corresponding to that moment. The logical interpretation interface, through the reverse attribution analysis module, detects that in the generated standardized decision result object, the generation of a certain position reduction suggestion is mainly driven by the top three negative features in terms of feature contribution (e.g., the interest rate rise expectation component, the industry valuation premium component, and the trading volume divergence component). The semantic alignment unit then performs cluster analysis on the dimension with the highest feature contribution and, based on the associated physical meaning description, generates a summary of the key decision factors containing the content "Core risk factor driven: rising interest rates lead to increased valuation pressure".

[0061] Preferably, the detailed implementation of the interaction logic between the logic interpretation interface and the core model layer alleviates the technical deficiency of "attribution difficulty" caused by the extreme complexity of individual decision-making logic. By introducing reverse attribution analysis based on gradient path accumulation, this application can reconstruct the nonlinear decision-making logic captured by the model during fine-tuning with high confidence, and transform these abstract weight distributions into a summary of key decision-making factors with business guidance significance. This deep attribution mechanism not only provides logical support for subsequent buyers to verify the performance of the decision logic container, but also, by aligning feature contribution with the standardized output results, technically constructs a transparent and verifiable cognitive ability assetization framework. The generated summary of key decision-making factors, because it closely follows the feature field distribution of the original input data, reflects a causal relationship with high technical certainty, thereby greatly enhancing the application adaptability of the encapsulated decision logic in heterogeneous business environments.

[0062] Optionally, before fine-tuning the decision neural network model, the method further includes: performing temporal feature aggregation processing on the multi-dimensional interactive behavior sequence data to extract the operational rhythm features and feature attention weights of the target object when making decisions; mapping the operational rhythm features and feature attention weights into a structured decision feature matrix; and using the structured decision feature matrix as a fine-tuning constraint term for the decision neural network model.

[0063] Preferably, before fine-tuning the decision neural network model, the specific implementation process for extracting the operational rhythm features and the feature attention weights is as follows: The designed temporal interaction feature deconstruction operator is used to obtain the multi-dimensional interactive behavior sequence data. The temporal interaction feature deconstruction operator performs discrete derivative operations in the time domain on each atomic event contained in the multi-dimensional interactive behavior sequence data to determine the time interval distribution between adjacent operation instructions. Subsequently, the temporal interaction feature deconstruction operator performs statistical analysis processing based on information entropy on the time interval distribution, extracting feature components characterizing the volatility of operation frequency and the decision hesitation cycle, thereby determining the operational rhythm features (e.g., in a high-frequency financial trading decision-making scenario, characterizing the temporal jump pattern from observing market reports to executing order instructions). Simultaneously, using a built-in feature scanning component, feature space mapping is performed on the coordinate pointing component and dwell time component in the multi-dimensional interactive behavior sequence data to identify the visual capture or operational focus duration of the target object for different information dimensions (e.g., characterizing financial indicators, market sentiment, or policy fluctuations) when performing tasks, in order to generate the feature attention weights. By analyzing the deep time-frequency characteristics of the original interaction flow, the invisible decision-making behavior style is transformed into quantifiable physical feature indicators, providing a feature foundation with high fidelity for the subsequent construction of logical profiles of target objects.

[0064] Preferably, in the specific technical implementation of mapping the operational rhythm features and the feature attention weights into a structured decision feature matrix: the designed high-dimensional tensor mapping logic is used to obtain the aforementioned operational rhythm features and the feature attention weights. The high-dimensional tensor mapping logic establishes a numerical array with a two-dimensional topological structure, thereby generating the structured decision feature matrix. Specifically, the rows of the structured decision feature matrix represent different execution stages of the standardized decision task (e.g., data retrieval stage, risk assessment stage, and final decision stage), and the columns represent the differentiated behavioral dimensions exhibited by the target object (e.g., temporal rhythm dimension, information gain dimension, and confidence level dimension). The elements at the intersection of rows and columns represent the coupling influence coefficients of the set behavioral dimensions under the corresponding execution stage (e.g., values ​​within the range of -1.0 to +1.0). By performing the tensor product operation of the operational rhythm features and the feature attention weights, and filling the result into the corresponding index address of the numerical array, the scattered rhythm and weight features are integrated into the structured decision feature matrix with spatiotemporal consistency. This matrix representation effectively reveals the dynamic weight distribution of decision logic under different time slices, providing a highly deterministic mathematical constraint for the adaptation of neural networks.

[0065] Preferably, in a scenario, when the structured decision feature matrix is ​​used as a fine-tuning constraint term for the decision neural network model, the following steps are taken: The structured decision feature matrix is ​​obtained using a designed regularization constraint injection operator, and the global loss function of the decision neural network model is simultaneously obtained. The regularization constraint injection operator performs a penalty term transformation on the structured decision feature matrix based on the Frobenius norm (i.e., the square root of the sum of squares of the matrix elements) to generate a logical alignment constraint component. Subsequently, the logical alignment constraint component is superimposed as a regularization term onto the global loss function to form a target-guided composite loss function. During gradient descent updates, the parameter evolution process of the decision neural network model is subject to the mandatory constraints of the target-guided composite loss function. Specifically, this constraint mechanism mandates that the neural network, while fitting the selection result, must simultaneously fit the operational style and focus represented by the structured decision feature matrix, thereby ensuring that the generated target-fit parameter set not only matches the target object in the result dimension but also maintains a high degree of coordination with the target object's operational habits in the evolution trajectory of the inference logic.

[0066] Preferably, the above-mentioned construction and constraint application of the structured decision feature matrix alleviates the technical defects of model overfitting or preference drift caused by the relatively sparse amount of fine-tuning data. By introducing rhythmic features and attention weights representing decision-making styles before fine-tuning, this application can provide directional guidance to the hidden state of the decision neural network model during the fine-tuning process. Compared with conventional schemes that rely solely on the final selection result for fine-tuning, this application solidifies the implicit rhythmic pattern into a matrix form and injects it into the loss function, enabling the target adaptation parameter set to carry the "thinking rhythm" and "feature attention habits" of the target object when performing complex decision-making tasks. This technique of fitting from both the execution process and the execution result improves the logical restoration of the decision logic container in heterogeneous decision-making scenarios, providing a technical closed loop with rigorous logical support for the standardization and high-quality encapsulation of cognitive ability assets. Because the generated target adaptation parameter set has the imprint of this underlying behavioral feature, its decision suggestions will be more in line with the real reasoning paradigm of the target object, thereby achieving a deep digital mapping of complex cognitive logic.

[0067] Optionally, the fine-tuning of the decision neural network model includes: fixing the shallow network weights representing the extraction of general logic in the decision neural network model, and performing gradient updates on the deep network weights representing the mapping of decision preferences based on the multi-dimensional interaction behavior sequence data, so as to generate the target adaptation parameter set.

[0068] Preferably, in the specific technical implementation of step 2: a preset general decision neural network model is obtained using a loading algorithm. The general decision neural network model is divided into a low-order logic layer for performing feature space mapping and a high-order mapping layer for performing action policy output, based on layer depth. Subsequently, using a designed weight locking operator, the shallow network weights representing the extraction of general logic contained in the low-order logic layer are obtained. The weight locking operator modifies the gradient variable flag of the shallow network weights in the backpropagation path, setting their physical state to read-only mode (e.g., setting the learning rate parameter of gradient descent to a zero component). By freezing the shallow logic, the general common-sense decision logic (e.g., the ability to extract commonalities from basic financial market indicators) is fully preserved, ensuring that the model will not suffer from knowledge forgetting due to a small sample size during subsequent fine-tuning.

[0069] Preferably, in the parameter evolution process of step 2: the multi-dimensional interaction behavior sequence data collected in the preceding steps is acquired and input into the low-order logic layer. The low-order logic layer performs high-dimensional feature extraction on the input raw signal based on the shallow network weights representing the general logic extraction, to generate a general logic feature vector. Then, the generated general logic feature vector is used as a driving stimulus and passed to the activated high-order mapping layer. The high-order mapping layer performs a nonlinear transformation on the general logic feature vector based on the deep network weights representing the decision preference mapping contained within it. In this processing action, by comparing the output value of the high-order mapping layer with the actual decision results in the multi-dimensional interaction behavior sequence data, the preference mapping bias gradient reflecting the decision difference is calculated using the designed cost function evaluation operator.

[0070] Preferably, in one scenario, when performing differential fine-tuning on the higher-order mapping layer: the gradient-guided update operator is used to obtain the aforementioned generated preference mapping deviation gradient. The gradient-guided update operator bypasses the shallow links that are locked and applies the preference mapping deviation gradient to the deep network weights. By performing parameter correction actions based on the chain rule, the values ​​of the deep network weights are driven to undergo state transition in the direction of fitting the decision style of the target object, thereby producing an updated target fitting parameter set. Since this process only performs gradient injection on the deep decision boundary, it can capture the unique judgment habits of the target object under a set logical dimension (e.g., risk aversion tendency when facing fluctuations in the debt-to-equity ratio) with lower computational cost (e.g., reducing computational overhead by 60% to 80% compared to a full update).

[0071] Preferably, the aforementioned synergistic approach of fixing shallow weights and updating deep weight gradients fundamentally alleviates the technical contradiction between "accurate fitting of decision logic" and "guarantee of computational stability." Since shallow weights carry widely applicable basic logic, a physical locking mechanism prevents damage to the general knowledge structure; while deep weights, through sensitive capture of multi-dimensional interactive behavior sequence data, achieve a deep digital mapping of the target object's specific decision-making habits. This layered processing technique ensures that the generated target adaptation parameter set not only possesses a solid logical generalization foundation but also exhibits extremely high fitting accuracy in specific decision-making scenarios. The final target adaptation parameter set is solidified and output in the form of a weight matrix, providing the core algorithm kernel for constructing a decision logic container with independent operational capabilities in subsequent steps.

[0072] like Figure 2The illustration shows a transaction verification method for a decision logic container according to an embodiment of this application, applied to a distributed trading platform. The method includes: receiving the encapsulated decision logic container and recording its hash identification code in a blockchain ledger; responding to a verification request from a buyer, allocating a trusted hardware execution environment as a sandbox verification environment for the decision logic container; loading the decision logic container into the sandbox verification environment and inputting a test decision scenario to obtain the corresponding standardized decision result object; feeding back the standardized decision result object to the buyer, and preventing the sandbox verification environment from externally reading or writing to the target adaptation parameter set during the verification process.

[0073] Optionally, it also includes: a smart contract for deploying and executing transaction logic, wherein the smart contract is pre-set with the asset identifier of the decision logic container, the royalty allocation ratio, and the access token required to decrypt the core model layer; in response to the smart contract meeting the execution conditions, the access token is distributed to the buyer to authorize the buyer to obtain full access to the decision logic container.

[0074] 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 stored in the memory to implement the steps of the above-mentioned decision logic encapsulation method or the steps of the transaction verification method.

[0075] like Figure 4 As shown in the figure, a decision logic encapsulation system for a target object is provided in an embodiment of this application. The system includes: a behavior access terminal for collecting multi-dimensional interactive behavior sequence data; a parameter mapping server for generating a target-adaptive parameter set; and a container encapsulation module for generating a decision logic container.

[0076] Preferably, this application employs a system architecture based on physical hardware coupling. The system includes an interactive signal acquisition terminal comprised of a multi-dimensional sensing matrix, a local buffer, and a network physical interface. This interactive signal acquisition terminal physically implements the functionality of a behavior access terminal. Specifically, the interactive signal acquisition terminal utilizes an internal level detection circuit to capture external input pulses (such as button closure signals or touch coordinate displacement voltages) triggered by the target object when performing standardized decision-making tasks in real time. The local buffer performs digital sampling of the captured raw physical signals with a preset bit width, generating a behavioral electrical signal recording stream corresponding to the time axis. Subsequently, the network physical interface, according to the loaded communication protocol stack, encapsulates the behavioral electrical signal recording stream into data packets that can be transmitted on a high-speed bus, thereby achieving precise conversion of subjective interactive actions into physical digital signals.

[0077] Preferably, Figure 4 The parameter mapping server mentioned above is implemented through a heterogeneous tensor acceleration cluster. This heterogeneous tensor acceleration cluster acquires behavioral data packets output by the aforementioned interactive signal acquisition terminal via a physical data link. The heterogeneous tensor acceleration cluster includes an instruction stream scheduling array, a high-bandwidth memory array, and a massively parallel multiply-accumulate operation core. Specifically, during execution, the instruction stream scheduling array retrieves the initial weight matrix of a preset decision neural network model from the high-bandwidth memory array. Subsequently, the massively parallel multiply-accumulate operation core performs gradient backpropagation operations on the hidden layer parameters of the model within the physical memory address space based on the received behavioral data packets. In this processing action, the hardware core adjusts the state of the charge storage unit to perform preference adaptation updates on the weight distribution in a high-dimensional space, thereby generating a target adaptation parameter set corresponding to the data distribution in the physical storage medium.

[0078] Preferably, Figure 4 The container encapsulation module is implemented by a firmware image generation controller. This controller includes a dedicated non-volatile storage control chip and a cryptographic security operator with hardware encapsulation capabilities. The firmware image generation controller obtains the target adaptation parameter set generated by the aforementioned heterogeneous tensor acceleration cluster and retrieves the binary instruction segment of the standardized application programming interface from the local read-only memory. Using the non-volatile storage control chip, the system writes the target adaptation parameter set and the binary instruction segment to consecutive physical sectors of the storage medium according to a preset file offset, and the cryptographic security operator adds a unique hash identifier generated based on a hardware fingerprint. Through this storage mapping and instruction linking technology, a decision logic container with independent bootstrapping capabilities is ultimately generated on the physical storage medium.

[0079] Preferably, the aforementioned hardware entities achieve data interconnection and timing coordination through a system-level backplane switching bus. Because the entire system achieves a complete evolution from external behavioral signal capture to the physical solidification of internal logical parameters through the physical cascading of the interactive signal acquisition terminal, heterogeneous tensor acceleration cluster, and firmware image generation controller, this hardware entity-based encapsulation scheme alleviates the technical bottleneck of physically isolating and assetizing behavioral patterns due to the intangible nature of individual cognitive logic. The generated decision logic container is directly solidified in the storage medium, providing a physically deterministic logical carrier for its subsequent atomic deployment and controlled execution on heterogeneous computing devices, improving the security and cross-system compatibility of decision assets in complex business networks.

[0080] The above Figures 2-4 For an exemplary description, please refer to the above. Figure 1 .

Claims

1. A method for encapsulating decision logic for a target object, characterized in that, Includes the following steps: Step 1: Collect multi-dimensional interaction behavior sequence data of the target object during the execution of standardized decision-making tasks; Step 2: Based on the multi-dimensional interaction behavior sequence data, fine-tune the decision neural network model to generate a target fitting parameter set that fits the distribution of the multi-dimensional interaction behavior sequence data. Step 3: Extract quantitative metadata describing the decision-making effectiveness of the target adaptation parameter set under different logical dimensions, and configure a standardized application programming interface that conforms to predefined communication specifications. Step 4: Encapsulate the target adaptation parameter set, the quantized metadata, and the standardized application programming interface into a decision logic container that conforms to the set software package specifications, so as to achieve standardized encapsulation of the decision logic of the target object.

2. The decision logic encapsulation method for a target object according to claim 1, characterized in that, The multi-dimensional interactive behavior sequence data includes at least two of the following: selection result data, operation trajectory data, response time data, and decision confidence data under a preset decision-making scenario.

3. The decision logic encapsulation method for a target object according to claim 1, characterized in that, The decision neural network model adopts either the Transformer architecture or the Long Short-Term Memory network architecture.

4. The method for encapsulating decision logic for a target object according to claim 1, characterized in that, The decision logic container conforms to the OpenContainerInitiative specification or the WebAssembly module specification.

5. The method for encapsulating decision logic for a target object according to claim 1, characterized in that, The decision logic container includes a metadata layer, an interface layer, and a core model layer; the metadata layer defines the attribute description information of the target adaptation parameter set; and the interface layer configures the communication protocol for external systems to access the core model layer. The algorithm weights for performing decision calculations are encapsulated through the core model layer.

6. The method for encapsulating decision logic for a target object according to claim 5, characterized in that, The metadata layer includes a unique hash identifier generated based on the binary content of the decision logic container, a creator identity identifier processed by asymmetric encryption, and a quantized capability vector representing the decision dimension of the target adaptation parameter set.

7. The method for encapsulating decision logic for a target object according to claim 1, characterized in that, The standardized application programming interface includes an initialization interface, a prediction and decision interface, and a logic interpretation interface; the initialization interface is used to load runtime environment configuration information; the prediction and decision interface is used to receive a structured decision scenario description and, based on the target adaptation parameter set, outputs a standardized decision result object containing a list of action options, a preference probability distribution, and a decision confidence level.

8. A transaction verification method for the decision logic container according to any one of claims 1-7, characterized in that, The method, applied to a distributed trading platform, includes: receiving the encapsulated decision logic container and recording its hash identification code in a blockchain ledger; responding to a buyer's verification request, allocating a trusted hardware execution environment as a sandbox verification environment for the decision logic container; loading the decision logic container into the sandbox verification environment and inputting a test decision scenario to obtain the corresponding standardized decision result object; feeding back the standardized decision result object to the buyer, and preventing the sandbox verification environment from externally reading and writing to the target adaptation parameter set during the verification process.

9. The decision logic transaction verification method according to claim 8, characterized in that, Also includes: The smart contract is associated with the deployment and execution of transaction logic. The smart contract is pre-set with the asset identifier, royalty allocation ratio, and access token required to decrypt the core model layer of the decision logic container. In response to the smart contract meeting the execution conditions, the access token is distributed to the buyer to authorize the buyer to obtain full access to the decision logic container.

10. A decision logic encapsulation system for a target object, characterized in that, include: A behavior access terminal is used to implement the operation of collecting multi-dimensional interactive behavior sequence data in any one of claims 1-8; A parameter mapping server is used to implement the operation of generating a target adaptation parameter set in any one of claims 1-8; A container encapsulation module is used to implement the operation of generating a decision logic container as described in any one of claims 1-8.

Images