A method and system for constructing a low-intrusion intelligent agent for government office work

By combining capability contracts and execution receipts, a versioned task node graph is generated and parameter verification and compliance gating are performed, which solves the synchronization problem between heterogeneous systems in government office systems and achieves low-intrusion construction and efficient government office process management.

CN123118041APending Publication Date: 2026-11-17CHINA UNICOM (JIANGSU) IND INTERNET CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611620835.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-20
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

The existing government office system lacks synchronous correlation in the process of interface calls and policy verification between heterogeneous systems. This leads to a disconnect between the execution results and the business status when the page is upgraded, the return format changes, or the network domain changes, resulting in problems such as duplicate writing, process interruption, and difficulty in restoring the operation basis.

Method used

By unifying the interface calls and authorization interface operations through capability contracts, a versioned task node graph is generated. Parameters are bound to each node and the business primary key is verified using execution receipts. Successful nodes are frozen, and affected unexecuted nodes are replaced. Compliance gating is performed in conjunction with policy document versions and risk levels to form a task-level audit package.

Benefits of technology

It enables low-intrusion construction of government office systems, reduces the impact of underlying modifications, lowers the risk of duplicate writes and process interruptions, keeps the policy version and authorization conditions of execution nodes synchronized with business status, and improves the traceability of results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118041A_ABST
    Figure CN123118041A_ABST
Patent Text Reader

Abstract

This invention relates to the field of information processing technology, specifically to a low-intrusion method and system for constructing intelligent agents for government office work. This invention unifies interface calls and authorized interface operations into business capabilities through capability contracts, enabling task plans generated by large language models to uniformly call heterogeneous interfaces or authorized interfaces. When adding new government business systems, only field mappings, control features, and postconditions need to be configured on the execution plane; the control plane does not need to access the target database or change the original approval logic, thereby reducing the impact of underlying modifications on the stability of existing systems and achieving low-intrusion construction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of information processing technology, specifically relating to a low-intrusion method and system for constructing intelligent agents for government office work. Background Technology

[0002] Government affairs processing typically begins with a designated person submitting a request for processing. This process involves identity verification, document retrieval, policy inquiries, cross-system data entry, and process triggering, ultimately resulting in a draft document, approval record, or business completion outcome. The government affairs systems involved in this process—including document processing, personnel management, and approval systems—may have been built by different vendors at different times, resulting in differences in their interface protocols, page structures, data formats, authentication methods, and security domains. Furthermore, some networks may be physically or logically isolated from each other.

[0003] Existing technologies typically involve writing separate interface programs for each government business system. When a usable interface is unavailable, a front-end interface automated script is used, and the calling parameters and execution order are generated all at once by a large language model or rule program. The interface program depends on the established fields and return formats of each system, the interface script depends on the position of controls and a fixed operation order, and cross-system tasks depend on a pre-arranged call queue. Policy verification is often performed independently after the text is generated, and tool call records, policy evidence, and manual confirmation records are also saved separately.

[0004] However, page upgrades, changes in return formats, missing materials, or changes in network domains can cause fixed adaptation relationships to deviate from the current business objects, resulting in a disconnect between the actual results of preceding actions and the parameters and paths of actions that have not yet been executed. Existing one-time call queues separate adaptation, execution, and verification, but do not continuously correct unexecuted actions based on subsequent business states, policy versions, and authorization conditions. Partial success, lost responses, or rule conflicts can easily be misjudged as continuing execution. When policy verification and action execution are performed separately, there is a lack of synchronization between the policy version and authorization conditions used by the execution node and the actual business state. Furthermore, execution results, policy evidence, and authorization status are scattered across different records, causing deviations between the release conditions of subsequent actions and the current business state. All of these factors combined result in heterogeneous government business systems lacking a continuous processing procedure to adjust subsequent actions based on execution feedback and retain compliance evidence while maintaining the original underlying architecture. This leads to consequences such as duplicate writes, invalid policy references, process interruptions, and difficulty in restoring operational basis.

[0005] To address the above problems, this invention proposes a low-intrusion method and system for constructing intelligent agents for government office work. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, the present invention aims to provide a low-intrusion method and system for constructing intelligent government affairs agents.

[0007] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: A method for constructing a low-intrusion intelligent agent for government office work includes: Obtain the user's natural language instructions, initiator identity reference, authorized role, attachment reference, and target security domain, and replace the credentials with credential references to form a task envelope; Based on the task envelope retrieval capability contract registry, select the capability contract that matches the business object, operation type, read / write attributes, and security domain, and generate a versioned task node graph through dependency analysis; Perform parameter validation on the nodes in the versioned task node graph and generate node execution requests; Within the security domain of the government business system, based on the node execution request and capability contract, existing interfaces or operation authorization interfaces are invoked to generate execution receipts; Based on the execution receipt, feedback replanning is performed, successful nodes that meet the expected post-conditions are frozen, affected unexecuted nodes are replaced, and compensation is processed. Compliance gates are implemented based on policy evidence and task risks to generate a one-time execution permit; Verify the final execution receipt and solidify the task-level audit package.

[0008] Preferably, the capability contract records the input field type, preconditions, execution method, expected postconditions, idempotent fields, compensation actions, and security domains; the capability contract is divided into interface-type capabilities and interface-type capabilities; interface-type capabilities describe existing interfaces with endpoint references, request field mappings, response field mappings, and error categories; interface-type capabilities describe authorized interfaces with target control semantic names, accessibility attributes, adjacent tags, and page structure summaries; each node in the versioned task node diagram records dependent nodes, conditional branches, bound capability contracts, parameter sources, expected postconditions, risk levels, and compensation nodes.

[0009] Preferably, parameter validation is performed on the nodes in the versioned task node graph, including: Obtain the parameters declared by the node from the task envelope, existing execution receipts, and authorized data sources. Check the type, format, validity period, source, and operator permissions of each parameter to obtain the parameter integrity rate. When the parameter integrity rate of a read-only node is not lower than the first integrity rate threshold, it enters the execution preparation stage; when the parameter integrity rate of a write node is not lower than the second integrity rate threshold, it enters the pre-run stage; when it is lower than the corresponding threshold, it enters the waiting for supplementation stage. The verified node is converted into a node execution request. The node execution request carries the node identifier, plan version, capability contract identifier, verified parameters, authorization ticket reference, execution time limit and expected postconditions.

[0010] Preferably, calling an existing interface or operation authorization interface includes: For interface-based capabilities, call existing interfaces according to the interface field mapping; For interface-type capabilities, the interface location matching degree is obtained by the degree of consistency between the current control characteristics and the stable characteristics existing in the capability contract. The interface location matching degree is obtained by weighted summation of semantic name value, accessibility attribute value, adjacent tag value and page structure summary value. If the interface positioning matching degree is not lower than the first positioning threshold, a pre-run is allowed. If it is lower than the first positioning threshold but not lower than the second positioning threshold, another set of adjacent labels and accessibility attributes are used for repositioning. The operation will continue only if the two positioning points to the same business control. If it is lower than the second positioning threshold, the interface operation will be stopped and an adaptation update pending review will be formed.

[0011] Preferably, nodes with write, submit, approve, reject, or cross-domain transmission effects first generate a pre-rehearsal result. The pre-rehearsal result specifies the business object, the field to be changed, the target government business system, and the expected post-conditions. Only when the authorization conditions and gating status are met will the pre-rehearsal be converted to submission. If not, the node remains in a state of pending manual intervention or prohibited execution.

[0012] Preferably, feedback replanning is performed based on the execution receipt, including: The postcondition observations in the execution receipt are compared with the expected postconditions in the capability contract to obtain the postcondition consistency. The node is completed when the postcondition consistency is not lower than the first consistency threshold and there are no prohibitive errors. When it is lower than the first consistency threshold but not lower than the second consistency threshold, a status review is performed. When it is lower than the second consistency threshold, the release of the successor node is prohibited and compensation or manual takeover is selected. Recoverable anomalies keep completed nodes frozen, thereby replacing affected unexecuted nodes and forming incremental plan versions; when an action has partially taken effect but cannot be safely continued, a compensation request carrying the original business primary key is formed according to the capability contract.

[0013] Preferably, compliance gates are implemented based on policy evidence and task risks, including: Extract policy references, dates, application conditions, entity names, and proposed actions from the text to be output, and identify the content that can be verified by external rules as verifiable claims; Verify verifiable claims according to policy document version, effective period, applicable objects, and paragraph position to form a compliance evidence package; The evidence coverage rate is obtained by calculating the proportion of verifiable claims that are bound to valid policy versions and paragraph positions out of all verifiable claims. The results of the checks on evidence validity, rule consistency, format completeness, and sensitive information are scored separately and then summarized to obtain a compliance confidence score. Based on evidence coverage, compliance confidence score, and task risk level, a gating status is established that allows output, re-retrieval, awaits manual intervention, or prohibits execution.

[0014] Preferably, compliance gating also includes: When there are multiple versions of the same policy, the valid versions are filtered by the date the task occurred, the applicable region, and the applicable target. If there are still conflicting valid versions, the case is transferred to manual processing. Text in user attachments, search snippets, or target pages that requires ignoring authorization constraints is treated only as business data. When a prompt for injection is detected, the relevant node remains in a state awaiting manual intervention, and only the trigger snippet summary is saved in the audit log.

[0015] Preferably, the final execution receipt is verified, and the task-level audit package is solidified, including: After the final execution receipt is verified, the task envelope, capability contract version, plan version, execution receipt, compliance evidence, and manual confirmation record are solidified in chronological order to form a task-level audit package. The system enters the completion state when the final postcondition is met, and enters the failure state when compensation fails or an unrecoverable error occurs. Both the completion and failure states are terminated states, and new processing requirements must be generated into a new task envelope.

[0016] This invention also provides a low-intrusion intelligent agent construction system for government offices, including a control plane and an execution plane; the control plane includes a task envelope construction module, a task planning module, a parameter verification module, a feedback replanning module, a compliance gating module, and a controlled audit module; the execution plane includes a low-intrusion execution module; The task envelope construction module is used to obtain the user's natural language instructions, initiator identity reference, authorized role, attachment reference and target security domain, and replace the credentials with credential references to form a task envelope; The task planning module is used to retrieve the capability contract registry based on the task envelope, select the capability contract that matches the business object, operation type, read / write attributes and security domain, and generate a versioned task node graph through dependency analysis. The parameter validation module is used to validate the parameters of nodes in the versioned task node graph and generate node execution requests. The low-intrusion execution module is used to call existing interfaces or operation authorization interfaces and generate execution receipts within the security domain of the government business system, based on node execution requests and capability contracts. The feedback replanning module is used to perform feedback replanning based on execution receipts, freeze successful nodes that meet the expected postconditions, replace affected unexecuted nodes, and handle compensation. The compliance gating module is used to perform compliance gating based on policy evidence and task risks, and generate a one-time execution permit; The controlled audit module is used to verify the final execution receipt and solidify the task-level audit package.

[0017] Beneficial effects: 1. This invention uses capability contracts to uniformly express interface calls and authorized interface operations as business capabilities, enabling task plans generated by large language models to uniformly call heterogeneous interfaces or authorized interfaces; when adding a new government business system, only field mapping, control features and postconditions need to be configured on the execution plane, and the control plane does not need to contact the target database or change the original approval logic, thereby reducing the impact of underlying transformation on the stability of the existing system and achieving low-intrusion construction.

[0018] 2. This invention binds parameters to each node based on a versioned task node graph, and uses execution receipts to verify business primary keys, return status, and postconditions, so that page drift, return format changes, and response loss can be identified before actual submission; successful nodes are frozen, and only affected unexecuted nodes are replaced, preserving the business facts that have already occurred, avoiding replay of the entire process, thereby reducing the risk of duplicate writing and process interruption, and forming a recoverable processing process.

[0019] 3. This invention utilizes policy document versions, effective periods, and paragraph positions to constrain verifiable claims. Combined with risk levels and manual confirmation for issuing execution permits, it separates the model-generated content from administrative decisions and ensures that the policy versions and authorization conditions used at execution nodes are synchronized with the actual business status. Task envelopes, capability contract versions, plan versions, execution receipts, policy evidence, and manual confirmation records are solidified over time to form an audit package, thereby improving the traceability of result sources and anomaly recovery paths and reducing the consequences of invalid policy references and operational basis being difficult to restore. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system module diagram of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely for explaining the invention and are not intended to limit the scope of protection of the invention.

[0022] Example 1: Please refer to Figure 1 As shown in the figure, this embodiment provides a low-intrusion method for constructing a government office intelligent agent, which includes the following steps: S1. Construct task envelopes and generate versioned task node graphs based on capability contracts; The system retrieves the user's natural language instructions, initiator's identity reference, authorized role, attachment reference, and target security domain. It then replaces the credentials themselves with credential references that can be parsed within the authorized environment, forming a task envelope. This task envelope serves as the boundary for a single government office task. It distinguishes tasks and business actions using a task identifier and a task idempotent key, and records the target matter, authorized scope, attachment location, and input summary. These fields are subsequently read during capability selection, parameter validation, and duplicate write checks.

[0023] When a task envelope enters the capability selection, it retrieves the capability contract registry by business object, operation type, read / write attribute, and security domain to obtain the capability contract that matches the current task. The capability contract provides a machine-readable description of an executable business capability, recording input field types, preconditions, execution method, expected postconditions, idempotent fields, compensation actions, and security domains; authentication information only stores references and not the password or token plaintext.

[0024] The control plane is deployed in a security domain that allows the operation of large language models and orchestration services. A capability contract registry is set up in the control plane and stores configuration versions of interface-type and UI-type capabilities. This capability contract registry does not replace the processing functions of the task planning module; it is only used to provide capability contract configuration. The task planning module connects to the capability contract registry to select capability contracts that match the business object, security domain, and read / write attributes based on the task envelope.

[0025] Capability contracts are categorized into interface-type capabilities and UI-type capabilities based on their execution method. Interface-type capabilities describe existing interfaces using endpoint references, request field mappings, response field mappings, and error categories; endpoint references are resolved within the deployment environment. UI-type capabilities describe authorized interfaces using the semantic names of target controls, accessibility attributes, adjacent tags, and page structure summaries; the page structure summaries consist of stable features in the allowed document object structure, accessibility tree, or control tree.

[0026] The selected capability contracts and task objectives, after dependency analysis, form a versioned task node graph. This versioned task node graph represents a set of nodes with planned versions. Each node records dependent nodes, conditional branches, bound capability contracts, parameter sources, expected postconditions, risk levels, and compensation nodes. When checks reveal parameters without sources, dependency loops, or cross-domain disabled actions, the current task remains in a pending planning state, and conditions are required to be completed.

[0027] Preferably, the planning state exits when missing parameter sources, dependencies, or cross-domain authorization conditions are filled in and pass the checks, and restarts dependency analysis from the versioned task node graph; if the checks still fail, it continues to wait until the deadline recorded in the task envelope or the authorization ticket expires, at which point automatic planning terminates and switches to manual processing. If the filled content changes the task objective or security domain, an incremental plan version is generated before dependency checks are performed. The large language model can only generate node suggestions through the task planning module; it cannot bypass subsequent parameter validation to directly trigger low-intrusion execution, nor can it bypass compliance gating and controlled auditing to complete high-risk submissions.

[0028] S2. Perform parameter validation on the nodes in the versioned task node graph and generate node execution requests.

[0029] After the versioned task node graph completes the initial check, it releases nodes whose dependencies have been satisfied and keeps the remaining nodes in a dependency waiting state.

[0030] During parameter validation, the parameter validation module connects to the task planning module and the authorized data source. It obtains the current node and its parameter source from the versioned task node graph output by the task planning module, and retrieves the declared parameters from the task envelope, existing execution receipts, and the authorized data source. For each parameter, it checks its type, format, validity period, source, and operator permissions to obtain the parameter completeness rate. The parameter completeness rate represents the percentage of required fields that have passed type and source validation out of the total number of required fields, expressed as a percentage. Read-only nodes enter execution preparation when their parameter completeness rate is not lower than the first completeness rate threshold; write nodes enter pre-execution when their parameter completeness rate is not lower than the second completeness rate threshold; and enter a waiting-to-be-supplemented state when their parameter completeness rate is lower than the corresponding threshold.

[0031] Preferably, the first integrity rate threshold is 95%, and the second integrity rate threshold is 100%.

[0032] Among them, the authorized data source refers to the existing data source that is allowed to be accessed by the parameter verification module after the authorized role, operator permissions and target security domain recorded in the task envelope have been verified. It includes, but is not limited to, electronic certificate database, basic information database, item material database, historical case database, approval result database, and authorized query interface opened by the data sharing and exchange platform.

[0033] Verified nodes are converted into node execution requests, which carry node identifiers, plan versions, capability contract identifiers, verified parameters, authorization ticket references, execution time limits, and expected postconditions. Preferably, the node execution time limit is 120 seconds. The plan version starts from the initial version and increments when execution feedback changes unexecuted paths; nodes that have been successfully verified retain their original business primary keys and receipt summaries, and parameters that have been manually confirmed are not modified due to replanning.

[0034] S3. Perform low-intrusion execution on nodes within the security domain of the government business system and generate execution receipts.

[0035] The execution plane is deployed in the security domain of each government business system. The low-intrusion execution module is located in the execution plane and connected to the interface endpoints or authorized interfaces that are authorized to be accessed within the security domain. The control plane and the execution plane transmit task requests, one-time execution licenses and execution receipts through the existing approved exchange channels. The control plane does not directly connect to the target business database. When the original government data is not allowed to leave the domain, it only returns the field existence, the de-identified business status and the receipt summary.

[0036] The execution plane selects an interface call or authorized interface operation based on the capability contract. The reading action calls an existing interface according to the interface field mapping, or locates the authorized interface according to the stable characteristics recorded by the interface-type capability. Then, the actual execution method, de-identification result and business status are written into the execution receipt.

[0037] For interface-type capabilities, the interface positioning matching degree is obtained by the consistency between the current control characteristics and the stable characteristics stored in the contract. The interface positioning matching degree is used to determine the impact of page drift on the operation target. Specifically, the target control semantic name is set to 1 if it is the same as the capability contract record after standardization or hits the specification mapping, otherwise it is set to 0; the accessibility attribute is set to 1 if the role, control type, and read / write status are all the same as the capability contract record, otherwise it is set to 0; the neighboring tags are set to 1 if they are the same as the capability contract record after standardization, otherwise it is set to 0; the page structure summary is the number of contract stable characteristics hit around the current control divided by the total number of stable characteristics in the capability contract record, with a value range of 0 to 1; when the total number of stable characteristics is 0, the current interface positioning is directly judged as a mismatch and the operation is stopped.

[0038] The interface location matching score is calculated by multiplying the semantic name value by 0.35, the accessibility attribute value by 0.30, the neighboring tag value by 0.20, and the page structure summary value by 0.15, and then summing the results. The summation result is limited to 0 to 1. This result is then compared to a first location threshold and a second location threshold, and corresponding pre-run, secondary location, or stop operation decisions are made. If the interface location matching score is not lower than the first location threshold, one pre-run is allowed. If it is lower than the first location threshold but not lower than the second location threshold, another set of neighboring tags and accessibility attributes is used for re-location. Relocation continues only if both locations point to the same business control. If it is lower than the second location threshold, interface operation is stopped, and an adaptation update awaiting review is generated. Preferably, the first location threshold is 0.92, and the second location threshold is 0.80.

[0039] The adaptation update pending review will exit when the reviewer approves or rejects the version; if approved, the new capability contract version will be frozen and restarted from the interface location; if rejected, the original version will be retained and the update will be put into manual review; if there is still no review conclusion by the task deadline, the update will also be put into manual review.

[0040] Nodes with write, submit, approve, reject, or cross-domain sending effects first generate a pre-rehearsal result. The pre-rehearsal result specifies the business object, the field to be changed, the target government business system, and the expected post-conditions. The pre-rehearsal is only converted to submission when the authorization conditions and gating status are both met; otherwise, it remains in a state of waiting for manual intervention or prohibiting execution, and no final write request is sent to the target government business system.

[0041] S4. Based on the execution receipt, perform feedback replanning, freeze successful nodes, and process compensation; After a tool action is completed, an execution receipt is obtained, containing node status, business primary key, anonymized output, postcondition observations, error type, start and end times, and a receipt summary. The execution receipt associates a business action with the task idempotency key, node identifier, target system identifier, and business object primary key. The receipt summary is used to verify record integrity; the business primary key, node status, and postcondition observations are actually read during duplicate checks and result verification.

[0042] After the execution receipt is generated, the capability contract used in the actual execution is checked against the capability contract bound to the current node according to the capability contract identifier. If the two are consistent, the execution receipt is retained and the post-condition verification is initiated. If the two are inconsistent, the current node is kept in the pending verification state and the subsequent node is not released.

[0043] After the execution receipt is returned, the action status is first queried using the task idempotent key and the business object primary key. If a successful record is found and the receipt summary matches, the existing receipt is accepted directly. If a record in progress is found, the status is kept pending verification. Only if the query confirmation is not effective is a retry allowed within the agreed number of attempts. If the query result is still uncertain, the system enters a pending manual state and remains in a write paused state. The agreed number of attempts can be 1 to 3; preferably, 1 retry is allowed.

[0044] The pending verification status exits upon obtaining a new execution receipt consistent with the frozen capability contract identifier, or upon determining the action status using the task idempotency key and the business object primary key. If it is determined to be effective, verification restarts from the post-condition verification stage; if it is determined to be ineffective, it restarts from the limited retries of the node execution request. If it remains in progress, it remains pending verification. After the node execution time limit is reached, it queries only once more. If the status is still in the pending verification range, it transitions to manual verification and terminates automatic retries. If the actual duration exceeds the node execution time limit, it first queries the business status corresponding to the task idempotency key. If it is confirmed to be ineffective, it performs limited retries. If it is confirmed to be effective, an effective receipt is generated. If the confirmation result is still in an uncertain state, it transitions to manual verification.

[0045] The postcondition observations in the execution receipt are compared with the expected postconditions in the capability contract to obtain the postcondition consistency degree. The postcondition consistency degree represents the proportion of satisfied expected postconditions to the total number of expected postconditions. A node completes when the postcondition consistency degree is not lower than a first consistency threshold and there are no prohibitive errors. If it is lower than the first consistency threshold but not lower than a second consistency threshold, only a status review is performed. If it is lower than the second consistency threshold, releasing the successor node is prohibited, and compensation or manual takeover is selected. Preferably, the first consistency threshold is 0.9, and the second consistency threshold is 0.7.

[0046] The status review exits when the target government business system returns a new business status or a new postcondition observation value; when the postcondition consistency reaches the first consistency threshold, it is released from the successor node and restarts; when it is below the second consistency threshold, it restarts from the compensation or manual takeover judgment; when it is between the two thresholds, the status review is maintained; after the node execution time limit is reached, a second review is conducted and if it still does not exceed the range, it is switched to waiting for manual intervention.

[0047] During feedback replanning, the feedback replanning module connects the low-intrusion execution module and the task planning module, receives the execution receipt returned by the low-intrusion execution module, reads the task idempotent key, business primary key, receipt summary and postcondition observation, forms the next batch of valid nodes, new plan version, compensation request or manual pending task, and returns the replanning results to the task planning module.

[0048] An execution receipt that meets the expected postconditions transitions the current node to a completed state and releases subsequent nodes whose dependencies have been met. Recoverable exceptions keep completed nodes frozen, thereby replacing affected unexecuted nodes and creating incremental plan versions. When an action has partially taken effect but cannot be safely continued, a compensation request carrying the original business primary key is generated according to the capability contract. Irreversible submissions, approvals, or outsourcing actions are converted to manual pending tasks.

[0049] After a compensation request is generated, the execution agent within the target security domain where the original node resides invokes its specified compensation action according to the frozen capability contract version. Before invoking, it checks whether compensation has already been performed using the task idempotent key, the original node identifier, and the original business primary key. The compensation request only references the original business object and does not create a new business object. After the compensation action is completed, a compensation receipt is generated, recording the task idempotent key, the original node identifier, the original business primary key, the capability contract identifier and version, the compensation status, and the postcondition observations. It also includes the error type, start and end times, and a receipt summary, and is retained alongside the original execution receipt.

[0050] When compensation is successful and the observed values ​​meet the expected postconditions recorded in the capability contract, the compensated nodes are frozen and parameter binding restarts from the affected unexecuted nodes. If compensation fails, the node execution time limit is exceeded, or the business status is uncertain, the process transitions to manual intervention. Unrecoverable errors cause the task to enter a failed state and terminate automatic execution. Before re-executing the plan version, the parameter sources and authorization conditions of newly added nodes are re-checked. Successful nodes unrelated to anomalies continue to reference the original execution receipts, while unexecuted nodes and draft content with changed dependencies are marked as invalid. Based on this, parameter binding is only performed on valid nodes in the new plan version to avoid mixing old parameters with new results.

[0051] S5. Conduct compliance gatekeeping and generate execution permits based on policy evidence and task risks; After a regular business node is completed, policy references, dates, processing conditions, entity names, and proposed actions are extracted from the text to be output, and the content that can be verified by external rules is identified as verifiable claims.

[0052] During compliance gating, the compliance gating module connects to the compliance knowledge base and verifies verifiable claims from the knowledge base based on policy document version, effective period, applicable objects, and paragraph position. The compliance knowledge base is set up on the control plane and stores policy document versions, effective periods, and rules, but does not replace the processing functions of the compliance gating module. Each verifiable claim is bound to a policy document identifier, version, effective period, applicable objects, and paragraph position, and together with the search segment summary, rule hit results, risk level, and manual confirmation record, constitutes a compliance evidence package. The aforementioned policy version and paragraph position are read during gating determination and audit solidification.

[0053] The evidence coverage rate is the percentage of verifiable claims in the compliance evidence package that are bound to valid policy versions and paragraph positions out of all verifiable claims. When the evidence coverage rate is not lower than the evidence coverage rate threshold and there are no rule conflicts, the content enters content gating. If it falls below the threshold, a new search is conducted. If the rate remains below the threshold after consecutive searches, the process is transferred to manual verification, and content lacking evidence is not added as fact. The evidence coverage rate threshold can be between 95% and 100%, with the upper limit of this range used for tasks involving policy clauses, processing conditions, or time limits; preferably, the evidence coverage rate threshold is 100%.

[0054] The results of the checks on evidence validity, rule consistency, format integrity, and sensitive information are each assigned a score and then aggregated to obtain a compliance confidence score. The compliance confidence score is an integer value from 0 to 100, representing the degree to which the current content meets government regulations and output constraints. Specifically, the score for evidence validity is determined by the weight of the evidence validity component, the maximum compliance confidence score, and the proportion of evidence with valid versions and paragraph positions bound to the total number of examined evidence. The score for rule consistency is determined by the weight of the rule consistency component, the maximum compliance confidence score, and the proportion of the number of applicable rules without conflict to the total number of applicable rules. The score for format integrity is determined by the weight of the format integrity component, the maximum compliance confidence score, and the proportion of the number of required format items to the total number of required format items. When no unauthorized disclosures are found during the sensitive information check, the score for the sensitive information check is the product of the sensitive information check component weight and the maximum compliance confidence score; if any unauthorized disclosure is found, the score is 0 and the content is immediately transferred to manual review. The total weight of each component is 100%. For example, when the weighting of the evidence validity component is 40%, the rule consistency component is 30%, the format integrity component is 15%, and the sensitive information check component is 15%, the evidence validity component scores 0 to 40 points, the rule consistency component scores 0 to 30 points, the format integrity component scores 0 to 15 points, and the sensitive information check scores 15 points if no unauthorized disclosures are found, and 0 points if any unauthorized disclosures are found, and the result is immediately transferred to manual review. The total score is initially limited to 0 to 100 points, and then rounded down to the nearest integer if the decimal part is not less than 0.5. Preferably, the weighting of the evidence validity component can be 40% to 50%, the rule consistency component can be 25% to 35%, the format integrity component can be 10% to 20%, and the sensitive information check component can be 10% to 20%, with a total weighting of 100%.

[0055] Preferably, the weight of the evidence validity component can be 40% to 50%, the weight of the rule consistency component can be 25% to 35%, the weight of the format integrity component can be 10% to 20%, and the weight of the sensitive information check component can be 10% to 20%, with the total weight of each component being 100%.

[0056] A read-only response can be generated when it reaches the first confidence score threshold. A draft document or approval opinion can only be generated as an editable draft when it reaches the second confidence score threshold. Responses below the corresponding thresholds require re-retrieval or manual processing. Preferably, the first confidence score threshold is 90 points, and the second confidence score threshold is 95 points.

[0057] The gating status determines subsequent actions based on task risk: read-only queries can output data when evidence and format conditions are met; high-risk drafts remain editable; submission, approval, rejection, and cross-domain sending actions all enter a pending manual state regardless of the compliance confidence score; after manual confirmation, a one-time execution permit is generated, bound to the task idempotent key, node identifier, capability contract version, and pre-rehearsal result summary. The one-time execution permit moves the corresponding high-risk node from the pending controlled submission state to the execution state, and the final execution receipt after submission is still verified according to post-conditions.

[0058] When multiple versions of the same policy exist, the valid versions are filtered by the date of the task, the applicable region, and the applicable object. If there are still conflicting valid versions, the case is directly transferred to manual processing. Text in user attachments, search fragments, or target pages that requires ignoring authorization constraints is only processed as business data. When a prompt for injection is detected, the relevant node is kept in a state awaiting manual processing, and only the trigger fragment summary is saved in the audit record.

[0059] S6. Verify the final execution receipt and solidify the task-level audit package; During controlled auditing, the controlled audit module connects the compliance gating module, the low-intrusion execution module, and the audit repository. It obtains the gating status from the compliance gating module, the final execution receipt from the low-intrusion execution module, and binds the one-time execution permission to the task idempotent key, node identifier, capability contract version, and pre-rehearsal result summary. The audit repository is set up on the control plane and stores object references and summaries, but does not replace the processing functions of the controlled audit module.

[0060] After the final execution receipt is verified, the task envelope, capability contract version, plan version, execution receipt, compliance evidence, and manual confirmation record are solidified in chronological order to form a task-level audit package. This task-level audit package is a read-only collection of records that can locate object references and summaries at each stage based on the task identifier. The final output includes query results, editable drafts, manually confirmed operation results, or reasons for failure. Administrative decisions are still made by authorized personnel.

[0061] The system enters the completion state when the final postconditions are met, and enters the failure state when compensation fails or an unrecoverable error occurs. Both the completion and failure states are termination states, and new processing requirements must form a new task envelope; the model cannot resume execution on its own.

[0062] The control plane is deployed in a security domain that allows the operation of large language models and orchestration services, while the execution plane is deployed in the security domain where each government business system is located. The two transmit task requests, one-time execution licenses, and execution receipts through existing approved exchange channels. The control plane does not directly connect to the target business database. When the original government data is not allowed to leave the domain, it only returns the existence of fields, the de-identified business status, and a receipt summary.

[0063] Example 2: Please refer to Figure 2 As shown, this embodiment provides a low-intrusion intelligent agent construction system for government offices, including a control plane and an execution plane. The control plane is deployed in a security domain that allows the operation of large language models and orchestration services, while the execution plane is deployed in the security domain where each government business system is located. The two transmit task requests, one-time execution permissions, and execution receipts through existing approved exchange channels. The control plane does not directly connect to the target business database; when the original government data is not allowed to leave the domain, it only returns field existence, anonymized business status, and a receipt summary. The control plane includes a task planning module, a parameter validation module, a feedback replanning module, a compliance gating module, and a controlled audit module; the execution plane includes a low-intrusion execution module.

[0064] The task planning module is used to construct a versioned task node graph based on capability contracts. This module receives task envelopes, connects to the capability contract registry to select capability contracts that match the business object, security domain, and read / write attributes, and then writes node dependencies, conditional branches, parameter sources, and planned versions into the versioned task node graph.

[0065] The parameter validation module is used to bind and validate node parameters. This module connects to the task planning module and the authorized data source, reads the parameter source from the task envelope and the previous execution receipt, and forms a validated node or a node awaiting supplementation status based on field type, completeness, validity period, and permissions.

[0066] The low-intrusion execution module is used to call existing interfaces or authorized interfaces based on verified nodes and generate execution receipts. This module is located on the execution plane and connects to the interface endpoints or authorized interfaces that are authorized to access within this security domain. It forms node execution requests according to the capability contract. Nodes with high-risk write effects first generate a pre-rehearsal result and obtain a one-time execution permission before submitting the final action.

[0067] The feedback replanning module is used to freeze successful nodes and replace affected unexecuted nodes based on execution receipts. This module connects the low-intrusion execution module and the task planning module, reads the task idempotent key, business primary key, receipt summary, and postcondition observations, and generates the next batch of valid nodes, new plan versions, compensation requests, or manual pending tasks.

[0068] The compliance gating module is used to establish gating statuses based on policy evidence and task risks. This module connects to the compliance knowledge base, verifies verifiable claims according to policy document version, effective period, applicable objects, and paragraph position, and determines the status of allowed output, re-retrieval, awaiting manual intervention, or prohibited execution based on the risk level.

[0069] The controlled audit module is used to issue execution licenses, verify final receipts, and generate audit packages after gating authorization. This module connects the compliance gating module, the low-intrusion execution module, and the audit repository, binding one-time execution licenses to task idempotent keys, node identifiers, capability contract versions, and pre-rehearsal result summaries, and solidifying references to objects at each stage after the final receipt meets the post-conditions.

[0070] The control plane also includes a capability contract registry, a compliance knowledge base, and an audit repository. The capability contract registry stores configuration versions of interface-type and UI-type capabilities; the compliance knowledge base stores policy document versions, effective periods, and rules; and the audit repository stores object references and summaries. This storage does not replace the processing functions of the corresponding modules. The large language model can only generate node suggestions through the task planning module; it cannot bypass the parameter validation module to directly trigger the low-intrusion execution module, nor can it bypass the compliance gating module and the controlled audit module to complete high-risk submissions.

[0071] The versioned task node graph generated by the task planning module enters the parameter verification module; verified nodes reach the low-intrusion execution module and generate an execution receipt; after the execution receipt is returned to the replanning module, successful nodes retain the original plan, and affected unexecuted nodes re-enter the parameter verification module through the new plan version; ordinary business results reach the compliance gating module, and when the gating status meets the risk requirements, the controlled audit module issues a one-time execution permit, and finally, after the execution receipt is verified, a task-level audit package is formed.

[0072] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for constructing a low-intrusion intelligent agent for government office work, characterized in that, include: Obtain the user's natural language instructions, initiator identity reference, authorized role, attachment reference, and target security domain, and replace the credentials with credential references to form a task envelope; Based on the task envelope retrieval capability contract registry, select the capability contract that matches the business object, operation type, read / write attributes, and security domain, and generate a versioned task node graph through dependency analysis; Perform parameter validation on the nodes in the versioned task node graph and generate node execution requests; Within the security domain of the government business system, based on the node execution request and capability contract, existing interfaces or operation authorization interfaces are invoked to generate execution receipts; Based on the execution receipt, feedback replanning is performed, successful nodes that meet the expected post-conditions are frozen, affected unexecuted nodes are replaced, and compensation is processed. Compliance gates are implemented based on policy evidence and task risks to generate a one-time execution permit; Verify the final execution receipt and solidify the task-level audit package.

2. The method for constructing a low-intrusion government office intelligent agent according to claim 1, characterized in that, Capability contracts record input field types, preconditions, execution methods, expected postconditions, idempotent fields, compensation actions, and security domains; capability contracts are divided into interface-type capabilities and UI-type capabilities; interface-type capabilities describe existing interfaces using endpoint references, request field mappings, response field mappings, and error categories; Interface-type capabilities describe the authorized interface using the semantic name of the target control, accessibility attributes, adjacent tags, and a summary of the page structure; each node in the versioned task node diagram records dependent nodes, conditional branches, bound capability contracts, parameter sources, expected postconditions, risk levels, and compensation nodes.

3. The method for constructing a low-intrusion government office intelligent agent according to claim 1, characterized in that, Perform parameter validation on nodes in the versioned task node graph, including: Obtain the parameters declared by the node from the task envelope, existing execution receipts, and authorized data sources. Check the type, format, validity period, source, and operator permissions of each parameter to obtain the parameter integrity rate. When the parameter integrity rate of a read-only node is not lower than the first integrity rate threshold, it enters the execution preparation stage; when the parameter integrity rate of a write node is not lower than the second integrity rate threshold, it enters the pre-run stage; when it is lower than the corresponding threshold, it enters the waiting for supplementation stage. The verified node is converted into a node execution request. The node execution request carries the node identifier, plan version, capability contract identifier, verified parameters, authorization ticket reference, execution time limit and expected postconditions.

4. The method for constructing a low-intrusion government office intelligent agent according to claim 1, characterized in that, Calling existing interfaces or operation authorization interfaces, including: For interface-based capabilities, call existing interfaces according to the interface field mapping; For interface-type capabilities, the interface location matching degree is obtained by the degree of consistency between the current control characteristics and the stable characteristics existing in the capability contract. The interface location matching degree is obtained by weighted summation of semantic name value, accessibility attribute value, adjacent tag value and page structure summary value. If the interface positioning matching degree is not lower than the first positioning threshold, a pre-run is allowed. If it is lower than the first positioning threshold but not lower than the second positioning threshold, another set of adjacent labels and accessibility attributes are used for repositioning. The operation will continue only if the two positioning points to the same business control. If it is lower than the second positioning threshold, the interface operation will be stopped and an adaptation update pending review will be formed.

5. The method for constructing a low-intrusion government office intelligent agent according to claim 1, characterized in that, Nodes with write, submit, approve, reject, or cross-domain sending capabilities first generate a pre-rehearsal result. The pre-rehearsal result specifies the business object, the field to be changed, the target government business system, and the expected post-conditions. Only when the authorization conditions and gating status are both met will the pre-rehearsal be converted to submission; otherwise, it will remain in a state of pending manual intervention or prohibited execution.

6. The method for constructing a low-intrusion government office intelligent agent according to claim 1, characterized in that, Based on the execution feedback, feedback and replanning are carried out, including: The postcondition observations in the execution receipt are compared with the expected postconditions in the capability contract to obtain the postcondition consistency. The node completes when the postcondition consistency is not lower than the first consistency threshold and there are no prohibitive errors; when it is lower than the first consistency threshold but not lower than the second consistency threshold, a status review is performed; when it is lower than the second consistency threshold, the release of the successor node is prohibited and compensation or manual takeover is selected. Recoverable anomalies keep completed nodes frozen, thereby replacing affected unexecuted nodes and forming incremental plan versions; when an action has partially taken effect but cannot be safely continued, a compensation request carrying the original business primary key is formed according to the capability contract.

7. The method for constructing a low-intrusion government office intelligent agent according to claim 1, characterized in that, Compliance gates are implemented based on policy evidence and task risks, including: Extract policy references, dates, application conditions, entity names, and proposed actions from the text to be output, and identify the content that can be verified by external rules as verifiable claims; Verify verifiable claims according to policy document version, effective period, applicable objects, and paragraph position to form a compliance evidence package; The evidence coverage rate is obtained by calculating the proportion of verifiable claims that are bound to valid policy versions and paragraph positions out of all verifiable claims. The results of the checks on evidence validity, rule consistency, format completeness, and sensitive information are scored separately and then summarized to obtain a compliance confidence score. Based on evidence coverage, compliance confidence score, and task risk level, a gating state is formed that allows output, re-retrieval, awaits manual intervention, or prohibits execution.

8. The method for constructing a low-intrusion government office intelligent agent according to claim 7, characterized in that, Compliance gating also includes: When there are multiple versions of the same policy, the valid versions are filtered by the date the task occurred, the applicable region, and the applicable target. If there are still conflicting valid versions, the case is transferred to manual processing. Text in user attachments, search snippets, or target pages that requires ignoring authorization constraints is treated only as business data. When a prompt for injection is detected, the relevant node remains in a state awaiting manual intervention, and only the trigger snippet summary is saved in the audit log.

9. The method for constructing a low-intrusion government office intelligent agent according to claim 1, characterized in that, Verify the final execution receipt and solidify the task-level audit package, including: After the final execution receipt is verified, the task envelope, capability contract version, plan version, execution receipt, compliance evidence, and manual confirmation record are solidified in chronological order to form a task-level audit package. The system enters the completion state when the final postcondition is met, and enters the failure state when compensation fails or an unrecoverable error occurs. Both the completion and failure states are terminated states, and new processing requirements must be generated into a new task envelope.

10. A low-intrusion intelligent agent construction system for government offices, characterized in that, It includes a control plane and an execution plane; the control plane includes a task envelope construction module, a task planning module, a parameter validation module, a feedback replanning module, a compliance gating module, and a controlled audit module; the execution plane includes a low-intrusion execution module. The task envelope construction module is used to obtain the user's natural language instructions, initiator identity reference, authorized role, attachment reference and target security domain, and replace the credentials with credential references to form a task envelope; The task planning module is used to retrieve the capability contract registry based on the task envelope, select the capability contract that matches the business object, operation type, read / write attributes and security domain, and generate a versioned task node graph through dependency analysis. The parameter validation module is used to validate the parameters of nodes in the versioned task node graph and generate node execution requests. The low-intrusion execution module is used to call existing interfaces or operation authorization interfaces and generate execution receipts within the security domain of the government business system, based on node execution requests and capability contracts. The feedback replanning module is used to perform feedback replanning based on execution receipts, freeze successful nodes that meet the expected postconditions, replace affected unexecuted nodes, and handle compensation. The compliance gating module is used to perform compliance gating based on policy evidence and task risks, and generate a one-time execution permit; The controlled audit module is used to verify the final execution receipt and solidify the task-level audit package.