A Strategy-Constrained Routing and Result Verification Method for Multi-Execution Ends of Scientific Research Tasks

By constructing runtime strategy objects and capability declaration objects, and combining hard constraint admission and soft selection indicators, the problem of inconsistent call paths at the execution end in scientific research task processing was solved, achieving consistency between front-end and back-end states and accurate handling of data boundaries, thereby improving task recoverability and execution accuracy.

CN123117749APending Publication Date: 2026-11-17SOUTHWEAT UNIV OF SCI & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611302768.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-26
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In scientific research task processing, existing technologies have difficulty maintaining consistent call constraints and data boundaries during cross-stage execution, making it difficult to ensure that the front-end configuration and the actual back-end call path are consistent. This results in inconsistencies between the interface state and the back-end state, and makes it difficult to ensure that the execution end's capabilities and data processing meet the task requirements.

Method used

The system constructs operational strategy objects, establishes capability declaration objects and scientific research asset objects, generates a set of allowed execution paths through hard constraint admission judgments and soft selection indicators, verifies the consistency between expected execution descriptions and actual execution descriptions, and ensures the accuracy and consistency of task execution through chained audit records.

Benefits of technology

It achieves consistency in the correspondence between front-end configuration and back-end call path in scientific research task processing, prevents automatic switching to unauthorized backup endpoints when the execution end is insufficient, ensures the accuracy and consistency of data boundary processing, reduces execution deviation, and improves task recoverability and state consistency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123117749A_ABST
    Figure CN123117749A_ABST
Patent Text Reader

Abstract

This application relates to a strategy-constrained routing and result verification method for multi-execution endpoints in scientific research tasks. The method involves: constructing a runtime strategy object; creating a capability declaration object for each callable execution endpoint; breaking down the scientific research task into multiple processing stages and generating stage requirement objects; representing task input data as scientific research asset objects; performing hard-constraint admission judgments on each stage and each candidate execution endpoint, generating a set of allowed execution paths, and disabling automatic endpoint switching when the set is empty; generating an outbound list to control the outbound field when the call leaves the controlled boundary; generating expected execution descriptions and actual execution descriptions before and after the call, respectively, and performing consistency checks to determine whether the result is written into the formal workflow or an isolated area; and connecting key events using pre-order hashes to form a chain of audit records. This ensures that the execution path of the scientific research task is consistent with the user-confirmed runtime strategy, preventing unauthorized outbound actions and misuse of backup endpoints.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of computer data access control and task scheduling technology, and in particular to a method for policy constraint routing and result verification for multiple execution ends of scientific research tasks. Background Technology

[0002] Scientific research task processing systems typically require access to project knowledge bases, text processing models, vector retrieval services, academic literature retrieval tools, and document parsing programs. For tasks that include project documents, research task cards, search terms, and intermediate results, in addition to the result content, it is also necessary to determine which execution terminal will handle each stage, whether external networks are accessed, which fields can leave the project boundary, and whether the established operating mode can be changed if the execution terminal is unavailable.

[0003] Existing first-type technologies typically set options such as "local," "cloud," or "network connection required" on the front end and pass the corresponding flags to the back end when the task starts. While these settings represent user choices, they are often only used as interface states or single configuration items, without establishing a corresponding relationship with specific execution endpoints, retrieval tools, input data, and processing stages. When a specified service is unavailable or the task exceeds its processing scope, the back end may call a default or backup endpoint, resulting in a discrepancy between the interface-shown operation and the actual call path.

[0004] The second type of existing technology utilizes model gateways or load balancers to select the execution end based on cost, latency, health status, and concurrency. This type of technology primarily addresses service availability issues, typically treating requests as a uniform payload without distinguishing the usage boundaries of project documents, public literature snippets, search terms, and runtime logs. Consequently, it is difficult to determine whether an execution end, even if it meets processing capabilities, is permitted to receive specific fields at the current stage.

[0005] The existing third type of technology uses network access control, field masking, or audit logs to reduce the risk of data leakage. These measures are usually implemented separately: network control cannot identify the stage of the research workflow, field processing programs may not know the minimum input required by the target endpoint, and audit logs do not link the running strategy, execution end capability snapshot, outbound fields, actual call records, and result release status into the same verification chain.

[0006] The existing fourth type of technology uses a phased workflow to complete problem structuring, knowledge retrieval, candidate hypothesis generation, evidence preparation, and research plan output. Different processing modules may call models or tools separately; if each module reads the environment configuration independently or uses its own fallback rules, task-level constraints are prone to failure when passed across phases. When the front-end stops polling, background tasks may still continue to run, resulting in inconsistencies between the interface state, background state, and resource usage state.

[0007] Therefore, in related technologies, when scientific research tasks call project knowledge bases, text processing models, academic data retrieval tools and other processing components, there is an urgent need for a way to maintain consistent calling constraints and data boundaries across stages of execution, to ensure that the front-end configuration and the actual back-end calling path are consistent, and to ensure that the front-end display status is consistent with the actual back-end status. Summary of the Invention

[0008] Therefore, it is necessary to provide a strategy constraint routing and result verification method for multiple execution ends of scientific research tasks to address the above-mentioned technical problems.

[0009] Firstly, this application provides a method for policy constraint routing and result verification for multi-execution ends of scientific research tasks. The method includes: Construct a runtime strategy object, which includes constraints on the computation location of the research task, network access constraints, tool usage constraints, data outbound constraints, and alternative path rules; A capability declaration object is created for each callable execution terminal. The capability declaration object contains the deployment location of the execution terminal, the supported data types, the upper limit of processing capacity, the available tools, and the health status. If the capability declaration expires or the health check fails, the execution terminal will not participate in the routing. The research task is broken down into multiple processing stages, and a stage requirement object is generated for each processing stage. The stage requirement object includes the execution capabilities, tool types, network access requirements, and input data types required for that stage. The task input data is represented as a scientific research asset object, which includes the data's sensitivity level, original or derived markers, allowed receiver range, and allowed transformation methods. For each processing stage and each candidate execution terminal, a hard constraint admission judgment is made based on the stage requirement object, the capability declaration object, the operation strategy object, and the scientific research asset object. Only when the execution terminal's capability meets the stage requirement, the execution terminal's call conforms to the operation strategy constraint, and the input asset is allowed to be received by the execution terminal, is the execution terminal included in the set of allowed execution paths. When the set of allowed execution paths is empty, automatic switching to endpoints not included in the set is prohibited. When the call from the selected execution end needs to leave the controlled boundary of the project, an outbound list object is generated according to the running strategy object. The outbound list object contains the range of fields to be sent, and fields not included in the outbound list object are not included in the actual request. Before invoking the execution end, an expected execution description is generated, which includes the expected execution end, the expected tool type, the running strategy version, and the request package identifier; After the call is completed, an actual execution description is collected, which includes the actual provider, the actual tool version, the actual network access records, the actual request packet identifier, and the output identifier. The actual execution description is checked for consistency with the expected execution description and the running strategy object. If the check passes, the phase result is written into the formal workflow. If the check fails, the phase result is written into the isolation area. Key events during task execution are linked together using a preceding hash to form a chain of audit records.

[0010] Optionally, in one embodiment of this application, the hard constraint admission judgment requires that the available values ​​of each capability of the candidate execution terminal meet the minimum requirements of the processing stage for the corresponding capability, that the use of each resource by the candidate execution terminal conforms to the permission value of the operation policy object for the resource behavior, and that each asset in the input asset set of the processing stage is allowed to be received by the candidate execution terminal; if any condition is not met, the candidate execution terminal is excluded.

[0011] Optionally, in one embodiment of this application, when the set of allowed execution paths contains multiple candidate execution endpoints, an execution endpoint is selected within the set of allowed execution paths based on a soft selection metric. The soft selection metric includes at least one or more combinations of task adaptation quality, health reliability, predicted latency, invocation cost, and outbound risk.

[0012] Optionally, in one embodiment of this application, when the call of the selected execution end needs to leave the controlled boundary of the project, the data to be sent is subjected to minimization processing according to the running strategy object and the scientific research asset object. The minimization processing includes one or more of the following: field deletion, field projection, text truncation, numerical range conversion, alias replacement and irreversible summary. The processing result forms a derived asset and retains the parent asset identifier and transformation chain.

[0013] Optionally, in one embodiment of this application, the consistency check compares each behavior value in the actual execution description with the range allowed by the expected execution description and the running strategy object. The check passes when all behaviors are within the allowed range and the request packet and output pass the integrity check; otherwise, the check fails.

[0014] Optionally, in one embodiment of this application, the circumstances under which the verification fails include at least the following: the actual provider does not match the expected execution end, network access occurs during the prohibited network phase, the actual request packet does not match the expected request packet, an unauthorized backup endpoint is used, and the actual sent field exceeds the scope of the outbound list object.

[0015] Optionally, in one embodiment of this application, a background task control step is also included: the task status is recorded by the background state machine; when the user requests to pause, the orchestrator writes a checkpoint object at the processing stage boundary or after the call is completed and enters the pause state; when resuming, the capability declarations of each execution end are re-collected and the current running strategy is re-verified; if the original execution end is unavailable, the execution end is re-selected from the new set of allowed execution paths.

[0016] Optionally, in one embodiment of this application, the checkpoint object includes a task identifier, a stage identifier, a completed step identifier, an input asset identifier, an output asset identifier, a running policy version, a capability snapshot identifier, a current routing decision, and an audit chain status; during recovery, the validity of the checkpoint object is also verified, and the validity includes policy version consistency, asset integrity, audit chain continuity, and idempotency conditions.

[0017] Optionally, in one embodiment of this application, the hash value of each audit node in the chained audit record is calculated from the hash value of the previous audit node, the current event type, and the current event payload, and is used to identify situations where intermediate records are modified or missing.

[0018] Optionally, in one embodiment of this application, an execution proof digest is generated after all processing stages are completed. The execution proof digest includes the requested execution mode, the actual execution mode, the type of tool used, the network access type, the type of external asset, the execution strategy version, and the audit chain status information.

[0019] The above-mentioned method for policy constraint routing and result verification for multi-execution ends of scientific research tasks has the following advantages compared with existing technologies: Compared to simply setting a "local, cloud, or network" switch on the front end, this invention writes the user's selection into a runtime policy object with a version number and explicit permission fields, and re-performs access control at each processing stage, thus establishing a verifiable correspondence between the runtime mode shown on the interface and the backend call path.

[0020] This invention generates a set of permissible execution paths based on the capability declaration object and the stage requirement object. It first performs hard constraint screening based on execution capabilities, strategies, and data boundaries, and then compares quality, reliability, latency, and cost within the compliant set. This prevents automatic switching to unpermitted backup endpoints when the execution end lacks sufficient capability, and also avoids selecting endpoints that cannot receive current research data as execution paths.

[0021] This invention controls the scope of outbound requests by using research asset objects, field whitelists, derived assets, and outbound lists, ensuring that research task cards, project documents, knowledge fragments, search queries, and operation logs are processed according to different boundaries. Outbound requests retain their transformation chain and hash before being sent, and fields not included in the outbound list are not allowed to enter into actual requests.

[0022] This invention collects the actual provider, actual model or tool version, network tool events, request packet hash, and output hash after invocation, and performs consistency verification with the expected execution description and running strategy. When a strategy violation is detected, the stage result enters an isolation zone and is not passed to subsequent stages, thereby reducing execution deviations caused by misuse of backup endpoints, configuration changes, and unexpected modifications to request packets.

[0023] This invention utilizes chained audit object connection strategy confirmation, capability snapshots, routing selection, data outbound, call records, checkpoints, and result release events to identify situations where intermediate records have been modified or are missing, and to generate a summary of the request strategy and the actual execution method. This process does not require retraining of the text processing model.

[0024] In the implementation of background task status control and checkpoint mechanisms, pausing, resuming, and canceling are achieved through background state transitions; upon resumption, capabilities and policies are re-verified, and stopping polling on the interface is not used as a basis for background suspension. When the execution end changes, it is only reselected from the new compliant path, thereby maintaining the recoverability and state consistency of long-running tasks.

[0025] The aforementioned effects are generated by a combination of operational strategies, capability and asset constraints, phased access control, outbound control, post-call verification, and chain auditing, and are not necessarily based on unverified accuracy, the correctness of scientific research conclusions, or a fixed efficiency improvement ratio. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating a strategy constraint routing and result verification method for multiple execution ends of a scientific research task in one embodiment. Figure 2 This is a connection diagram between each processing module and the controlled execution object in one embodiment; Figure 3 Here is a flowchart of stage-level compliance path generation and model routing in one embodiment; Figure 4 This is a flowchart of the scientific research asset exit access control and minimization process in one embodiment; Figure 5 This is a diagram illustrating the actual execution of verification, chained auditing, and background state control in one embodiment. Figure 6 This is a flowchart illustrating the controlled execution of a research task based on a project knowledge base and public searches, as shown in one embodiment. Figure 7 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0028] In this embodiment, the method is executed by a controlled orchestration program deployed on a computing device. The computing device can communicate with one or more of the following: a user interface, a project knowledge base, a text processing model service, an academic document retrieval interface, a document parsing component, a task queue, and an audit storage. It can also connect to a local model service as needed. The controlled orchestration program includes a capability declaration module, a runtime strategy construction module, a research asset grading module, a phased requirements generation module, a routing access module, a data outbound control module, an execution invocation module, an actual execution verification module, a checkpoint control module, a result release module, and a chained audit module. These modules can be executed by a unified task orchestration program or distributed to multiple processing units for execution.

[0029] Figure 1 The overall flow of this method is illustrated. Instead of directly executing tasks based on the schema name in the user interface, this method first converts it into a runtime policy object and establishes expiration-terminal capability claims for each execution endpoint, retrieval tool, and processing component. After the research task is broken down into multiple processing stages, each stage performs capability access, policy access, and data access respectively. If no compliant path exists, a blocking state is entered, and backup endpoints not permitted by the runtime policy are not used.

[0030] Figure 2 The diagram illustrates the connections between each processing module and the controlled execution object. The runtime strategy object stores the user-confirmed computation location and data boundaries; the capability declaration object describes the functions and limitations provided by each execution endpoint; the research asset object describes the types, sensitivity levels, and permissible transformations of task inputs and intermediate data; the phase requirement object describes the capabilities and tools required for the current phase; the routing decision object records candidate endpoints, reasons for exclusion, and the final selection; the outbound list object records data leaving the project boundary; the expected execution description and actual execution description record the invocation plan and invocation facts, respectively; the checkpoint object is used for task state saving, restoration, and idempotent retries; and the audit chain object connects critical events.

[0031] In this embodiment, the method first constructs a runtime strategy object. Upon receiving user confirmation of the runtime mode and project-level usage requirements, the runtime strategy object is constructed. This object includes constraints on the computational location, network access, tool usage, data outbound transmission, and alternative path rules for the research task. Specifically, the runtime strategy object includes at least a strategy identifier, project identifier, version number, data classification level, allowed computational locations, allowed service providers, allowed network access types, allowed tool sets, original data outbound transmission permissions, derived data outbound transmission permissions, allowed outbound transmission fields, retention period, regional restrictions, encryption requirements, de-identification requirements, alternative path rules, audit level, confirmation status, and effective time. The interface can display runtime modes such as local processing, controlled networking, or external service calls as needed by the project; however, actual execution should read the specific constraint fields in the runtime strategy object and should not infer permissions solely based on the mode name.

[0032] A capability declaration object is created for each callable execution endpoint. This object includes the endpoint's deployment location, supported data types, maximum processing capacity, available tools, and health status. Execution endpoints will not participate in routing if their capability declarations expire or health checks fail. Execution endpoints may include project knowledge base retrieval services, text processing models, vector retrieval services, academic document retrieval tools, document parsers, and local processing services configured according to deployment requirements. The capability declaration object must include at least the endpoint identifier, provider, deployment location, service or tool version, supported data types, maximum input size, structured output format, available network tools, data retention declaration, region information, encryption capabilities, task control granularity, health status, concurrency capacity, declaration generation time, and expiration time.

[0033] The research task is broken down into multiple processing stages, and a stage requirement object is generated for each stage. This stage requirement object includes the execution capabilities, tool types, network access requirements, and input data types needed for that stage. The research workflow is divided into several stages, which can correspond to problem structuring, knowledge integration, candidate hypothesis generation, evidence preparation, research plan generation, and result output, or other internal divisions can be used. Each stage forms a stage requirement object, which includes at least a stage identifier, a set of input assets, required data types, minimum context size, structured output requirements, allowed tools, internet access requirements, original document usage requirements, output asset type, maximum allowable delay, checkpoint location, and failure handling method. The specific research processing methods for each stage are not limited by this invention.

[0034] Task input data is represented as research asset objects, which include the data's sensitivity level, original or derived markers, permitted recipient range, and permitted transformation methods. Original questions, research task cards, project documents, public literature, evidence fragments, search terms, model requests, phase results, and operation logs are each represented as research asset objects. In specific deployments, the assets may also include experimental parameters or image files authorized for use by the project. Research asset objects include at least the following: asset identifier, project identifier, media type, storage location, content hash, source type, original or derived markers, sensitivity level, identifiability level, permitted use phase, permitted recipient range, permitted transformation set, external release status, retention period, and manual confirmation status.

[0035] For each processing stage and each candidate execution terminal, a hard constraint admission judgment is made based on the stage requirement object, the capability declaration object, the operation strategy object, and the scientific research asset object. The execution terminal is included in the set of allowed execution paths only when the execution terminal capability meets the stage requirement, the execution terminal call conforms to the operation strategy constraint, and the input asset is allowed to be received by the execution terminal. When the set of allowed execution paths is empty, automatic switching to endpoints not included in the set is prohibited. Figure 3 This illustrates the phased compliance path generation and model routing process. (For each phase...) and candidate execution end The system checks capabilities, policies, and data conditions respectively. (Definition) For the execution end In the Available values ​​for the item's capabilities For the stage Minimum requirements for this ability; For the execution strategy to the first The permission value for a resource behavior. To determine whether the corresponding resource is used for this path; Assets Is it allowed by the execution end? Received. The hard constraint admission result can be calculated using the following formula. :

[0036] in, This is an indicator function; it takes the value 1 if the condition is met, and 0 otherwise. For the stage Required set of input assets. Only if... When equal to 1, the execution end Enter the set of allowed execution paths Failure to meet any hard constraint cannot be compensated for by increasing the quality score or reducing costs. Simultaneously, the excluded endpoints and their corresponding reasons, such as insufficient capabilities, policy prohibitions, data prohibitions, or expired declarations, should be stored. When multiple execution endpoints are involved, the system calculates the soft selection metric only within that set. Preferably, endpoint selection utility... Quality adapted to the task Health and reliability Prediction delay Calling cost and exit risks composition:

[0037] in, to Non-negative weights. Selected execution endpoint. for middle Higher endpoints. The soft selection metric is only used for comparison between endpoints that both satisfy the hard constraints and does not change... Defined data and network boundaries. When When empty, the system outputs the reason for non-execution and optional adjustment items, such as adding required services, narrowing the processing scope, adjusting the allowed sending fields, or requiring the user to explicitly change the operating strategy; other endpoints should not be used automatically before user confirmation.

[0038] When the selected execution endpoint's call needs to leave the project's controlled boundary, an outbound manifest object is generated based on the runtime strategy object. This outbound manifest object contains a range of fields to be sent; fields not listed in the outbound manifest object are not included in the actual request. When the selected execution endpoint is outside the project's controlled boundary, or the call needs to traverse an external network, an outbound manifest object is generated based on the stage input. This object includes at least the task identifier, stage identifier, target execution endpoint, target region, asset identifier to be sent, fields to be sent, original or derived state, transformation steps, hashes before and after transformation, sending purpose, network tools, retention requirements, risk level, approval method, and validity period.

[0039] Before invoking the execution endpoint, an expected execution description is generated. This description includes the expected execution endpoint, expected tool type, runtime policy version, and request packet identifier. After routing is determined, the expected execution description is generated. This description includes at least a stage identifier, expected execution endpoint, expected model or tool type, expected deployment location, allowed network tools, request packet hash, capability snapshot hash, runtime policy version, maximum number of retries, allowed set of backup endpoints, and expected output type. The call is initiated based on the expected execution description. Any retries or backup switching should trigger a re-execution of the stage-level admission judgment.

[0040] After the call is completed, an actual execution description is collected. This description includes the actual provider, actual tool version, actual network access records, actual request packet identifier, and output identifier. After the call ends, another actual execution description is collected. This description includes at least the actual provider, actual model or tool version, endpoint category, actual network access records, actual request packet hash, model identifier returned in the response, start time, end time, status code, number of retries, actual standby failover, output hash, and error code. Fields that cannot be obtained from external endpoints are marked as unobservable and must not be considered verified solely based on configuration values.

[0041] The actual execution description is compared with the expected execution description and the runtime strategy object for consistency verification. If the verification passes, the stage result is written to the formal workflow; if the verification fails, the stage result is written to the isolation area. The actual execution description, expected execution description, and runtime strategy object are compared item by item. Define the first... Actual behavior Does it belong to the set of expected behaviors and strategies? This results in a violation count. :

[0042] when equal And when the request packet hash, output hash, and execution end return information pass the integrity check, the stage result is written into the formal workflow; when Greater than At that time, the results are placed in the isolation zone and cannot be used in subsequent stages, and a policy violation record is generated.

[0043] Key events during task execution are linked together using a preceding hash to form a chain of audit records. Figure 5 The actual execution verification, chain auditing, and backend status control processes are illustrated. The chain auditing module generates audit nodes for events such as policy confirmation, capability snapshots, asset classification, routing selection, outbound processing, request sending, response receiving, checkpoint writing, status changes, violation handling, and result release. In one embodiment of this application, the hash value of each audit node in the chain audit record is calculated from the hash value of the previous audit node, the current event type, and the current event payload, used to identify situations where intermediate records have been modified or are missing. The specific calculation formula is as follows:

[0044]

[0045] in, For cryptographic hash functions, This is a summary of the corresponding structured event payload. After saving the hash of the previous node, any modification, deletion, or insertion of an intermediate event will cause changes in subsequent chain values. This audit chain is used to verify the continuity between system records; for unobservable behavior within external endpoints, the audit record should indicate the source of information and the scope of observability.

[0046] In one embodiment of this application, when the selected execution end's call needs to leave the project's controlled boundary, a minimization process is performed on the data to be sent based on the running strategy object and the research asset object. This minimization process includes one or more of the following: field deletion, field projection, text truncation, numerical range formatting, alias replacement, and irreversible summarization. The processing result forms a derived asset while retaining the parent asset's identifier and transformation chain. Specifically, the processing method is selected based on the running strategy and the research asset object, including deleting local paths and internal numbers, field whitelist projection, text fragment selection, numerical range formatting, variable aliasing, retrieval-based abstraction, pseudonym replacement, and irreversible summarization. When the input includes an image and corresponding processing components are deployed, image cropping, feature extraction, or local statistical extraction can also be used. The processing result forms a new derived asset object, retaining the parent asset's identifier and transformation chain, without overwriting the original asset. Figure 4 The document illustrates the exit access control and minimization process for research assets.

[0047] To compare the external risks associated with different handling methods, asset sensitivity can be considered. Identifiability Transformation reversibility Risks remain at the target end and de-identification sufficiency calculate :

[0048] Among them, each item is normalized to to , to As the weights are non-negative, the calculation results can be truncated and normalized to each component. to Scope. Risk value does not replace prohibitive constraints in the operating strategy: even if the risk value is low, as long as... Even if project documents, internal fields, or knowledge fragments are explicitly prohibited from being shared externally, the system still must not send them. The risk value is used to compare request packet size and processing adequacy when allowing the sharing of derived data.

[0049] In one embodiment of this application, the consistency check compares each behavior value in the actual execution description item by item to see if it falls within the allowed range of the expected execution description and the runtime policy object. The check passes when all behaviors fall within the allowed range and both the request packet and output pass the integrity check; otherwise, the check fails. Specifically, for each behavior in the actual execution description (such as the actual provider, the actual tool version, the actual network access record, etc.), it must fall within the allowed set of the expected execution description and the runtime policy object. In addition, the actual request packet hash must match the request packet hash in the expected execution description, and the output hash must pass the integrity check. The validation passes when all of the above conditions are met; otherwise, it fails.

[0050] In one embodiment of this application, the verification failure scenarios include at least the following: the actual provider does not match the expected execution end; network access occurs during the network access prohibition phase; the actual request packet does not match the expected request packet; an unauthorized backup endpoint is used; and the actual sent fields exceed the range of the outbound list object. All of the above scenarios are... The specific manifestation of a value greater than 0 is that, in any of the following situations, the stage result enters the isolation zone, is not available for subsequent stages to reference, and a strategy violation record is generated.

[0051] In one embodiment of this application, a background task control step is also included: the task status is recorded by the background state machine; when a user requests to pause, the orchestrator writes a checkpoint object after processing a stage boundary or completing a call and enters a paused state; upon resumption, the capability declarations of each execution end are re-collected and the current running strategy is re-verified. If the original execution end is unavailable, an execution end is reselected from the new set of allowed execution paths. In the implementation of configuring background task control, the task status is recorded by the background state machine, not solely based on the interface local identifier. The status includes at least Created, Pause Request, Paused, Resume Request, Running, Cancellation Request, Cancelled, and Blocked. When a user requests to pause, the pause request status is written first; the orchestrator stops at a safe point after a stage boundary, retrieval batch boundary, or completion of the current model call and writes a checkpoint object. The front end can only display that the task is paused after the background status enters the paused state. After the cancellation request enters the cancellation request state, the orchestrator stops initiating new calls, cancels tasks that have not yet been sent, releases temporary resources, and retains the audit records that have been formed. Closing the page only changes the user interface and is not equivalent to pausing or canceling the background task.

[0052] In one embodiment of this application, the checkpoint object includes a task identifier, a stage identifier, a completed step identifier, an input asset identifier, an output asset identifier, a running policy version, a capability snapshot identifier, a current routing decision, and an audit chain status. During recovery, the validity of the checkpoint object is also verified, including policy version consistency, asset integrity, audit chain continuity, and idempotency conditions. The checkpoint object includes at least a task identifier, a stage identifier, a completed step, an input asset hash, an output asset hash, a running policy version, a capability snapshot hash, a current routing decision, an outbound list identifier, an audit chain head value, an idempotency key, and a step to be executed. During recovery, the original execution endpoint is not directly reused; instead, capability declarations are re-collected and the current policy is verified. If the original execution endpoint fails but other compliant paths exist, a new route can be selected and recorded from the set of allowed execution paths. If no compliant path exists, the process remains paused or enters a blocked state.

[0053] The effectiveness of checkpoints can be determined by policy revalidation, asset integrity, audit chain continuity, and idempotency conditions.

[0054] when If the value is 1, recovery from the checkpoint is allowed; otherwise, return to the most recent valid checkpoint or re-execute the corresponding phase.

[0055] In one embodiment of this application, an execution proof digest is generated after all processing stages are completed. This execution proof digest includes the requested execution method, the actual execution method, the type of tool used, the network access type, the type of outsourced assets, the execution policy version, and audit chain status information. After all stages are completed, the violation count for each stage is checked. The system checks whether the value is 0, whether the outbound list is closed, whether the audit chain is continuous, whether the policy version is consistent, and whether there are any unprocessed blocked or canceled request states. After passing these checks, the system outputs the research task results and an execution proof summary. The execution proof summary should at least record the requested execution method, the actual execution method, the type of model and tool used, the network access type, the type of outsourced assets, the data processing performed, the execution policy version, the audit chain head value, and unobservables. The front-end page, downloaded files, and interface responses for the same research task should reference the same execution proof summary. The "local execution," "paused," or "not connected" status displayed on the front-end should be based on the actual status and effective execution policy returned by the controlled orchestration program. The scientific correctness of the research results is the responsibility of the corresponding research methods and manual verification; this invention addresses the consistency of execution boundaries, routing selection, and actual execution records. Specific Implementation Example 1 Figure 6 This demonstrates a controlled execution process for research tasks based on a project knowledge base and public searches.

[0057] In one specific embodiment, the user proposes a research problem on ultra-precision machining in text form, requiring an analysis of the potential impact of machining parameters, tool condition, material properties, and thermal stability conditions on machining quality, and a summary of the current research status, candidate research directions, and subsequent research plans. The task input primarily consists of text-based problems and project knowledge base information.

[0058] The research questions input by the user are represented as research asset objects, authorized research materials such as PDFs and DOCX files in the project knowledge base are represented as research asset sets, and academic search queries formed based on research tasks are represented as research asset sets. Whether user input and project documents are allowed to be sent to external endpoints is determined by the operational policy. This embodiment connects the project knowledge base retrieval component, the configured text processing model, and optional academic search tools, and establishes capability declaration objects for each. Execution terminals that are not configured, fail health checks, or have expired capability declarations do not participate in this round of routing.

[0059] Confirm the operational strategy. The operational strategy stipulates: the use of the configured text processing model and project knowledge base is allowed; the retrieval of publicly available information through the academic retrieval interface is allowed; the use of unauthorized general web search tools is prohibited; only field-filtered search queries are sent to the academic retrieval interface; only the necessary task fields and filtered knowledge fragments for the current stage are sent to the external model.

[0060] In the problem understanding phase, candidate execution ends are assessed for capability, strategy, and data access based on the phase's requirements. After passing the access assessment, the text processing model generates a structured research task card based on the original problem. This task card includes at least the research object, target phenomenon, candidate variables, constraints, expected output, and fields requiring clarification. The task card, after user confirmation, serves as the unified input for subsequent phases. In the project knowledge retrieval phase, the project knowledge base is searched based on the task card to obtain candidate knowledge fragments related to the research object, variables, mechanisms, methods, boundary conditions, and counterexamples. The project knowledge base retrieval can employ one or more of the following methods: keyword retrieval, vector retrieval, hybrid retrieval, and re-ranking. This phase does not access external networks.

[0061] During the public data retrieval phase, a field whitelist and an outbound list are generated according to the operational strategy. The outbound list only includes filtered search queries and does not contain internal project paths, user identity information, internal document numbers, or other unauthorized fields. After passing through the outbound access control, academic search tools are invoked to obtain public data candidates with titles, authors, years, source addresses, or other traceable information. In the comprehensive analysis phase, the content required for the current stage is selected from the research task card, project knowledge base candidate fragments, and public data candidates to form a request package. The request package does not contain complete project documents and preferably includes structured task fields, filtered knowledge fragments, source identifiers, and necessary boundary conditions. The request package is sent to a text processing model allowed by the operational strategy to form candidate hypotheses, supporting and conflicting evidence, methodological comparisons, and subsequent research plans.

[0062] Before each model or tool call, an expected execution description is generated, recording the expected provider, model or tool category, runtime policy version, allowed network tools, capability snapshot hash, and request packet hash. After the call, observable actual providers, actual model or tool versions, network tool events, actual request packet hash, output hash, and call status are collected and compared with the expected execution description and runtime policy. When the actual behavior conforms to the runtime policy and the request packet integrity check passes, the phase result is written to the formal workflow; when unauthorized tool calls, fields sent exceeding the outbound list, mismatched execution providers, or use of unauthorized alternate endpoints are detected, the phase result is placed in an isolation zone, not to be referenced by subsequent phases, and a policy violation record is generated.

[0063] When academic search tools are unavailable or network access is prohibited by the operational policy, the system must not automatically switch to unauthorized network search tools. Instead, it must continue execution using only the project knowledge base, recording any insufficient coverage of publicly available information as items to be supplemented. If no compliant execution path exists at the current stage, the system enters a policy blocking state and returns the reason. After all stages are completed, the system outputs research task results and an execution proof summary. The summary records the types of models and tools actually used, access to the academic search interface, asset types sent to external endpoints, request packet hashes, operational policy versions, and unobservables. Specific Implementation Example 2 In another embodiment, the execution policy prohibits external network access, uses only the project knowledge base and permitted text processing models to complete tasks, and records any lack of external information as items to be supplemented. If the project does not have a local model configured, the capability declaration module marks it as unavailable, and the routing admission module selects the configured external model only when the execution policy allows it.

[0065] In the above embodiments, the text processing model can be the Qianwen model or other models capable of outputting natural language or structured data; the task can be executed by a unified orchestration program, or by multiple processing modules collaboratively under the same operational strategy constraints; the external interface can be academic document retrieval or other controlled knowledge services. This invention does not limit the specific model, retrieval algorithm, or report generation method.

[0066] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0067] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 7 As shown, the computer device includes a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When executed by the processor, the computer program implements a strategy-constrained routing and result verification method for multi-execution terminals in scientific research tasks. The display screen can be an LCD screen or an e-ink display screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0068] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0069] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0070] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0071] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0072] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0073] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0074] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0075] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for strategy-constrained routing and result verification for multi-execution terminals of scientific research tasks, characterized in that, The method includes: Construct a runtime strategy object, which includes constraints on the computation location of the research task, network access constraints, tool usage constraints, data outbound constraints, and alternative path rules; A capability declaration object is created for each callable execution terminal. The capability declaration object contains the deployment location of the execution terminal, the supported data types, the upper limit of processing capacity, the available tools, and the health status. If the capability declaration expires or the health check fails, the execution terminal will not participate in the routing. The research task is broken down into multiple processing stages, and a stage requirement object is generated for each processing stage. The stage requirement object includes the execution capabilities, tool types, network access requirements, and input data types required for that stage. The task input data is represented as a scientific research asset object, which includes the data's sensitivity level, original or derived markers, allowed receiver range, and allowed transformation methods. For each processing stage and each candidate execution terminal, a hard constraint admission judgment is made based on the stage requirement object, the capability declaration object, the operation strategy object, and the scientific research asset object. Only when the execution terminal's capability meets the stage requirement, the execution terminal's call conforms to the operation strategy constraint, and the input asset is allowed to be received by the execution terminal, is the execution terminal included in the set of allowed execution paths. When the set of allowed execution paths is empty, automatic switching to endpoints not included in the set is prohibited. When the call from the selected execution end needs to leave the controlled boundary of the project, an outbound list object is generated according to the running strategy object. The outbound list object contains the range of fields to be sent, and fields not included in the outbound list object are not included in the actual request. Before invoking the execution end, an expected execution description is generated, which includes the expected execution end, the expected tool type, the running strategy version, and the request package identifier; After the call is completed, an actual execution description is collected, which includes the actual provider, the actual tool version, the actual network access records, the actual request packet identifier, and the output identifier. The actual execution description is checked for consistency with the expected execution description and the running strategy object. If the check passes, the phase result is written into the formal workflow. If the check fails, the phase result is written into the isolation area. Key events during task execution are linked together using a preceding hash to form a chain of audit records.

2. The method for strategy constraint routing and result verification of multi-execution terminals in scientific research tasks according to claim 1, characterized in that, The hard constraint admission criteria require that the available values ​​of all capabilities of the candidate execution terminal meet the minimum requirements of the processing stage for the corresponding capabilities, that the candidate execution terminal's use of all resources conforms to the permission value of the operation strategy object for the resource behavior, and that each asset in the input asset set of the processing stage is allowed to be received by the candidate execution terminal; if any condition is not met, the candidate execution terminal is excluded.

3. The method for strategy constraint routing and result verification of multi-execution terminals in scientific research tasks according to claim 1, characterized in that, When the set of allowed execution paths contains multiple candidate execution endpoints, an execution endpoint is selected from the set of allowed execution paths based on soft selection criteria. The soft selection criteria include at least one or more combinations of task adaptation quality, health reliability, predicted latency, invocation cost, and outbound risk.

4. The method for strategy constraint routing and result verification of multi-execution terminals in scientific research tasks according to claim 1, characterized in that, When the selected execution end needs to leave the project's controlled boundary, the data to be sent is subjected to minimization processing according to the running strategy object and the scientific research asset object. The minimization processing includes one or more of the following: field deletion, field projection, text truncation, numerical range conversion, alias replacement, and irreversible summarization. The processing result forms a derived asset and retains the parent asset identifier and transformation chain.

5. The method for strategy constraint routing and result verification of multi-execution terminals in scientific research tasks according to claim 1, characterized in that, The consistency check compares each behavior value in the actual execution description with the range allowed by the expected execution description and the running strategy object. The check passes when all behaviors are within the allowed range and the request packet and output pass the integrity check; otherwise, the check fails.

6. The method for strategy constraint routing and result verification of multi-execution terminals in scientific research tasks according to claim 5, characterized in that, The circumstances under which the verification fails include at least the following: the actual provider does not match the expected execution end; network access occurs during the prohibited network phase; the actual request packet does not match the expected request packet; an unauthorized backup endpoint is used; and the actual sent fields exceed the scope of the outbound list object.

7. The method for strategy constraint routing and result verification of multi-execution terminals in scientific research tasks according to claim 1, characterized in that, It also includes background task control steps: the task status is recorded by the background state machine; when the user requests to pause, the orchestrator writes the checkpoint object at the processing stage boundary or after the call is completed and enters the pause state; when resuming, the capability declarations of each execution end are re-collected and the current running strategy is re-verified. If the original execution end is unavailable, the execution end is re-selected from the new set of allowed execution paths.

8. The method for strategy constraint routing and result verification of multi-execution terminals in scientific research tasks according to claim 7, characterized in that, The checkpoint object includes a task identifier, a stage identifier, a completed step identifier, an input asset identifier, an output asset identifier, an operating policy version, a capability snapshot identifier, a current routing decision, and an audit chain status. During recovery, the validity of the checkpoint object is also verified, and the validity includes policy version consistency, asset integrity, audit chain continuity, and idempotency conditions.

9. The method for strategy constraint routing and result verification of multi-execution terminals in scientific research tasks according to claim 1, characterized in that, The hash value of each audit node in the chained audit record is calculated from the hash value of the previous audit node, the current event type, and the current event payload, and is used to identify situations where intermediate records have been modified or are missing.

10. The method for strategy constraint routing and result verification of multi-execution terminals in scientific research tasks according to claim 1, characterized in that, After all processing stages are completed, an execution proof digest is generated. The execution proof digest includes the requested execution method, the actual execution method, the type of tool used, the network access type, the type of external asset, the execution strategy version, and the audit chain status information.