Parallel orchestration and non-blocking execution methods and systems for cross-model auditing

By employing a parallel orchestration and non-blocking execution method for cross-model auditing, the stability problem of multi-voice deep auditing under a single-call non-renewable hard timeout protocol is solved. This enables plug-and-play functionality for heterogeneous voices and fault isolation for single voices, ensuring the reliability and repeatability of audit results.

CN123114652APending Publication Date: 2026-11-17SHENZHEN TOUCHU JIUXUAN NETWORK TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611231644.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-13
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Under a protocol with a single call that cannot be renewed due to a hard timeout, how can we complete multi-part deep auditing that far exceeds the hard timeout, avoid the waiting party from mistakenly killing the background auditing task due to timeout, achieve plug-and-play functionality for heterogeneous parts and isolation of single-part faults, avoid missing parts causing the entire batch of audits to fail or the billing parts to run in vain, and ensure that audit results can be retrieved across processes and are not re-delivered?

Method used

By employing a parallel orchestration and non-blocking execution method for cross-model auditing, an audit panel is constructed using a voice registry and a unified transport dispatch table. Voice availability gating is implemented, background audit tasks are dispatched in parallel within separate subprocesses, independent exception capture boundaries and timeout limits are set, and persistent storage of audit results and cross-process restart are achieved through task masking and single-time restricted waiting mechanisms.

Benefits of technology

It enables plug-and-play functionality for heterogeneous audio components, ensuring the robustness of the audit panel, avoiding mid-process failures caused by missing audio components, allowing the final results to be retrieved in stages after timeout, preventing the background audit task from being mistakenly killed, and supporting cross-process restart and deduplication of results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123114652A_ABST
    Figure CN123114652A_ABST
Patent Text Reader

Abstract

This invention discloses a parallel orchestration and non-blocking execution method and system for cross-model auditing. The method includes: verifying the dispatchability of each preset voice combination during the voice registry loading period; receiving the audited material and voice combination parameters; selecting a heterogeneous voice set; executing voice availability gating, returning a failure state and ending the call when the panel is empty; allocating audit flags and creating background audit tasks; the background audit task executing in separate subprocesses and parallelly dispatching each voice through a unified transmission dispatch table; implementing single-time limited waiting through task masking; and persistently storing the in-transit status and final result in two separate directories. This invention achieves plug-and-play heterogeneous voices, single-voice fault isolation, service process recycling without canceling audits, audits exceeding a single call hard timeout allowing the final result to be retrieved in stages via audit flags, and audit results being retrieved upon cross-process restart.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence auditing technology, specifically to a parallel orchestration and non-blocking execution method and system for cross-model auditing. Background Technology

[0002] In software engineering and code quality assurance scenarios, using artificial intelligence models to audit materials has become a common practice. Traditional single-model auditing methods submit materials to a single AI inference agent for review. However, due to the training distribution and cognitive blind spots of a single model, it is difficult to detect its inherent biases and omissions.

[0003] To overcome the aforementioned shortcomings, existing technologies have developed schemes that submit the same audited material to multiple different models for independent review. Parallel independent multi-model review is common knowledge in the field prior to the filing date. However, under the constraints of a call protocol with a single-call non-renewable hard timeout, how to complete multi-voice deep auditing that far exceeds the hard timeout, while ensuring that the waiting party's own timeout does not mistakenly terminate the background auditing task that is already being executed, remains an unsolved technical problem.

[0004] Furthermore, the parallel job orchestration of continuous integration systems is geared towards deterministic inspection tools and does not involve the adaptation of differentiated transmission channels for heterogeneous vendor models. Its result retrieval mechanism is also different from the calling protocol constraints of this solution, so it cannot be directly applied to cross-model auditing scenarios. Summary of the Invention

[0005] The technical problems to be solved by this invention include: how to complete multi-part deep auditing that far exceeds the hard timeout in a call protocol with a single call that cannot be renewed; how to avoid the waiting party from accidentally killing the background audit task due to its own timeout; how to achieve plug-and-play functionality of heterogeneous parts and isolation of single-part faults; how to avoid the failure of the entire batch of audits or the empty running of billing parts due to missing parts; and how to ensure that audit results can be retrieved across processes and are not re-delivered.

[0006] This invention provides a parallel orchestration and non-blocking execution method and system for cross-model auditing to solve the above-mentioned technical problems.

[0007] This invention provides a parallel orchestration and non-blocking execution method for cross-model auditing. The method is applied to an audit orchestration device that interacts with the caller via a calling protocol. This calling protocol has a non-renewable hard timeout for each single call. Before responding to any call, during the voice registry loading period, the method verifies that the voice identifiers referenced by each preset voice combination in the voice registry exist in the voice registry, and that the registered transmission channels have corresponding transmission executors in the unified transmission dispatch table. This verification does not check the existence of command-line tools and keys. The following steps are executed only after the verification passes; if the verification fails, no further auditing services are provided. Receive the audited materials and sound part combination parameters submitted by the caller; Based on the sound part combination parameters, a heterogeneous sound part set is selected from the preset sound part combination corresponding to the sound part registry. The heterogeneous sound part set is composed of models belonging to different transmission channels and mapped to different transmission actuators through the unified transmission dispatch table, forming an audit panel. Perform voice availability gating on the audit panel: determine whether the command line tools and keys required for each voice exist, and remove the voice if any are missing; if the audit panel after gating is empty, return to the empty panel failure state and end this call; otherwise, execute the following steps; Assign an audit identifier to this audit, first write it to the status file, then create a background audit task, and return the audit identifier and running status to the caller; The background audit task is executed in a subprocess separate from the service process of the audit orchestration device. The subprocess is started with an independent session, so that the background audit task is not canceled when the service process of the audit orchestration device exits or is recycled. The background audit task is dispatched to each voice part in parallel via the unified transmission dispatch table, and independent exception capture boundaries and voice part timeout limits are set for each voice part. This ensures that any exception or timeout of any voice part is captured at the boundary and set to a failure state, without being thrown upwards or canceling the parallel execution of the remaining voice parts. By separating the handle of the background audit task from the waiting operation through task masking, in response to the caller's continued query request carrying the audit identifier, a single-time limited wait is executed with a configurable timeout that is strictly less than the single call hard timeout and reserves the margin required to complete the processing and transmission of this response: if the background audit task has not reached the final state, the running state is returned and a prompt message indicating to continue to initiate continued queries is displayed, and the timeout of this single-time limited wait only ends the waiting operation without propagating cancellation to the background audit task; if the background audit task has reached the final state, the complete multi-part result is returned; The in-transit status and final result of this audit are stored persistently in two separate directories. The final result is written to the storage medium with the audit identifier as the key and retained for a preset lifespan, so that the final result can still be retrieved based on the audit identifier after the process of the background audit task is restarted.

[0008] In this technical solution, the selected heterogeneous sound part set is served by models belonging to different transmission channels and mapped to different transmission actuators through the unified transmission dispatch table. The composition of the audit panel is uniquely determined by the sound part combination parameters and the registration content of the sound part registry. The selection is not based on the identifier of the model that initiated this call as input.

[0009] In this technical solution, the availability gating of the acoustic part further includes: adding regional acoustic parts to the audit panel based on the existence of a key; wherein billing acoustic parts that require a dedicated key are also gated by removal when the key does not exist; the retention, removal and addition results of the gating are disclosed to the caller.

[0010] In this technical solution, the step of reading the final result from the storage medium based on the audit identifier further includes: determining whether the audited material exceeds a preset size level, and returning a failure state and disclosing the reason if it does; the determination is performed separately for the audited material and for the acoustic availability gating of the acoustic operating environment.

[0011] In this technical solution, after the parallel dispatch call, the method further includes: aggregating the independent decisions of each voice part into a parallel adjudication structure. When there are multiple voice parts, the top-level conclusion field of the parallel adjudication structure is left empty and specific findings are not merged at the top level. The independent decisions of each voice part are fully preserved in a parallel manner and are accompanied by adjudication instructions, panel size, completion count and failure count.

[0012] In this technical solution, the retrieval further includes: determining whether the delivery mark corresponding to the audit identifier exists; if it does not exist, reading the final result in the storage medium, returning the final result and setting the delivery mark; if it exists, intercepting the delivery.

[0013] In this technical solution, the transmission executor includes a command line subprocess executor and an interface client executor. The command line subprocess executor runs in a one-time empty directory allocated to the voice part, and the directory is discarded after use. The interface client executor only sends and receives text. When the unified transmission dispatch table encounters an unregistered transmission channel, it sets the corresponding voice part to a failure state without interrupting the parallel dispatch of the remaining voice parts.

[0014] In this technical solution, a restricted execution environment is imposed on at least one voice part running as a command-line subprocess. The restricted execution environment is constructed by the process isolation mechanism provided by the operating system, which denies file reading by default and only explicitly allows access to pre-configured whitelisted paths, thereby restricting the voice part subprocess from reading files outside its working directory.

[0015] In this technical solution, the timeout limit of the sound part is configured separately for each transmission channel, or uniformly configured for all sound parts.

[0016] In this technical solution, the receiving also includes receiving intent information, which is used to determine the assembly relationship between the prompt templates issued to each voice part and the output structure; when the intent is for defect review, the assembly includes the defect review prompt template and the multi-dimensional defect output field set; when the intent is for hypothesis demonstration, the assembly includes the hypothesis prompt template and the epistemological output field set containing falsifiability, null hypothesis, basis law and confidence update.

[0017] This invention also provides a parallel orchestration and non-blocking execution system with cross-model auditing, comprising: The registration verification module is used to complete the verification during the loading period of the voice registry, and to prevent the system from entering the service state that can receive calls when the verification fails. The receiving module is used to receive the audited materials and voice combination parameters submitted by the caller through the calling protocol when the system is in a service state that can receive calls. The calling protocol has a non-renewable hard timeout for a single call. The panel construction module, connected to the receiving module, is used to select a heterogeneous set of acoustic parts from the preset acoustic parts combination corresponding to the acoustic parts combination in the acoustic parts registry based on the acoustic parts combination parameters. The heterogeneous set of acoustic parts is composed of models belonging to different transmission channels and mapped to different transmission actuators through the unified transmission dispatch table, thus forming an audit panel. The gating module, connected to the panel construction module, is used to determine whether the command line tools and keys required for each voice part exist. If any are missing, the voice part will be removed. When the audit panel after gating is empty, an empty panel failure state will be output and the current call will be terminated. The task initialization module, connected to the gating module, is used to assign an audit identifier to the current audit when the audit panel behind the gating is not empty. It first writes the identifier to the status file, then creates a background audit task, and returns the audit identifier and running status to the caller. The parallel dispatch module, connected to the task initialization module, is used to enable the background audit task to be executed in a subprocess separate from the service process of the audit orchestration device. The background audit task dispatches calls to each voice part in parallel via the unified transmission dispatch table. The unified transmission dispatch table maps each voice part to the corresponding transmission executor according to the transmission channel registered in the voice part registry. It also sets independent exception capture boundaries and voice part timeout limits for each voice part, so that any exception or timeout of any voice part is captured at the boundary and set to a failure state, without being thrown upwards or canceling the parallel execution of the remaining voice parts. A non-blocking response module, connected to the task initialization module, is used to separate the handle of the background audit task from the waiting operation through task masking. In response to a follow-up query request initiated by the caller carrying the audit identifier, it executes a single-time limited wait with a configurable timeout strictly shorter than the single-call hard timeout and reserving the processing and transmission margin required to complete the current response. If the background audit task has not reached its final state, it returns a running state and a prompt indicating to continue initiating follow-up queries. The timeout of the single-time limited wait only ends the current waiting operation and does not propagate cancellation to the background audit task. If the background audit task has reached its final state, it returns a complete multi-part result. The persistence module, connected to the parallel dispatch module, the non-blocking response module, and the task initialization module, is used to persistently store the in-transit status and the final result in two separate directories. The final result is written to the storage medium with the audit identifier as the key and retained for a preset lifespan to support retrieval based on the audit identifier during cross-process restart.

[0018] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described parallel orchestration and non-blocking execution method for cross-model auditing.

[0019] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the above-described parallel orchestration and non-blocking execution method for cross-model auditing.

[0020] The present invention also provides a computer program product, including computer program code or instructions, which, when executed on a computer, enable the above-described cross-model auditing parallel orchestration and non-blocking execution method to be implemented.

[0021] The beneficial effects of this invention include: This invention shields the underlying transmission differences of heterogeneous voice parts through a unified transmission dispatch table, enabling voice parts from different transmission channels to be dispatched in parallel in the same way, thus achieving plug-and-play functionality for heterogeneous voice parts. Through single-voice fault isolation, any abnormality in one voice part does not affect the parallel aggregation of other voice parts, ensuring the robustness of the audit panel.

[0022] This invention enables reproducible degradation of the panel by gating the availability of the sound parts, avoiding the loss of the entire batch of results or the empty running of the billing sound parts when there are missing parts. By coupling the audit identifier, the protected background task and the single-time limited wait, the audit that exceeds the single call hard timeout can still retrieve the final result in stages, and the background audit task is not canceled when the waiting party times out.

[0023] This invention executes background auditing tasks in separate subprocesses, ensuring that the auditing tasks are not canceled when the service process exits or is recycled, directly addressing specific issues not addressed by general asynchronous task queues; it persists results to disk so that they can be retrieved across process restarts, and avoids duplicate deliveries through delivery deduplication; and it uses a parallel, non-merging adjudication support structure to ensure that discoveries made by only one voice are not discarded because other voices have not mentioned them, while the caller session with complete context acts as the adjudicator, avoiding same-source anchoring bias. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 A schematic flowchart illustrating the parallel orchestration and non-blocking execution method for cross-model auditing provided by this invention.

[0026] Figure 2 This is a schematic flowchart illustrating the overall process.

[0027] Figure 3 A schematic block diagram illustrating parallel dispatching and single-voice fault isolation.

[0028] Figure 4 A sequence diagram for audit identifiers, protected background tasks, and single restricted waits.

[0029] Figure 5 This is a schematic flowchart illustrating the process for determining the availability gating of acoustic parts and the grading of material dimensions.

[0030] Figure 6 This is a schematic diagram illustrating the deduplication status of dual directories on disk, cross-restart retrieval, and delivery.

[0031] Figure 7 A schematic diagram of the parallel orchestration and non-blocking execution system for cross-model auditing provided by the present invention.

[0032] Figure 8 This is a hardware structure diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0033] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0034] It should be noted that the following terms in this invention have the following meanings: In this invention, a "voice" refers to an independent audit execution unit that participates in the execution of an audit task. Each voice has independent audit roles, invocation entry points, independent exception capture boundaries, and output results. The voice can correspond to an artificial intelligence model, inference service, command-line tool, program component, or other execution unit capable of generating audit results.

[0035] The "voice registry," as used in this invention, refers to a data structure used to record the configurations of voices that can participate in auditing tasks. The voice registry records at least: voice identifier, voice type, invocation method, transmission channel, and output format. By querying the voice registry, voices participating in the current auditing task can be selected based on preset voice combinations.

[0036] Preset voice combination: refers to the voice combination that is pre-configured in the voice combination registry and is used to correspond to the voice combination parameters submitted by the caller.

[0037] Voice part combination parameters: These are the selection parameters submitted by the caller during the call, used to determine the set of voice parts required for this audit from the voice part registry.

[0038] In this invention, "heterogeneous voice set" refers to a collection of multiple voices with different execution characteristics. These differences can manifest in: different model types, different calling protocols, different runtime environments, and different toolchains.

[0039] The "audit panel" as used in this invention refers to an audit execution set composed of multiple voices selected from the voice registry according to the current audit requirements. The audit panel is used to describe the scope of voices involved in the execution of the current task and the corresponding execution configuration. The audit panel does not refer to a graphical interface displayed to the user.

[0040] The "audit identifier" as used in this invention refers to a data identifier used to uniquely associate an audit task. The audit identifier is used to associate: background audit task, voice part execution status, audit result, and persistent storage data. The caller obtains the corresponding audit status and result through the audit identifier.

[0041] Background auditing tasks, as described in this invention, refer to auditing execution tasks that can run independently even after the caller's request has been returned. Background auditing tasks have independent lifecycles and do not depend on the caller's current waiting process.

[0042] In this invention, "single-timed limited wait" refers to a configurable timeout period that is strictly less than the hard timeout of a single call in the calling protocol, and allows sufficient margin for processing and transmission of the current response. If the wait is not completed, it returns to a state where it can be continued; if it is completed, it returns immediately upon completion. The timeout of the single-timed limited wait only terminates the current wait operation.

[0043] Task shielding, as described in this invention, refers to a mechanism that isolates the lifecycle of a background auditing task from the waiting lifecycle of the caller. Through task shielding, the caller's own timeout, cancellation, or connection interruption will not propagate cancellation operations to the background auditing task.

[0044] Independent exception capture boundary, as described in this invention, refers to the exception and timeout capture boundary for a given voice part, such that any exception or timeout of a voice part is captured at this boundary and converted to a failure state, without being thrown upwards or canceling the parallel execution of other voice parts.

[0045] The parallel adjudication structure, as described in this invention, refers to a data organization structure in which the output results of multiple voice parts are stored and presented independently according to the identity of each voice part. This structure does not automatically perform majority fusion of the results from multiple voice parts, but rather retains the independent judgment results of each voice part.

[0046] Empty panel failure state: refers to the predetermined failure state returned when the audit panel after gating is empty.

[0047] Running state: refers to the state in which the background audit task has not yet reached the completion or failure final stage.

[0048] Final state: refers to the state where the background audit task has been written into the final result or has been determined to be a failure.

[0049] Complete multi-voice results: refers to the audit results that include the success or failure records of each selected voice.

[0050] Restricted execution environment: refers to the execution environment constructed by the process isolation mechanism provided by the operating system, which denies file reading by default and only explicitly allows access to pre-configured whitelisted paths.

[0051] This invention provides a parallel orchestration and non-blocking execution method for cross-model auditing, such as... Figure 1 As shown, the following explanation uses an audit orchestration device to illustrate this method.

[0052] Step 1: During the loading of the voice part registry, verify that the voice part identifiers referenced by each preset voice part combination in the voice part registry exist in the voice part registry, and that the transmission channels registered therein have corresponding transmission executors in the unified transmission dispatch table; if the verification fails, prevent the system from entering the service state that can receive calls.

[0053] The voice registry performs this verification during load; any undistributable entries prevent the system from entering a callable service state, rather than failing midway through the audit. This verification does not check for the existence of command-line tools and keys; those checks are performed by runtime gating.

[0054] Step 2: The audit orchestration device receives the audited material and the sound combination parameters. The audited material can be a code file, a text document, or other data to be audited.

[0055] The voice combination parameters are used to determine the set of voices required for this audit from the voice registry. Optionally, the audit orchestration device also receives intent information, which is used to determine the assembly of prompts and output structures. For example, if the intent is defect review, a defect review prompt and a multi-dimensional defect output structure are assembled; if the intent is hypothesis testing, a hypothesis prompt and an epistemological output structure are assembled. The risk level is consumed only as a prompt assembly parameter of this engine itself.

[0056] Step 3: The audit orchestration device selects a heterogeneous set of voice parts from the preset voice part combinations corresponding to the voice part combination in the voice part registry based on the voice part combination parameters, and forms the audit panel.

[0057] Step 4: Perform acoustic availability gating on the audit panel.

[0058] The system checks whether the required command-line tools and keys for each voice unit exist. If either is missing, the corresponding voice unit is removed from the audit panel. If the gated audit panel is empty, the system returns an empty panel failure state and terminates the call; otherwise, execution continues. Voice units that pass the above checks are retained, and regional voice units can be added based on the key. The retention, removal, or addition of gating results are disclosed to the caller. Billing voice units requiring a dedicated key are also gated by removal if the key is missing, to prevent the panel from crashing midway or a billing voice unit from running in vain.

[0059] Step 5: The audit orchestration device assigns an audit identifier to this audit, first writes it to the status file, then creates a background audit task, and returns the audit identifier and running status to the caller.

[0060] The audit identifier serves as a unique identifier for this audit task. The order of writing to the status file first, followed by creating the background audit task, is used to prevent concurrent race conditions. The audit identifier and running status are then immediately returned to the caller without waiting for the audit to complete.

[0061] In one implementation, the background audit task is executed in a subprocess separate from the service process of the audit orchestration device. This subprocess is started as an independent session, ensuring that the background audit task is not canceled when the service process of the audit orchestration device exits or is reclaimed. This measure directly addresses a specific problem that general asynchronous task queues do not face: the process hosting the orchestration engine itself may be reclaimed by the client, while the workers of general task queues are independent of the client's lifecycle and do not face this risk of consequential cancellation.

[0062] Step 6: The audit orchestration device dispatches and calls each part in the audit panel in parallel from the background audit task through a unified transmission and dispatch table.

[0063] The unified transmission dispatch table maps each voice segment's transmission channel to the corresponding transmission executor based on the channel's registration in the voice segment registry. Transmission executors include command-line subprocess executors and interface client executors. The command-line subprocess executor runs in a one-time empty directory, while the interface client executor only sends and receives text. Any unregistered transmission channel is considered a failure. Independent exception capture boundaries and configurable voice segment timeouts are set for each voice segment. This ensures that any exception in any voice segment, or if a voice segment fails to return within its timeout period, is captured at its independent exception capture boundary and marked as a failure for that voice segment. This failure is not propagated upwards, nor does it cancel the parallel execution of other voice segments. Therefore, the audit panel does not fail as a whole due to the failure of a single voice segment; the results of completed voice segments are returned intact.

[0064] Step 7: Separate the handle of the background audit task from the waiting operation by task masking; in response to the caller's follow-up query request with the audit identifier, execute a single restricted wait with a configurable time limit that is strictly less than the hard timeout of a single call and reserves the margin required to complete the processing and transmission of this response.

[0065] If the background audit task has not reached the final state, it returns to the running state and prompts the user to continue the query. The timeout of this single limited wait only ends the current wait operation and does not propagate the cancellation to the background audit task. The background audit task continues to run for the next query to retrieve the result. If the background audit task has reached the final state, it returns the complete multi-part result.

[0066] The caller initiates a single-time limited wait with a configurable timeout that is strictly less than the hard timeout of the calling protocol for a single call, and reserves the margin required to complete the processing and transmission of this response. If the audit is not completed, a status indicating that it is still running and a prompt message indicating that it should continue to initiate a follow-up query are returned, and the caller initiates a wait again accordingly; the complete multi-part result is returned the moment the audit is completed.

[0067] By separating the handle of the background audit task from the caller's waiting operation through task masking, the caller's own timeout is not propagated to the background audit task for cancellation. Therefore, even if a waiting call times out, the background audit task continues to run, ready for retrieval on the next subsequent query.

[0068] Step 8: The audit orchestration device will persistently store the in-transit status and the final result of this audit in two separate directories.

[0069] The final result is written to the storage medium with the audit identifier as the key and retained for a preset lifespan, so that the final result can still be retrieved based on the audit identifier after the process where the background audit task is located is restarted.

[0070] The audit's ongoing status and final results are stored in two separate directories. The ongoing directory records the audit status during operation, including real-time updates of the completion count, failure count, actual model and time consumed for each part, as well as panel size, completion count, and failure count. Results are written to storage media using the audit identifier as the key and retained for a preset lifespan. Upon reaching the final state, the audit is persistently written to the results directory via a completion callback.

[0071] The caller initiates retrieval using the audit identifier. During retrieval, the system checks if a delivered flag exists. If not, it reads the result from disk, returns the result, and sets the flag. If the flag exists, it intercepts duplicate deliveries. Since the in-transit state disappears with the process, while the result is independently stored with the audit identifier as the key, the caller can still retrieve the result using the audit identifier after the service process restarts or the session switches, without the original session being present.

[0072] like Figure 2As shown, the method generally includes: receiving the audited materials, voice part combination parameters, intent, and risk level; selecting a heterogeneous voice part set according to a self-verified registry based on the voice part combination parameters to form an audit panel; performing voice part availability gating on the audit panel; assigning audit identifiers, writing to a status file first, and then creating a background audit task; the background audit task being executed in a separate subprocess, and each voice part being dispatched in parallel via a unified transmission dispatch table; single-time restricted waiting; aggregation into a parallel adjudication structure; disk persistence, cross-restart retrieval, and delivery deduplication. The adjudication is made by the calling session that is not a member of the panel; this engine does not generate a final conclusion.

[0073] like Figure 3 As shown, in the audit panel behind the gate, the same audit material is handed over to each voice, and each voice is invisible to the others and has no order of precedence. Therefore, the judgment of the later voice will not be anchored by the conclusion of the earlier voice.

[0074] The unified transmission dispatch table maps the transmission channels registered in the registry by each voice unit to the corresponding transmission executor. The command-line subprocess executor runs in a one-time empty directory, and the interface client executor only sends and receives text; if an unregistered transmission channel is encountered, it is considered a failure without interrupting the dispatch.

[0075] When any voice part times out or encounters an anomaly, the anomaly is captured at the independent anomaly capture boundary of that voice part and transitions to a failure state. It is not thrown upwards or cancels the parallel execution of other voice parts. Therefore, the audit panel does not fail as a whole due to the failure of a single voice part. The results of completed voice parts are returned in their entirety, along with the panel size, completion count, and failure count.

[0076] The resulting structure is a parallel adjudication structure: in multi-voice scenarios, the conclusion field at the top level is left blank, and specific findings are not merged at the top level. The independent adjudications of each voice are fully preserved in a parallel manner, accompanied by adjudication instructions. Thus, a finding proposed by only one voice is delivered as usual and is not discarded simply because other voices did not mention it. It should be noted that the panel size is not a fixed value; it varies with gating degradation and runtime failures. Majority decisions based on a variable denominator are not comparable across different machines and at different times, and majority decisions can cause a finding proposed by only one voice to be drowned out by the silence of other voices. Therefore, majority voting is not based on the number of voices.

[0077] like Figure 4 As shown, after the caller session initiates an audit to the orchestration engine, the engine first writes a status file, then creates a background audit task and delivers it for execution, and immediately returns the audit flag and running status to the caller without blocking or waiting.

[0078] The caller then initiates a single-time limited wait with a configurable timeout that is strictly less than the hard timeout for a single call in the calling protocol, and allows sufficient margin for processing and transmitting the response. If the audit is not completed, a status indicating that the process is still running and a prompt to continue querying are returned, prompting the caller to initiate another wait. The wait is awakened the moment the audit is completed, and the complete multi-part result is returned. The timeout of the single-time limited wait only ends the current wait operation.

[0079] The key lies in task shielding: separating the handle of the background audit task from the protected waiting operation prevents the waiting operation's own timeout from being propagated to the background audit task for cancellation. Therefore, even if a waiting call times out, the background audit task continues to run, ready for retrieval on the next subsequent query.

[0080] The aforementioned audit identifier, protected background task, and single-time restricted wait are interdependent and indispensable: without the audit identifier, there is no way to continue querying; without task shielding, the waiting party will mistakenly kill the ongoing audit as soon as it times out; without single-time restricted wait, the single call will be killed by the client's hard timeout. Only through the coupling of these three elements can, within a call protocol with non-renewable hard timeouts, the audit that exceeds the single call hard timeout still be able to retrieve the final result in stages via the audit identifier.

[0081] like Figure 5 As shown, the following judgment is performed on each candidate voice part one by one. First, it is determined whether the material under review exceeds all preset size levels; if it does, a failure state is returned and the reason is disclosed. This is a strategy for the material under review, which is a different feature from the usability gating for voice parts described below.

[0082] Next, it is determined whether the command-line tool required for the audio section exists, and whether the required key exists; if either is missing, the audio section is removed from the panel. Billing audio sections requiring a dedicated key are also gated by removal when the key is unavailable, to prevent the panel from crashing midway or a billing audio section from running in vain.

[0083] The voice parts determined above shall be retained, and regional voice parts may be added according to the key; the retention, removal or addition of gating results shall be disclosed to the caller.

[0084] Finally, it checks if the panel after gating is empty; if it is empty, it returns an empty panel failure state and ends the call, instead of silently returning an empty result. Otherwise, it enters parallel transmission dispatch. In this way, the panel can be reproducibly degraded.

[0085] It is important to distinguish that gating removes the sound from the panel during selection, which is a different process from a sound entering the panel during operation and then abnormally turning into a failure state.

[0086] like Figure 6As shown, the audit status in progress and the final result are placed in two separate directories. The in-progress directory records the audit status during operation, including the completion count, failure count, actual model and time consumed updated in real time for each part, as well as the panel size, completion count and failure count. After the audit reaches the final state, it is persisted to the result directory through the completion callback.

[0087] The results directory is written to the storage medium using the audit identifier as the key and retained for a preset lifespan. It stores the parallel decision structure, panel size, completion count, and failure count. After the lifespan expires, it is cleaned up and reclaimed. When the final state is first delivered to the caller, a "delivered" flag is set; the presence or absence of this flag is the deduplication criterion.

[0088] Because the in-transit status is retained independently with the audit identifier as the key after the process terminates, the caller can still initiate retrieval using the audit identifier after the service process restarts or the session switches, without the original session being present. During retrieval, the system checks if a delivered flag exists: if not, it reads the disk, returns the result, and sets the flag; if it exists, it intercepts duplicate deliveries. The lifespan and directory location are preset configurable values.

[0089] The following synergistic effects exist between the features of this invention, indicating that this solution is not a simple superposition of known methods: (i) The coupling of audit identifier, protected background task and single-time restricted wait: if any one of them is missing, it will either be killed by the client's single timeout or the waiting party will cancel the audit task when it times out. It is the coupling of the three that can prevent both timeout and false kill.

[0090] (ii) Coupling of load-period dispatchability verification and runtime key / command line gating: ensuring that "those selected for the panel must either be able to run or be explicitly downgraded", avoiding the loss of the entire batch of results or the empty running of billing due to midway failure.

[0091] (III) Single-voice fault isolation and parallel convergence coupling: Independent judgment of one voice part without affecting the other voice parts.

[0092] (iv) Coupling of variable panel size and parallel adjudication structure: Panel size changes due to gating degradation and operational failures. The majority decision based on a variable denominator is not comparable across different machines and at different times; moreover, majority decision can cause a discovery made by only one voice to be drowned out by the silence of the others. Therefore, "fault isolation / gating leading to variable panel size" and "parallel non-merging adjudication structure" are a pair of mutually conditional conditions. Removing either one will either render the other unnecessary or produce an unreproducible conclusion.

[0093] (v) Separating subprocess execution from task masking: The worker of the general asynchronous task queue is independent of the client's lifecycle and does not face the risk of cancellation due to the "process carrying the orchestration engine being reclaimed by the client". Separating subprocess execution directly addresses this specific issue and, together with task masking, constitutes an independent lifecycle guarantee for background audit tasks.

[0094] Removing any one of the couplings degrades the other—therefore, compared to the simple superposition of "separately known parallel distribution, background task polling, and key-based service selection," it has a non-obvious overall technical effect.

[0095] In an improvement process, an early design would mark the entire audit task as failed if a voice part reported an error or failure, wasting the results of voice parts that had already been completed normally. This was later changed to only mark the voice part as failed and include it in the failure count, while other results were used normally. Initially, the audit voice parts were fixed; due to rapid iteration of large models, this was changed to a pluggable panel, allowing for optimization of participating voice parts based on effectiveness and cost. To avoid the same model auditing its own output, an initial design followed the caller's voice parts, resulting in adaptations or additional checks on each client. Furthermore, testing revealed that even context-free models from the same source could provide effective findings during auditing. Therefore, this was changed to a caller-independent, context-free dialogue for all voice parts. It was found that some models took tens of minutes to return after being called via the application programming interface, making inference time unpredictable and uncontrollable. Therefore, a configurable waiting time limit was set, and the process was actively terminated if exceeded. Local command-line calls were used, but some models lacked local clients; therefore, the interface format was incorporated into the design, ultimately achieving compatibility between the two. The choice of a dual-directory file system instead of a database for persistence is to facilitate the review of audit results by another AI inference engine and to facilitate the delivery of audit results for a single task in file format.

[0096] In one implementation, the restricted execution environment is applied selectively based on the operating platform: a built-in sandbox execution mechanism is used on a first-type operating system, and a built-in namespace isolation mechanism is used on a second-type operating system. For example, the sandbox-exec seatbelt configuration file is used on macOS, and the bubblewrap namespace isolation mechanism is used on Linux. These measures only reduce the local read surface. The one-time empty directory applies to voices running as command-line subprocesses; remote voices invoked via interfaces have their request boundaries constrained separately.

[0097] The technical solution of the present invention will be illustrated below with several examples.

[0098] Example 1: Taking the audit of the requested voice part combination parameters as an example, the engine obtains the voice part set from the self-verified voice part registry; if the command-line tool required by a certain voice part is not installed or the key does not exist, the voice part is removed from the panel; the engine then returns the audit flag and running status without waiting for the audit to complete. The panel after gating enters the parallel dispatch stage.

[0099] Example 2: Taking a single call to the protocol with a non-renewable hard timeout as an example, the engine performs a deep audit lasting several minutes in a protected background task. The caller repeatedly initiates a single-time restricted wait with a configurable timeout that is strictly less than the hard timeout and reserves the processing and transmission margin required to complete the current response. For the first few times, the engine receives a prompt message indicating that the process is still running and instructing the caller to continue initiating a follow-up query. The audit is completed the moment the current round of waiting ends and a complete multi-part result is returned. During this period, if the waiting call itself times out, the background audit task is not canceled.

[0100] Example 3: Taking an anomaly in a certain voice part of the panel as an example, if the command-line tool required for that voice part is missing or the interface reports an error, that voice part will be marked as failed and added to the failure count. The results of the remaining voice parts will be returned in parallel as usual, and the panel will not fail as a whole. When aggregating, the top-level conclusion field is set to empty, and each voice part is independently adjudicated and retained in parallel.

[0101] Example 4: Taking service process restart as an example, the caller reads back the audit data from the persistent storage via the retrieval interface using the audit identifier; repeated retrievals of the same final state are intercepted by the deduplication flag.

[0102] Example 5: Taking an isolation sandbox as an example, the command-line subprocess runs in a one-time empty directory. Its readable scope is limited to this empty directory and pre-configured whitelisted paths, excluding project source code and keys. Another subprocess is further constrained by its execution environment, defaulting to denying file reading and only explicitly allowing access to pre-configured whitelisted paths, thus restricting its subprocess from reading files outside its working directory. The one-time empty directory applies to subprocesses running as command-line subprocesses; remote subprocesses invoked via APIs have their request boundaries further constrained.

[0103] Example 6: Taking multi-voice adjudication as an example, the same audited material has four voices on the panel when running on machine A. On machine B, where a certain voice command-line tool is not installed, that voice is removed by gating, and the panel has three voices. During operation, one voice interface reports an error and fails, and only two voices are actually completed. The number of voices involved differs in each of the three scenarios. In multi-voice scenarios, the engine always sets the top-level conclusion field to empty, submits the adjudications of each voice in parallel, and attaches the panel size, completion count, and failure count. Therefore, it does not give contradictory top-level conclusions due to changes in the number of participants; findings made by only one voice are delivered as is with its independent adjudication.

[0104] Example 7: Taking intent switching as an example, when the intent is hypothesis argumentation, the output is epistemological fields such as strengths, risks, counterarguments, hypothesis and falsifiability, null hypothesis, base rate, and confidence update; when the intent is defect review, the output is multidimensional defect discovery.

[0105] This invention also provides a parallel orchestration and non-blocking execution system for cross-model auditing, such as... Figure 7 As shown, the system includes a registration verification module, a receiving module, a panel construction module, a gating module, a task initialization module, a parallel dispatch module, a non-blocking response module, and a persistence module.

[0106] A registration verification module is used to complete the verification during the loading period of the voice registry, and to prevent the system from entering a service state that can receive calls if the verification fails; wherein... The receiving module is used to receive the audited materials and voice combination parameters submitted by the caller through the calling protocol when the system is in a service state that can receive calls. The calling protocol has a non-renewable hard timeout for a single call. The panel construction module, connected to the receiving module, is used to select a heterogeneous set of acoustic parts from the preset acoustic parts combination corresponding to the acoustic parts combination in the acoustic parts registry based on the acoustic parts combination parameters. The heterogeneous set of acoustic parts is composed of models belonging to different transmission channels and mapped to different transmission actuators through the unified transmission dispatch table, thus forming an audit panel. The gating module, connected to the panel construction module, is used to determine whether the command line tools and keys required for each voice part exist. If any are missing, the voice part will be removed. When the audit panel after gating is empty, an empty panel failure state will be output and the current call will be terminated. The task initialization module, connected to the gating module, is used to assign an audit identifier to the current audit when the audit panel behind the gating is not empty. It first writes the identifier to the status file, then creates a background audit task, and returns the audit identifier and running status to the caller. The parallel dispatch module, connected to the task initialization module, is used to enable the background audit task to be executed in a subprocess separate from the service process of the audit orchestration device. The background audit task dispatches calls to each voice part in parallel via the unified transmission dispatch table. The unified transmission dispatch table maps each voice part to the corresponding transmission executor according to the transmission channel registered in the voice part registry. It also sets independent exception capture boundaries and voice part timeout limits for each voice part, so that any exception or timeout of any voice part is captured at the boundary and set to a failure state, without being thrown upwards or canceling the parallel execution of the remaining voice parts. A non-blocking response module, connected to the task initialization module, is used to separate the handle of the background audit task from the waiting operation through task masking. In response to a follow-up query request initiated by the caller carrying the audit identifier, it executes a single-time limited wait with a configurable timeout strictly shorter than the single-call hard timeout and reserving the processing and transmission margin required to complete the current response. If the background audit task has not reached its final state, it returns a running state and a prompt indicating to continue initiating follow-up queries. The timeout of the single-time limited wait only ends the current waiting operation and does not propagate cancellation to the background audit task. If the background audit task has reached its final state, it returns a complete multi-part result. The persistence module, connected to the parallel dispatch module, the non-blocking response module, and the task initialization module, is used to persistently store the in-transit status and the final result in two separate directories. The final result is written to the storage medium with the audit identifier as the key and retained for a preset lifespan to support retrieval based on the audit identifier during cross-process restart.

[0107] Optionally, the above modules can implement their respective functions through hardware structures, software modules, or a combination of hardware structures and software modules. The integrated modules can be implemented in hardware, as software functional modules, or as functional modules combining hardware and software.

[0108] The orchestration process of this method can also be used in other scenarios where multiple execution units need to process the same material in parallel and independently.

[0109] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims. It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the present invention is defined by the appended claims rather than the foregoing description. Therefore, it is intended that all variations falling within the meaning and scope of the equivalents of the claims be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0110] Figure 8 The embodiments of this specification illustrate a hardware structure diagram of an electronic device for a cross-model auditing parallel orchestration and non-blocking execution method. For example... Figure 8As shown, the electronic device may include at least one processor, a memory (e.g., non-volatile memory), a RAM, and a communication interface, and the at least one processor, memory, RAM, and communication interface are connected together via a bus. At least one processor executes at least one computer-readable instruction stored or encoded in the memory.

[0111] It should be understood that the computer-executable instructions stored in memory, when executed, cause at least one processor to perform the above-described combinations in the various embodiments of this specification. Figure 1-7 The description includes various operations and functions.

[0112] In the embodiments of this specification, electronic devices may include, but are not limited to: personal computers, server computers, workstations, desktop computers, laptop computers, notebook computers, mobile electronic devices, smartphones, tablet computers, cellular phones, personal digital assistants (PDAs), handheld systems, messaging devices, wearable electronic devices, consumer electronic devices, etc.

[0113] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0114] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the above-described parallel orchestration and non-blocking execution method for cross-model auditing.

[0115] In one embodiment, the computer-readable storage medium includes a read-only memory (ROM), random access memory (RAM), magnetic disk, optical disk, or solid-state storage device. The computer program contains a sequence of instructions corresponding to the above method. When the sequence of instructions is loaded into the computer and executed by the processor, the computer performs the following: verifying that each voice part can be assigned during the voice part registry loading period; receiving the audited material and voice part combination parameters; selecting a heterogeneous voice part set and performing voice part availability gating; allocating audit identifiers and creating background audit tasks; having the background audit task execute in separate subprocesses, and using a unified transmission dispatch table to dispatch calls to each voice part in parallel, setting independent exception capture boundaries and voice part timeout limits; implementing single-time restricted waiting through task masking; and persistently storing the transit status and final result in two separate directories.

[0116] The present invention also provides a computer program product, characterized in that it includes computer program code or instructions, which, when executed on a computer, enable the above-mentioned parallel orchestration and non-blocking execution method for cross-model auditing to be implemented.

[0117] In one implementation, the computer program product is distributed as a software installer, executable file, or source code. When the computer program code or instructions are executed on a computer, the computer performs all the steps of the parallel orchestration and non-blocking execution method for cross-model auditing, including: separating the handle of the background auditing task from the waiting operation through task masking; responding to a follow-up query request initiated by the caller carrying an audit identifier; executing a single restricted wait with a configurable timeout strictly less than the hard timeout of a single call, and reserving the processing and transmission margin required to complete the current response; and the timeout of the single restricted wait itself is not propagated to the background auditing task for cancellation.

[0118] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for parallel orchestration and non-blocking execution across models of auditing, characterized in that, This method is applied to an audit orchestration device that interacts with the caller via a calling protocol. The calling protocol has a non-renewable hard timeout for each single call. Before responding to any call, during the loading period of the voice part registry, the method verifies that the voice part identifiers referenced by each preset voice part combination in the voice part registry exist in the voice part registry, and that the registered transmission channels have corresponding transmission executors in the unified transmission dispatch table. This verification does not check the existence of command-line tools and keys. The following steps are executed only after the verification passes; if the verification fails, no further auditing services are provided. Receive the audited materials and sound part combination parameters submitted by the caller; Based on the sound part combination parameters, a heterogeneous sound part set is selected from the preset sound part combination corresponding to the sound part registry. The heterogeneous sound part set is composed of models belonging to different transmission channels and mapped to different transmission actuators through the unified transmission dispatch table, forming an audit panel. Perform voice availability gating on the audit panel: determine whether the command line tools and keys required for each voice exist, and remove the voice if any are missing; if the audit panel after gating is empty, return to the empty panel failure state and end this call; otherwise, execute the following steps; Assign an audit identifier to this audit, first write it to the status file, then create a background audit task, and return the audit identifier and running status to the caller; The background audit task is executed in a subprocess separate from the service process of the audit orchestration device. The subprocess is started with an independent session, so that the background audit task is not canceled when the service process of the audit orchestration device exits or is recycled. The background audit task is dispatched to each voice part in parallel via the unified transmission dispatch table, and independent exception capture boundaries and voice part timeout limits are set for each voice part. This ensures that any exception or timeout of any voice part is captured at the boundary and set to a failure state, without being thrown upwards or canceling the parallel execution of the remaining voice parts. By separating the handle of the background audit task from the waiting operation through task masking, in response to the caller's continued query request carrying the audit identifier, a single-time limited wait is executed with a configurable timeout that is strictly less than the single call hard timeout and reserves the margin required to complete the processing and transmission of this response: if the background audit task has not reached the final state, the running state is returned and a prompt message indicating to continue to initiate continued queries is displayed, and the timeout of this single-time limited wait only ends the waiting operation without propagating cancellation to the background audit task; if the background audit task has reached the final state, the complete multi-part result is returned; The in-transit status and final result of this audit are stored persistently in two separate directories. The final result is written to the storage medium with the audit identifier as the key and retained for a preset lifespan, so that the final result can still be retrieved based on the audit identifier after the process of the background audit task is restarted.

2. The parallel orchestration and non-blocking execution method for cross-model auditing according to claim 1, characterized in that, The composition of the audit panel is uniquely determined by the sound part combination parameters and the registration content of the sound part registry. The selection is not based on the identifier of the model that initiated this call as input. The models in the heterogeneous sound part set belong to different manufacturers.

3. The parallel orchestration and non-blocking execution method for cross-model auditing according to claim 1, characterized in that, The availability gating of the acoustic part further includes: adding regional acoustic parts to the audit panel based on the existence of a key; wherein billing acoustic parts that require a dedicated key are also gated by removal when the key does not exist; the retention, removal and addition results of the gating are disclosed to the caller.

4. The parallel orchestration and non-blocking execution method for cross-model auditing according to claim 1, characterized in that, Also includes: Determine whether the material under review exceeds a preset size classification; if it does, return a failure status and disclose the reason. The judgment is performed on the audited material and the availability gating of the acoustic part is performed separately for the acoustic part's operating environment.

5. The parallel orchestration and non-blocking execution method for cross-model auditing according to claim 1, characterized in that, Following the parallel dispatch call, the method further includes: aggregating the independent decisions of each voice part into a parallel adjudication structure. When there are multiple voice parts, the top-level conclusion field of the parallel adjudication structure is left empty and specific findings are not merged at the top level. The independent decisions of each voice part are fully preserved in a parallel manner and are accompanied by adjudication instructions, panel size, completion count and failure count.

6. The parallel orchestration and non-blocking execution method for cross-model auditing according to claim 1, characterized in that, The step of reading the final result from the storage medium based on the audit identifier further includes: determining whether the delivery mark corresponding to the audit identifier exists; if it does not exist, reading the final result in the storage medium, returning the final result and setting the delivery mark; if it exists, intercepting the delivery.

7. The parallel orchestration and non-blocking execution method for cross-model auditing according to claim 1, characterized in that, The transmission executor includes a command-line subprocess executor and an interface client executor. The command-line subprocess executor runs in a one-time empty directory allocated to the voice part, and the directory is discarded after use. The interface client executor only sends and receives text. When the unified transmission dispatch table encounters an unregistered transmission channel, it sets the corresponding voice part to a failure state without interrupting the parallel dispatch of the remaining voice parts.

8. The parallel orchestration and non-blocking execution method for cross-model auditing according to claim 7, characterized in that, A restricted execution environment is imposed on at least one voice part running as a command-line child process. The restricted execution environment is constructed by the process isolation mechanism provided by the operating system, which denies file reading by default and only explicitly allows access to pre-configured whitelisted paths, thereby restricting the voice part child process from reading files outside its working directory.

9. The method for parallel orchestration and non-blocking execution of cross-model auditing according to claim 1, characterized in that, The timeout limit for each sound part can be configured separately for each transmission channel, or uniformly configured for all sound parts.

10. The method for parallel orchestration and non-blocking execution of cross-model auditing according to claim 1, characterized in that, The receiving also includes receiving intent information, which is used to determine the assembly relationship between the prompt templates issued to each voice part and the output structure; when the intent is for defect review, the assembly includes the defect review prompt template and the multi-dimensional defect output field set; when the intent is for hypothesis demonstration, the assembly includes the hypothesis prompt template and the epistemological output field set containing falsifiability, null hypothesis, basicity and confidence update.

11. A parallel orchestration and non-blocking execution system with cross-model auditing, characterized in that, include: The registration verification module is used to complete the verification during the loading period of the voice registry, and to prevent the system from entering the service state that can receive calls when the verification fails. The receiving module is used to receive the audited materials and voice combination parameters submitted by the caller through the calling protocol when the system is in a service state that can receive calls. The calling protocol has a non-renewable hard timeout for a single call. The panel construction module, connected to the receiving module, is used to select a heterogeneous set of acoustic parts from the preset acoustic parts combination corresponding to the acoustic parts combination in the acoustic parts registry based on the acoustic parts combination parameters. The heterogeneous set of acoustic parts is composed of models belonging to different transmission channels and mapped to different transmission actuators through the unified transmission dispatch table, thus forming an audit panel. The gating module, connected to the panel construction module, is used to determine whether the command line tools and keys required for each voice part exist. If any are missing, the voice part will be removed. When the audit panel after gating is empty, an empty panel failure state will be output and the current call will be terminated. The task initialization module, connected to the gating module, is used to assign an audit identifier to the current audit when the audit panel behind the gating is not empty. It first writes the identifier to the status file, then creates a background audit task, and returns the audit identifier and running status to the caller. The parallel dispatch module, connected to the task initialization module, is used to enable the background audit task to be executed in a subprocess separate from the service process of the audit orchestration device. The background audit task dispatches calls to each voice part in parallel via the unified transmission dispatch table. The unified transmission dispatch table maps each voice part to the corresponding transmission executor according to the transmission channel registered in the voice part registry. It also sets independent exception capture boundaries and voice part timeout limits for each voice part, so that any exception or timeout of any voice part is captured at the boundary and set to a failure state, without being thrown upwards or canceling the parallel execution of the remaining voice parts. The non-blocking response module, connected to the task initialization module, is used to separate the handle of the background audit task from the waiting operation through task masking, and to respond to the follow-up query request initiated by the caller carrying the audit identifier, to perform a single restricted wait with a configurable time limit that is strictly less than the single call hard timeout and reserves the margin required to complete the processing and transmission of this response. When the background audit task has not reached the final state, it returns to the running state and prompts the user to continue the query. The timeout of the single limited wait only ends the current wait operation and does not propagate the cancellation to the background audit task. When the background audit task has reached the final state, it returns the complete multi-part result. The persistence module, connected to the parallel dispatch module, the non-blocking response module, and the task initialization module, is used to persistently store the in-transit status and the final result in two separate directories. The final result is written to the storage medium with the audit identifier as the key and retained for a preset lifespan to support retrieval based on the audit identifier during cross-process restart.

12. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the parallel orchestration and non-blocking execution method for cross-model auditing as described in any one of claims 1 to 10.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the parallel orchestration and non-blocking execution method for cross-model auditing as described in any one of claims 1 to 10.

14. A computer program product, characterized in that, It includes computer program code or instructions that, when executed on a computer, enable the parallel orchestration and non-blocking execution method for cross-model auditing as described in any one of claims 1 to 10 to be implemented.