A Failure Elevation Scheduling Method Based on Knowledge Conversion Hierarchy Criteria
By establishing a capability hierarchy diagram of the Chomsky formal language and independent verifiers, the problems of high resource consumption of knowledge transformation operators and lack of independence of verification conclusions in existing technologies are solved, realizing the controllability and traceability of resource management and verification process of computer task execution system.
Patent Information
- Application Number
- CN202611403038.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-10
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies lack a description of the capability boundaries of knowledge transformation operators in computer task execution systems, leading to repeated calls, invalid switching, and high resource consumption. Furthermore, it is difficult to determine the minimum sufficient capability level and subsequent upgrade direction based on the inherent complexity of the task, and the verification conclusions lack independence and verifiability.
A capability hierarchy graph based on the Chomsky formal language hierarchy is established. Through the capability hierarchy graph of knowledge transformation operators, the minimum sufficient capability level is determined and targeted upgrades are performed. An independent verifier is configured to generate structured failure signatures. Resource budget judgment is introduced to achieve controllability and traceability of task processing.
It reduces resource consumption, decreases the probability of repeated calls and invalid switching, improves the reliability and traceability of task processing, and provides a resource management and verification process based on computational theory.
Smart Images

Figure CN123114767A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer task scheduling and knowledge processing technology, and in particular to a failure escalation scheduling method based on knowledge conversion hierarchy criteria. Background Technology
[0002] With the development of natural language processing, knowledge reasoning, and intelligent task scheduling technologies, computer task execution systems are typically configured with multiple knowledge transformation operators for format conversion, structure parsing, constraint solving, difference recognition, and semantic generation. Because different tasks vary in input type, structural complexity, contextual dependencies, and verifiability of results, the processing capabilities and resource consumption of different knowledge transformation operators also differ.
[0003] Existing technologies typically select knowledge transformation operators based on preset rules, model scores, or historical success rates, and then re-invoke the original operator or switch to another operator after execution failure. This approach relies heavily on fixed sequences or empirical rules, lacking a unified description of the operator's capability boundaries. In particular, existing technologies often fail to establish a correspondence between regular, context-independent, context-dependent, and unrestricted capability levels and different operator categories based on the Chomsky formal language hierarchy, making it difficult to determine the minimum sufficient capability level and subsequent upgrade directions based on the inherent complexity of the task.
[0004] Furthermore, existing technologies typically determine task success based on execution anomalies, result confidence levels, or simple verification status. In statistical confidence routing schemes, candidate results and their confidence evaluations may be generated by the same or similar models, making verification sources and error sources homogeneous and difficult to provide independent and verifiable evidence for verification conclusions. Simultaneously, existing technologies lack structured processing of failure types, counterexample information, and unverified constraints, making it difficult to select target operators based on specific capability gaps and easily leading to problems such as duplicate calls, invalid switching, or premature invocation of resource-intensive operators.
[0005] Some existing systems lack effective constraints on resource budgets such as execution time, memory usage, accelerator processor usage, and model call consumption. For general completeness verification involving the nontrivial semantic properties of general computable programs, the verification problem may fall within the undecidable range revealed by Rice's theorem. Existing solutions typically lack corresponding undecidability criteria and still mainly rely on confidence thresholds or the number of retries to determine whether to transfer to manual processing. In addition, existing systems also have difficulty fully recording each operator execution and verification process, which is not conducive to manual review and execution auditing.
[0006] To address this, a failure escalation scheduling method based on knowledge conversion hierarchy criteria is proposed. Summary of the Invention
[0007] The main objective of this invention is to provide a failure escalation scheduling method based on knowledge conversion hierarchy criteria, which can effectively solve the problems in the background art.
[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0009] A failure escalation scheduling method based on knowledge conversion hierarchy criteria, applied to a computer task execution system, includes the following steps:
[0010] S1. Establish a capability hierarchy graph for knowledge transformation operators based on the Chomsky formal language hierarchy. The capability hierarchy graph includes a Type 3 regular capability hierarchy, a Type 2 context-independent capability hierarchy, a Type 1 context-dependent capability hierarchy, and a Type 0 unrestricted capability hierarchy. The four preset categories of knowledge transformation operators correspond one-to-one with the four capability hierarchies and form a bijective relationship. The capability hierarchy graph includes multiple capability nodes and directed upgrade edges connecting different capability nodes. The directed upgrade edges point from capability nodes with lower capability hierarchies to capability nodes with higher capability hierarchies. Each capability node is configured with a knowledge transformation operator, an independent verifier, and an operator capability descriptor. The knowledge transformation operator is an executable processing unit with an input constraint interface and an output data structure. The independent verifier is a verification program that is independent of the output generation process of the corresponding knowledge transformation operator and verifies the output of the operator according to a preset verification predicate.
[0011] S2. Obtain the task to be processed and its resource budget, parse the natural language description, task metadata and formal representation of the task to be processed, and generate task constraint vector and verification requirement set;
[0012] S3. Based on the coverage relationship between the task constraint vector and each operator capability descriptor, and the inclusion relationship between the verification requirement set and the verification predicate set corresponding to each independent verifier, candidate capability nodes are determined; the lowest capability level corresponding to the candidate capability node is determined as the minimum sufficient capability level of the task to be processed; the capability node with the lowest expected resource consumption is selected from the candidate capability nodes corresponding to the minimum sufficient capability level as the initial capability node, and the initial capability node is set as the current capability node.
[0013] S4. Call the knowledge conversion operator configured in the current capability node to process the task to be processed, obtain the operator output, operator execution data and actual resource consumption, and update the remaining resource budget according to the actual resource consumption; the corresponding independent verifier verifies the operator output according to the verification requirement set, generates a verification result including verification status, failure type, counterexample information and unverified constraint set, and generates the execution credential for this execution according to the current operator execution data and verification data, and appends the execution credential for this execution to the execution credential chain;
[0014] S5. When the verification status indicates that the verification has failed, the failure type, counterexample information and unverified constraint set are combined into a structured failure signature. The capability gap set corresponding to the structured failure signature is determined according to the preset correspondence between failure type and capability gap. The target capability node is determined according to the capability gap set, directed upgrade edge, remaining resource budget and already visited capability nodes.
[0015] S6. When the target capability node exists and the preset termination condition is not met, update the target capability node to the current capability node;
[0016] S7. Repeat steps S4 to S6 until a verified task processing result is obtained or a preset termination condition is met; when a verified task processing result is obtained, output the task processing result and execution credential chain; when the current capability node belongs to the highest capability level and the verification requirement meets the preset undecidability criterion, or meets other preset termination conditions, stop automatic upgrade, and transfer the pending task, each operator output, verification result and execution credential chain to the manual review channel.
[0017] Preferably, the task constraint vector is represented as follows:
[0018]
[0019] in, This indicates that the task to be processed is in the... Constraint features in each constraint dimension The constraint dimensions include at least the input representation type, structural memory requirement, recursion depth, element dependency span, context constraints, minimum difference search range, semantic openness, and result verifiability.
[0020] No. The operator capability descriptor for each capability node is represented as:
[0021] in, Indicates the first The capability node in the ... The set of allowed values in each constraint dimension This represents the set of validation predicates for the corresponding independent validator. This represents the set of capability types that the corresponding knowledge transformation operator can process. Indicates the ability level value. Indicates the expected resource consumption. The verifiable decisionability flag is used to distinguish between mechanically complete decisionability, partially decisionability, and non-mechanically complete decisionability.
[0022] The four categories of knowledge transformation operators include: a first category of knowledge transformation operators for processing finite state structures, which corresponds to the third type of regularization capability level; a second category of knowledge transformation operators for processing stack-like nested structures, which corresponds to the second type of context-independent capability level; a third category of knowledge transformation operators for processing bounded context constraints, which corresponds to the first type of context-dependent capability level; and a fourth category of knowledge transformation operators for processing unrestricted rules and open semantics, which corresponds to the zeroth type of unrestricted capability level.
[0023] Preferably, the first is determined according to the following formula. Capability coverage determination value for each capability node:
[0024]
[0025] in, This represents the set of verification requirements. This represents an indicator function, which takes the value 1 when the condition inside the parentheses is true, and takes the value 0 otherwise.
[0026] The set of candidate capability nodes is determined according to the following formula:
[0027]
[0028] in, This represents the set of candidate capability nodes.
[0029] Preferred, the first The comprehensive resource cost of each candidate capability node is determined according to the following formula:
[0030]
[0031] in, , , and These represent the normalized estimated execution time, estimated peak memory usage, estimated accelerator processor usage, and estimated model call consumption, respectively. , , and It is a non-negative resource weight, and satisfies:
[0032]
[0033] The initial capability node is determined according to the following formula:
[0034]
[0035]
[0036] in, This represents the lowest capability level value among the candidate capability nodes, and the lowest capability level value corresponds to the minimum sufficient capability level. This refers to the initial capability node.
[0037] Preferred, the first The structured failure signature corresponding to a failed verification is as follows:
[0038]
[0039] in, Indicates the first Failure type when the second verification fails This indicates the corresponding counterexample information. This represents the counterexample encoding obtained by performing structured encoding on the counterexample information. This represents the set of unverified constraints.
[0040] Based on the preset correspondence between failure types and capability gaps, the structured failure signatures are mapped to a set of capability gaps. ;
[0041] The number is determined according to the following formula. The set of upgradable capability nodes when the second verification fails:
[0042]
[0043] in, Indicates the first The current capability node corresponding to the first failed verification. This represents the set of directed upgrade edges in the capability hierarchy graph. Indicates the first Remaining resource budget when the second verification fails This indicates that the expected resource consumption in each resource dimension will not exceed the corresponding remaining resource budget. This represents the set of capability nodes that the current task has already visited. Represents a set of upgradable capability nodes;
[0044] The target capability node is determined according to the following formula:
[0045]
[0046] in, This refers to the target capability node. Represents capability nodes The overall resource cost.
[0047] Preferably, the fourth type of knowledge conversion operator corresponding to the type 0 unrestricted capability level includes a semantic analogy knowledge conversion operator, wherein the semantic analogy knowledge conversion operator adopts a dual-channel fusion structure;
[0048] The first channel is a structure mapping channel, used to generate structure mapping results based on the structural correspondence between entities, attributes, and relationships between entities in the task to be processed; the second channel is a pluggable semantic channel, used to call the semantic model through a preset calling interface and generate semantic processing results based on domain semantic information; both the structure mapping results and the semantic processing results are appended with channel source flags and channel confidence levels;
[0049] When both channels pass the corresponding verification and the outputs of both channels meet the preset consistency conditions, the dual-channel enhancement mode is adopted; when the structure mapping channel passes the corresponding verification but the pluggable semantic channel fails the corresponding verification or is unavailable, the structure channel priority mode is adopted; when the structure mapping channel fails the corresponding verification or is unavailable, but the pluggable semantic channel passes the corresponding verification, the semantic channel independent fallback mode is adopted; when both channels fail the corresponding verification, or the outputs of both channels do not meet the preset consistency conditions and cannot be resolved according to the preset conflict resolution rules, a verification failure status is generated and the process is transferred to the manual review channel.
[0050] Preferably, when the output of the knowledge transformation operator corresponding to the current capability node is the minimum difference set, the corresponding independent verifier generates a verification result and a minimumity verification credential corresponding to the minimum difference set, uses the verification result and the minimumity verification credential as components of the execution credential, and appends the execution credential to the execution credential chain.
[0051] Preferably, the resource budget includes at least the execution time budget, the accelerator processor usage budget, the model call budget, the operator call count budget, and the peak memory limit;
[0052] The preset termination conditions include at least one of the following: there is no target capability node that meets the capability gap set; the remaining resource budget is insufficient to execute the next capability node; the actual peak memory usage exceeds the peak memory limit; the number of upgrades reaches the preset upgrade limit; or the current capability node belongs to the highest capability level and the verification requirement meets the preset undecidability criterion.
[0053] The preset undecidability criterion includes a semantic attribute criterion based on Rice's theorem. The semantic attribute criterion is used to identify verification requirements where the verification object is a non-trivial semantic attribute of a function computed by a general computable program, and a general complete determination of the non-trivial semantic attribute is required. The non-trivial semantic attribute is an output semantic attribute that is not satisfied by all general computable programs and is not satisfied by all general computable programs.
[0054] The execution credential chain includes multiple execution credentials arranged in execution order. Each execution credential includes at least a task digest value, a capability node identifier, a capability level identifier, a knowledge conversion operator version, an independent verifier version, an input digest value, an output digest value, a verification result, actual resource consumption, and a digest value of the previous execution credential.
[0055] A failure escalation scheduling system based on knowledge conversion hierarchy criteria includes:
[0056] The capability hierarchy graph management module is used to establish and maintain the capability hierarchy graph of knowledge transformation operators. The capability hierarchy graph is established based on the Chomsky formal language hierarchy, including Type 3 regular capability hierarchy, Type 2 context-independent capability hierarchy, Type 1 context-dependent capability hierarchy, and Type 0 unrestricted capability hierarchy. The four preset categories of knowledge transformation operators form a one-to-one bijective relationship with the four capability levels. The capability hierarchy graph includes multiple capability nodes and directed upgrade edges from capability nodes with lower capability level values to capability nodes with higher capability level values. Each capability node is configured with a knowledge transformation operator, an independent verifier, and an operator capability descriptor.
[0057] The task constraint parsing module is used to parse the tasks to be processed and generate task constraint vectors and a set of verification requirements.
[0058] The node determination module is used to determine the minimum sufficient capability level of the task to be processed based on the task constraint coverage relationship, the verification requirement inclusion relationship, the capability level value and the expected resource consumption, determine the initial capability node from the candidate capability nodes corresponding to the minimum sufficient capability level, and set the initial capability node as the current capability node.
[0059] The operator execution module is used to call the knowledge conversion operator corresponding to the current capability node to process the task to be processed, and generate operator execution data and actual resource consumption;
[0060] The independent verification module is used to verify the operator output according to the verification predicate corresponding to the current knowledge transformation operator, and generate verification results including verification status, failure type, counterexample information and unverified constraint set;
[0061] The failure escalation module is used to generate a structured failure signature based on the verification result, determine a set of capability gaps based on the preset correspondence between failure types and capability gaps, and determine a target capability node based on the set of capability gaps, directed upgrade edges, and remaining resource budget, and update the target capability node to the current capability node; the failure escalation module is also used to stop automatic escalation and transfer the pending task to the manual review channel when the current capability node belongs to the highest capability level and the verification requirement meets the preset undecidability criterion;
[0062] The resource budget module is used to record actual resource consumption, update the remaining resource budget, and determine whether the preset termination conditions are met.
[0063] The credential chain generation module is used to generate the execution credential for this execution based on the corresponding operator execution data and verification data after each knowledge transformation operator execution and independent verification is completed, and to append the execution credential for this execution to the execution credential chain in the execution order.
[0064] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method steps.
[0065] Compared with the prior art, the present invention has the following beneficial effects:
[0066] 1. This invention establishes a bijective correspondence between knowledge transformation operators and Chomsky's formal language hierarchy, mapping the processing capabilities of different knowledge transformation operators to Type 3 regularization, Type 2 context-independent, Type 1 context-dependent, and Type 0 unrestricted capability levels. Based on task constraints, a minimum sufficient capability level is determined, providing a formal language theoretical basis for the initial selection of capability nodes and failure escalation, reducing unnecessary resource consumption caused by directly calling high-capability-level operators. Under the aforementioned exemplary simulation conditions, compared to all tasks fixedly calling the highest Type 0 capability level, using the minimum sufficient capability level and performing targeted escalation after verification failure reduces the normalized total resource cost by 58%.
[0067] 2. This invention generates a structured failure signature based on the failure type, counterexample information, and unverified constraint set, and maps the structured failure signature to a capability gap set. Then, it selects a target capability node that can fill the corresponding capability gap and meet the remaining resource budget along the directed upgrade edge, thereby realizing targeted upgrade based on the failure reason and reducing the probability of repeated calls and invalid switching.
[0068] 3. This invention configures independent verifiers for different capability nodes, separate from the knowledge conversion operator output generation process, which can reduce verification bias caused by the knowledge conversion operator's self-generation and self-judgment. For the highest capability level semantic analogy knowledge conversion operator, dual-channel processing is performed through a structure mapping channel and a pluggable semantic channel. Based on the channel verification status and output consistency, a structure channel priority, dual-channel enhancement, or independent semantic channel fallback mode is selected to improve the reliability and traceability of open semantic task processing.
[0069] 4. This invention incorporates resource budgeting, including execution time, memory usage, accelerator processor usage, model call consumption, and operator call counts, into the capability node selection and termination judgment process. It also stops automatic escalation and transfers to manual review for verification requirements that meet preset undecidability criteria. This ensures that the transfer of undecidable tasks has a computational theoretical basis, rather than relying solely on confidence thresholds or retries. Simultaneously, by recording the execution data, verification data, and capability level flags of each capability node through an execution credential chain, it improves resource controllability, result traceability, and the efficiency of manual review during task execution. Attached Figure Description
[0070] Figure 1 This is a flowchart of the failure escalation scheduling method based on knowledge conversion hierarchy criteria of the present invention;
[0071] Figure 2 This is a block diagram of the failure escalation scheduling system based on knowledge conversion hierarchy criteria of the present invention;
[0072] Figure 3 This is a schematic diagram illustrating the bijective correspondence between the Chomsky formal language hierarchy and the knowledge transformation operator categories in this invention.
[0073] Figure 4 This is a schematic diagram illustrating the temporal relationship between the failure upgrade chain and the execution credential chain of this invention;
[0074] Figure 5 The semantic analogy operator of this invention A schematic diagram of the dual-channel fusion structure. Detailed Implementation
[0075] 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.
[0076] Example 1
[0077] like Figure 1-5As shown, a failure escalation scheduling method based on knowledge conversion hierarchy criteria is disclosed. This method is implemented by a computer task execution system, which includes a processor, memory, a task receiving interface, an operator runtime environment, and a verification runtime environment. The operator runtime environment is used to load and execute knowledge conversion operators, while the verification runtime environment, independent of the operator runtime environment, is used to execute independent verifiers. The computer task execution system may further include a resource monitoring unit, a credential storage unit, and a manual review interface, used to collect actual resource consumption during operator execution, store execution credential chains, and receive pending tasks transferred to the manual review channel, respectively.
[0078] The knowledge transformation operator can be a finite state transition program, a stack-based structure parser, a constraint solver, a graph structure processor, a pre-trained neural network model, or other executable program with a structured output interface. Each knowledge transformation operator is configured with an input constraint interface and an output data structure to limit the types of inputs that can be received and the data format of the output results.
[0079] This embodiment includes the following steps.
[0080] S1. Establish a capability hierarchy diagram for knowledge transformation operators.
[0081] Establish a capability hierarchy diagram in a computer task execution system:
[0082]
[0083] in, Represents a set of capability nodes. This represents the set of directed upgrade edges.
[0084] Each capability node corresponds to a knowledge transformation operator, an independent verifier, and an operator capability descriptor. Directed upgrade edges point from capability nodes with lower capability level values to capability nodes with higher capability level values. The capability hierarchy graph is preferably a directed acyclic graph.
[0085] This embodiment establishes four capability levels arranged from low to high based on the Chomsky formal language hierarchy, and establishes a one-to-one bijective relationship between the four knowledge transformation operator categories and the four capability levels, specifically including:
[0086] The first type of knowledge transformation operator is a pattern matching operator. Corresponding to the Type 3 regularization capability level, it is used to handle tasks that do not require stack memory through finite state structures, regular rules, or deterministic field patterns;
[0087] The second type of knowledge transformation operator is the rule induction operator. This corresponds to the Type 2 context-independent capability level and is used to process nested tasks through stack-based parsing, recursive descent, or context-independent production rules.
[0088] The third type of knowledge transformation operator is the minimum difference recognition operator. This corresponds to the Type 1 context-related capability level and is used to handle tasks with cross-element dependencies, context constraints, or minimum difference search requirements.
[0089] The fourth type of knowledge transformation operator is the semantic analogy operator. Corresponding to the Type 0 Unrestricted Capability Level, it is used to handle tasks such as unrestricted rule deduction, open semantic generation, or tasks lacking complete formal representation.
[0090] The aforementioned bijective relationship refers to the one-to-one correspondence between the four knowledge transformation operator categories and the four capability levels, rather than limiting each capability level to only one operator instance. Multiple capability nodes with different versions, processing scopes, or resource consumption can be configured under the same knowledge transformation operator category.
[0091] To avoid the type number and capability level value in the basic level of the Choms being inversely related, this embodiment can set the capability level values of the Type 3 regular capability level, the Type 2 context-independent capability level, the Type 1 context-dependent capability level, and the Type 0 unrestricted capability level to 1, 2, 3, and 4, respectively.
[0092] Directed upgrade edges can be set between the first and second ability levels, between the second and third ability levels, and between the third and fourth ability levels. Alternatively, directed upgrade edges can be set between the first and third ability levels or between the second and fourth ability levels, depending on the actual ability inclusion relationships.
[0093] No. The operator capability descriptor for each capability node is represented as:
[0094]
[0095] in, Indicates the first The capability node in the ... The set of allowed values in each constraint dimension This represents the set of validation predicates for the corresponding independent validator. This represents the set of capability types that the corresponding knowledge transformation operator can process. Indicates the ability level value. Indicates projected resource consumption; This indicates the verifiability flag, and its value is used to distinguish between mechanically complete, partially complete, and non-mechanically complete verifiability.
[0096] The independent verifier operates independently of the output generation process of the corresponding knowledge transformation operator. For example, the knowledge transformation operator generates the task processing result in the first running process, and the independent verifier reads the task processing result in the second running process and verifies it according to a preset verification predicate. The independent verifier does not use the intermediate state or output score of the knowledge transformation operator as the verification conclusion.
[0097] S2. Generate task constraint vectors and verification requirement set.
[0098] Obtain the tasks to be processed, task metadata, formal representation, and resource budget through the task receiving interface.
[0099] In a specific example, the task to be processed is: based on the input structured configuration data, identify the set of minimum configuration differences that can cause changes in the target processing result. The task to be processed includes a natural language description, task metadata, formal representation, and resource budget. The task metadata includes the input file type, data size, number of fields, and task identifier; the formal representation includes the input data pattern, field dependencies, result constraints, and minimumity constraints; the resource budget includes the execution time budget, peak memory limit, accelerator processor usage budget, model call budget, and operator call count budget.
[0100] The above data is parsed to generate a task constraint vector:
[0101]
[0102] in, This indicates that the task to be processed is in the [number]th [stage]. Constraint features across several constraint dimensions. In this embodiment, the constraint dimensions include input representation type, structural memory requirement, recursion depth, element dependency span, contextual constraints, minimum difference search range, semantic openness, and result verifiability.
[0103] For example, the input representation of the task to be processed is a JSON structure, the maximum recursion depth is 4, the element dependency span is 12 fields, a minimum difference search is required, the semantic openness is low, and the verifiability of the result is high.
[0104] Simultaneously, a set of verification requirements is generated based on the task's output requirements:
[0105]
[0106] in, This represents a data structure validation predicate. This represents the constraint consistency verification predicate. This indicates a validation predicate for the result difference. This indicates a minimality verification predicate.
[0107] When the task to be processed contains open semantic processing or general program semantic verification requirements, the verification object type, verification attribute type and verification scope are also extracted from the task description and formal representation to determine whether the corresponding verification requirement meets the preset undecidability criterion.
[0108] S3. Determine the initial capability nodes.
[0109] For each capability node in the capability hierarchy diagram, the capability coverage determination value is calculated according to the following formula:
[0110]
[0111] in, This represents the indicator function. The indicator function takes the value 1 when the condition within the parentheses is true; otherwise, it takes the value 0.
[0112] This determines the set of candidate capability nodes:
[0113]
[0114] A capability node will not be included in the candidate capability node set if it cannot handle the input representation type of the task to be processed, or if its independent validator cannot execute any of the necessary verification predicates in the set of verification requirements.
[0115] For each capability node that enters the candidate capability node set, its estimated execution time, estimated peak memory usage, estimated accelerator processor usage, and estimated model call consumption are obtained. These data can be used for prediction based on the historical execution records of the corresponding knowledge conversion operator, benchmark test results, and the current task's data scale.
[0116] No. The comprehensive resource cost of each candidate capability node is calculated according to the following formula:
[0117]
[0118] in, , , and These represent the normalized estimated execution time, estimated peak memory usage, estimated accelerator processor usage, and estimated model call consumption, respectively. , , and It is a non-negative resource weight, and satisfies:
[0119]
[0120] For example, if the importance of execution time, peak memory usage, accelerator processor usage, and model call consumption are set to 0.35, 0.25, 0.20, and 0.20 respectively, then:
[0121]
[0122] First, determine the lowest capability level value among the candidate capability nodes according to the following formula:
[0123]
[0124] The capability level corresponding to the lowest capability level value is determined as the minimum sufficient capability level of the task to be processed. The minimum sufficient capability level refers to the lowest capability level that satisfies the task constraint coverage relationship and the verification requirement inclusion relationship.
[0125] Then, the initial capability node is determined from the candidate capability nodes of the lowest capability level according to the following formula:
[0126]
[0127] Therefore, this embodiment first determines the lowest capability level that meets the sufficient conditions, and then selects the capability node with the lowest overall resource cost within that capability level, rather than directly selecting the node with the highest capability or the lowest resource cost from all capability nodes.
[0128] For example, the task to be processed has cross-field dependencies and a minimum difference search requirement, and both the third and fourth capability levels satisfy the coverage relationship. Since the capability level value of the third capability level is lower than that of the fourth capability level, the third capability level is determined as the minimum sufficient capability level. If the third capability level includes multiple capability nodes, then the comprehensive resource cost of one of them is selected. The smallest capability node is used as the initial capability node, and the initial capability node is set as the current capability node.
[0129] S4. Execute the current capability node and perform independent verification.
[0130] The system invokes the knowledge transformation operator corresponding to the current capability node to process the pending task, obtaining the operator output, operator execution data, and actual resource consumption. The operator execution data includes the current capability node identifier, knowledge transformation operator version, input summary value, output summary value, start time, and end time; the actual resource consumption includes the actual execution time, actual peak memory usage, actual accelerator processor usage, actual model call consumption, and actual number of operator calls.
[0131] No. The remaining resource budget before the next execution is represented as follows:
[0132] in, Indicates the first Budget for remaining execution time prior to the next execution; Indicates peak memory limit; This indicates the remaining accelerator processor usage budget; Indicates the remaining model call budget; This represents the budget for the remaining number of operator calls.
[0133] Let the first The actual execution time, actual accelerated processor usage, and actual model call consumption for each execution were respectively: , and Then the first The remaining resource budget after this execution is updated as follows:
[0134]
[0135] Peak memory limits are not cumulatively deducted; instead, the actual peak memory usage for this instance is factored in. Compare them.
[0136] After resource recording is complete, the operator output is sent to the corresponding independent validator. The independent validator performs verification according to the verification predicates in the verification requirement set, generating a verification result that includes verification status, failure type, counterexample information, and the set of unverified constraints. The actual resource consumption is obtained by the operator execution process and the resource monitoring unit and is not included in the verification result generated by the independent validator.
[0137] Validation status includes validation passed and validation failed. Failure types can include input structure mismatch, nested structure parsing failure, cross-field constraint failure, minimumity validation failure, semantic ambiguity, validation incompleteness determination, and resource execution exception.
[0138] Counterexample information is used to indicate the specific input location, field, constraint, or candidate result that caused the validation to fail. For example, when a cross-field constraint is not satisfied, counterexample information may include the field identifier, the actual value of the field, the expected value of the field, and the dependency distance between related fields.
[0139] After each knowledge transformation operator execution and independent verification is completed, an execution credential is immediately generated and appended to the execution credential chain according to the execution order. The digest value of each execution credential is generated according to the following formula:
[0140]
[0141] in, Represents the summary value of the previous execution credential; Represents the task summary value; Indicates the current capability node identifier; Indicates the version of the knowledge transformation operator; Indicates the standalone validator version; Indicates the input summary value; Indicates the output summary value; Indicates the verification result; This indicates the actual resource consumption in this instance; This indicates a data connection operation.
[0142] Therefore, regardless of whether the task is ultimately verified or transferred to manual review due to meeting the termination conditions, the system has formed a complete chain of execution credentials including each execution and verification process.
[0143] S5. Generate structured failure signatures and determine target capability nodes.
[0144] When the verification status indicates that the verification failed, the first generation is generated. Structured failure signature when the second verification fails:
[0145]
[0146] in, Indicates the failure type. Indicates counterexample information, This represents the counterexample encoding obtained by structured encoding of the counterexample information. This represents the set of unverified constraints.
[0147] Counterexample coding This can include error codes, input path summary values, actual value types, expected value types, dependency spans, and constraint identifiers. For example, if a negative example indicates that a second capability node cannot handle cross-field dependencies with a 12-field gap, the corresponding dependency span and constraint identifier will be recorded in the negative example code.
[0148] The system pre-stores the correspondence between failure types and capability gaps. For example, nested structure parsing failure corresponds to a stack-style nesting processing capability gap, cross-field constraint failure corresponds to a bounded context constraint processing capability gap, minimumity verification failure corresponds to a minimum difference search capability gap, and semantic ambiguity corresponds to an open semantic generation capability gap.
[0149] Based on the aforementioned correspondence, the structured failure signatures are mapped to a set of capability gaps:
[0150]
[0151] No. If the initial verification fails, the set of upgradable capability nodes will be determined according to the following formula:
[0152]
[0153] in, Indicates the first The current capability node corresponding to the first failed verification. Indicates the first Remaining resource budget after the next execution. This represents the set of capability nodes that the current task has already visited. This indicates that the expected resource consumption for each resource dimension will not exceed the corresponding remaining resource budget.
[0154] The target capability nodes are determined according to the following formula:
[0155]
[0156] For example, if the current second-level capability node fails the verification because it cannot handle cross-field dependencies, and there are capability nodes in the third and fourth capability levels that can supplement the bounded context constraint processing capability and meet the remaining resource budget, then the third-level capability node with the lower overall resource cost is selected as the target capability node.
[0157] S6. Update current capability nodes
[0158] When the set of upgradable capability nodes is not empty and the preset termination condition is not met, the target capability node will be... Update to the current capability node to be used in the next execution, i.e.:
[0159]
[0160] Then, the process returns to step S4, where the updated knowledge transformation operator corresponding to the current capability node is invoked. This avoids re-invoking the initial capability node after an upgrade and ensures that each upgrade proceeds along the directed upgrade edges in the capability hierarchy graph towards a higher capability level.
[0161] S7, repeat validation and upgrade, and perform result processing.
[0162] Repeat steps S4 to S6. When the operator output of the current capability node passes independent verification, determine the operator output as the final task processing result, and output the final task processing result and the execution credential chain.
[0163] When there are no target capability nodes that meet the capability gap set and remaining resource budget, the remaining resource budget is insufficient, the actual peak memory usage exceeds the peak memory limit, or the number of upgrades reaches the preset upgrade limit, automatic upgrades will stop, and the pending tasks, operator outputs, verification results, and execution credential chains will be sent to the manual review channel.
[0164] When the current capability node belongs to the Type 0 Unrestricted Capability Level, further determine whether the verification requirement meets the preset undecidability criterion. The preset undecidability criterion includes at least the following conditions:
[0165] First, the object of verification is the output semantic properties of functions computed by general computable programs, rather than just the format, numerical range, or structural integrity of limited input data.
[0166] Second, the output semantic attribute is a non-trivial semantic attribute, that is, there exists at least one function computed by a general computable program that satisfies the semantic attribute, and at the same time there exists at least one function computed by a general computable program that does not satisfy the semantic attribute.
[0167] Third, the verification requirement requires a general completeness determination of whether any general computable program satisfies the aforementioned nontrivial semantic properties.
[0168] When the above conditions are met, according to Rice's theorem, the corresponding verification requirement is marked as not being completely determined by the mechanical verifier, the verification decidability flag is set to not being mechanically complete, automatic attempts based on confidence threshold or number of repeated calls are stopped, and the pending task and its execution credential chain are transferred to the manual review channel.
[0169] Not all Type 0 unrestricted capability level tasks are directly judged as non-mechanically complete. When the verification requirement can be transformed into a finite candidate set enumeration, finite sample replay, structural constraint checking, or other verification process with definite termination conditions, the verification is still completed by the corresponding independent verifier.
[0170] As another concrete example, the task to be processed is to convert a policy configuration file with multi-level nesting and cross-level reference relationships into a normalized rule tree. The initial formal representation of the task characterizes the multi-level nesting relationships, and based on this, the system determines the Type 2 context-independent capability level as the initial minimum sufficient capability level, and selects the corresponding stack-based nested structure processing operator as the initial capability node.
[0171] After the stack-based nested structure processing operator completes the nested structure parsing, the corresponding independent verifier detects unresolved cross-level identifier references between different nested branches, generates a structured failure signature including cross-level reference paths, failure types, and unverified reference constraints, and maps the structured failure signature to a bounded context constraint processing capability gap.
[0172] Based on the directed upgrade edge and remaining resource budget, the system selects the target capability node at the type 1 context-related capability level. The knowledge transformation operator corresponding to the target capability node handles cross-level reference constraints. After the output of the target capability node passes independent verification, the system outputs a normalized rule tree and generates execution credentials from the verification failure data of the type 2 capability node and the verification success data of the type 1 capability node, respectively, and appends them to the execution credential chain in the execution order.
[0173] In this example, the upgrade of a capability node is triggered by the capability gap determined by an independent validator, rather than by the statistical confidence score output by the knowledge transformation operator.
[0174] To illustrate the impact of minimum sufficient capability level scheduling on resource consumption, a set of exemplary simulations is conducted. 100 tasks are set up, with 40 tasks corresponding to capability levels 3, 30 to 2, 20 to 1, and 10 to 0. The normalized resource cost per transaction for the knowledge conversion operator at each capability level is set to 1, 2, 4, and 8, respectively. If all tasks consistently call the highest capability level (Type 0), the normalized total resource cost is 800.
[0175] When using the method of this embodiment, each task first calls the corresponding minimum sufficient capability level, with an initial normalized total resource cost of 260. Further, 20 tasks are configured to undergo a targeted upgrade due to failure of independent verification. Of these, 10 are upgraded from Type 3 to Type 2, 6 from Type 2 to Type 1, and 4 from Type 1 to Type 0. The upgrade increases the normalized resource cost by 76, resulting in a final normalized total resource cost of 336, a 58% reduction compared to all tasks fixedly calling the highest capability level of Type 0.
[0176] The above data are exemplary simulation results obtained based on preset task distribution and normalized resource costs, used to illustrate the resource consumption relationship under different scheduling methods, and do not represent actual test results on specific hardware platforms.
[0177] Example 2:
[0178] This embodiment uses the semantic analogy knowledge transformation operator corresponding to the Type 0 unrestricted capability level. The semantic analogy knowledge conversion operator described herein employs a dual-channel fusion structure consisting of a structural mapping channel and a pluggable semantic channel.
[0179] The structure mapping channel receives entities, attributes, and relationships between entities from the task to be processed, and generates a structure mapping result based on the correspondence between the source entity structure and the target entity structure. The structure mapping result includes at least the entity correspondence, relationship preservation status, unmapped structures, and structure mapping score.
[0180] The pluggable semantic channel connects to a large language model or other semantic processing model through a preset calling interface, generating semantic processing results based on the natural language description, domain context, and output constraints of the task to be processed. The preset calling interface specifies unified input fields, output fields, timeout parameters, and model version fields, enabling different semantic processing models to be replaced without changing the scheduling interface.
[0181] The structure mapping channel and the pluggable semantic channel output the channel source flag, channel processing result, and channel confidence score, respectively. An independent validator validates the outputs of both channels to obtain the structure channel validation status. and semantic channel verification status .
[0182] When both channels pass verification, the consistency between the outputs of the two channels is calculated according to a preset consistency function:
[0183]
[0184] in, This represents the result of the structure mapping. This represents the result of semantic processing. This represents the consistency function. When the consistency threshold is not lower than a preset consistency threshold... At that time, a dual-channel enhancement mode is used to fuse the outputs of the two channels.
[0185] Specifically, this embodiment selects the processing mode according to the following rules:
[0186] when , and When using dual-channel enhancement mode, the fused processing result is output, and the channel source flag is set to dual-channel fusion.
[0187] when If the pluggable semantic channel fails verification or is unavailable, the structure channel priority mode is adopted, and the structure mapping result is determined as the candidate processing result.
[0188] When the structure mapping channel fails validation or is unavailable, In this case, the semantic channel independent fallback mode is adopted to determine the semantic processing result as the candidate processing result;
[0189] When both channels fail verification, or when both channels pass verification but the output does not meet the preset consistency conditions and cannot be resolved by the preset conflict resolution rules, a verification failure status is generated, and the pending task, the output of the two channels, the channel source flag, and the execution credential chain are transferred to the manual review channel.
[0190] Channel confidence scores serve only as supplementary information for dual-channel fusion and manual review, and are not used as the sole basis for verification. Each channel output still needs to be verified by an independent validator according to preset verification predicates, thereby preventing semantic channels from generating results on their own and judging the results as correct based on their own confidence scores.
[0191] When the verification requirement corresponding to the semantic channel output meets the preset undecidability criterion, even if the channel confidence level given by the semantic channel reaches the preset value, the verification will not be judged as passed, but will instead be transferred to the manual review channel. The model version, channel source flag, verification result, and fusion mode of each channel are all written into the corresponding execution credentials.
[0192] Example 3:
[0193] This embodiment illustrates how the minimum difference set output is recorded in the execution credential chain. When the knowledge transformation operator corresponding to the current capability node outputs the minimum difference set, the minimum difference set is sent to the corresponding independent verifier, which then generates a verification result and a minimumity verification credential according to a preset verification predicate.
[0194] The minimality verification credential includes at least the current output digest value, verification status, verification time, and independent verifier version. The system uses the verification result and the minimality verification credential as components of the current execution credential and appends the current execution credential to the execution credential chain in the execution order.
[0195] When the verification result indicates that the verification has failed, the system generates a structured failure signature based on the corresponding failure type, counterexample information, and unverified constraints, and determines the target capability node according to the aforementioned failure escalation scheduling process.
[0196] The method for generating the minimum difference set and the specific minimumity verification mechanism can be implemented by an external processing component connected to this scheduling process. This embodiment only describes the recording and association of the verification results and minimumity verification credentials in the failure escalation scheduling process.
[0197] Example 4:
[0198] This embodiment also provides a failure escalation scheduling system based on knowledge conversion hierarchy criteria, including a capability hierarchy graph management module, a task constraint parsing module, a node determination module, an operator execution module, an independent verification module, a failure escalation module, a resource budget module, and a credential chain generation module.
[0199] The capability hierarchy graph management module stores capability nodes and directed upgrade edges in the form of adjacency lists, graph databases, or relational data tables, and records the knowledge transformation operators, independent verifiers, and operator capability descriptors corresponding to each capability node. The module also stores the bijective relationships between the four types of knowledge transformation operators and the four formal language capability levels, as well as the verifiability decidability flags corresponding to each capability node.
[0200] The task constraint parsing module receives the natural language description, task metadata, and formal representation of the task to be processed, and generates a task constraint vector and a set of verification requirements. For tasks involving open semantics or general program semantics verification, the task constraint parsing module also extracts the verification object type, verification attribute type, and verification scope.
[0201] The node determination module performs capability coverage determination and comprehensive resource cost calculation. First, it determines the minimum sufficient capability level that satisfies the coverage and inclusion relationships, and then determines the initial capability node from the candidate capability nodes corresponding to the minimum sufficient capability level.
[0202] The operator execution module loads the knowledge transformation operator corresponding to the current capability node in an independent running process or container and stores the operator output as a preset output data structure. The operator execution module also generates operator execution data and sends the actual resource consumption to the resource budget module. For semantic analogy-based knowledge transformation operators at the Type 0 unrestricted capability level, the operator execution module can configure a structure mapping channel and a pluggable semantic channel.
[0203] The independent verification module executes the verification predicate in a separate process or container from the operator execution module, generating verification status, failure types, counterexample information, and a set of unverified constraints. The independent verification module does not directly use the confidence level output by the knowledge transformation operator itself as the verification pass conclusion.
[0204] The failure escalation module generates a structured failure signature based on the verification results, determines the set of capability gaps, and identifies the target capability node from the capability nodes pointed to by the directed escalation edges. The failure escalation module updates the target capability node to the current capability node, and stops automatic escalation and calls the manual review interface when the current capability node belongs to the highest capability level and the verification requirements meet the preset undecidability criteria.
[0205] The resource budget module records the actual resource consumption for each operator execution, updates the remaining resource budget, and determines whether the preset termination conditions are met.
[0206] The credential chain generation module generates the execution credential after each operator execution and independent verification, and connects the execution credentials in the execution order. When the task processing result passes verification, the final task processing result and execution credential chain are output; when the preset termination condition is met, the task to be processed, the output of each operator, the verification result, and the execution credential chain are sent to the manual review terminal.
[0207] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.
Claims
1. A failure escalation scheduling method based on knowledge conversion hierarchy criteria, characterized in that, Applied to computer task execution systems, it includes the following steps: S1. Establish a capability hierarchy graph for knowledge transformation operators based on the Chomsky formal language hierarchy. The capability hierarchy graph includes a Type 3 regular capability hierarchy, a Type 2 context-independent capability hierarchy, a Type 1 context-dependent capability hierarchy, and a Type 0 unrestricted capability hierarchy. The four preset categories of knowledge transformation operators correspond one-to-one with the four capability hierarchies and form a bijective relationship. The capability hierarchy graph includes multiple capability nodes and directed upgrade edges connecting different capability nodes. The directed upgrade edges point from capability nodes with lower capability hierarchies to capability nodes with higher capability hierarchies. Each capability node is configured with a knowledge transformation operator, an independent verifier, and an operator capability descriptor. The knowledge transformation operator is an executable processing unit with an input constraint interface and an output data structure. The independent verifier is a verification program that is independent of the output generation process of the corresponding knowledge transformation operator and verifies the output of the operator according to a preset verification predicate. S2. Obtain the task to be processed and its resource budget, parse the natural language description, task metadata and formal representation of the task to be processed, and generate task constraint vector and verification requirement set; S3. Based on the coverage relationship between the task constraint vector and each operator capability descriptor, and the inclusion relationship between the verification requirement set and the verification predicate set corresponding to each independent verifier, candidate capability nodes are determined; the lowest capability level corresponding to the candidate capability node is determined as the minimum sufficient capability level of the task to be processed; the capability node with the lowest expected resource consumption is selected from the candidate capability nodes corresponding to the minimum sufficient capability level as the initial capability node, and the initial capability node is set as the current capability node. S4. Call the knowledge conversion operator configured in the current capability node to process the task to be processed, obtain the operator output, operator execution data and actual resource consumption, and update the remaining resource budget according to the actual resource consumption; the corresponding independent verifier verifies the operator output according to the verification requirement set, generates a verification result including verification status, failure type, counterexample information and unverified constraint set, and generates the execution credential for this execution according to the current operator execution data and verification data, and appends the execution credential for this execution to the execution credential chain; S5. When the verification status indicates that the verification has failed, the failure type, counterexample information and unverified constraint set are combined into a structured failure signature. The capability gap set corresponding to the structured failure signature is determined according to the preset correspondence between failure type and capability gap. The target capability node is determined according to the capability gap set, directed upgrade edge, remaining resource budget and already visited capability nodes. S6. When the target capability node exists and the preset termination condition is not met, update the target capability node to the current capability node; S7. Repeat steps S4 to S6 until a verified task processing result is obtained or a preset termination condition is met; when a verified task processing result is obtained, output the task processing result and execution credential chain; when the current capability node belongs to the highest capability level and the verification requirement meets the preset undecidability criterion, or meets other preset termination conditions, stop automatic upgrade, and transfer the pending task, each operator output, verification result and execution credential chain to the manual review channel.
2. The failure escalation scheduling method based on knowledge conversion hierarchy criteria according to claim 1, characterized in that, The task constraint vector is represented as follows: in, This indicates that the task to be processed is in the... Constraint features in each constraint dimension The constraint dimensions include at least the input representation type, structural memory requirement, recursion depth, element dependency span, context constraints, minimum difference search range, semantic openness, and result verifiability. No. The operator capability descriptor for each capability node is represented as: in, Indicates the first The capability node in the ... The set of allowed values in each constraint dimension This represents the set of validation predicates for the corresponding independent validator. This represents the set of capability types that the corresponding knowledge transformation operator can process. Indicates the ability level value. Indicates the expected resource consumption. The verifiable decisionability flag is used to distinguish between mechanically complete decisionability, partially decisionability, and non-mechanically complete decisionability. The four categories of knowledge transformation operators include: a first category of knowledge transformation operators for processing finite state structures, which corresponds to the third type of regularization capability level; a second category of knowledge transformation operators for processing stack-like nested structures, which corresponds to the second type of context-independent capability level; a third category of knowledge transformation operators for processing bounded context constraints, which corresponds to the first type of context-dependent capability level; and a fourth category of knowledge transformation operators for processing unrestricted rules and open semantics, which corresponds to the zeroth type of unrestricted capability level.
3. The failure escalation scheduling method based on knowledge conversion hierarchy criteria according to claim 2, characterized in that, The number is determined according to the following formula. Capability coverage determination value for each capability node: in, This represents the set of verification requirements. This represents an indicator function, which takes the value 1 when the condition inside the parentheses is true, and takes the value 0 otherwise. The set of candidate capability nodes is determined according to the following formula: in, This represents the set of candidate capability nodes.
4. The failure escalation scheduling method based on knowledge conversion hierarchy criteria according to claim 3, characterized in that, No. The comprehensive resource cost of each candidate capability node is determined according to the following formula: in, , , and These represent the normalized estimated execution time, estimated peak memory usage, estimated accelerator processor usage, and estimated model call consumption, respectively. , , and It is a non-negative resource weight, and satisfies: The initial capability node is determined according to the following formula: in, This represents the lowest capability level value among the candidate capability nodes, and the lowest capability level value corresponds to the minimum sufficient capability level. This refers to the initial capability node.
5. The failure escalation scheduling method based on knowledge conversion hierarchy criteria according to claim 4, characterized in that, No. The structured failure signature corresponding to a failed verification is as follows: in, Indicates the first Failure type when the second verification fails This indicates the corresponding counterexample information. This represents the counterexample encoding obtained by performing structured encoding on the counterexample information. This represents the set of unverified constraints. Based on the preset correspondence between failure types and capability gaps, the structured failure signatures are mapped to a set of capability gaps. ; The number is determined according to the following formula. The set of upgradable capability nodes when the second verification fails: in, Indicates the first The current capability node corresponding to the first failed verification. This represents the set of directed upgrade edges in the capability hierarchy graph. Indicates the first Remaining resource budget when the second verification fails This indicates that the expected resource consumption in each resource dimension will not exceed the corresponding remaining resource budget. This represents the set of capability nodes that the current task has already visited. Represents a set of upgradable capability nodes; The target capability node is determined according to the following formula: in, This refers to the target capability node. Represents capability nodes The overall resource cost.
6. The failure escalation scheduling method based on knowledge conversion hierarchy criteria according to claim 2, characterized in that, The fourth type of knowledge conversion operator corresponding to the type 0 unrestricted capability level includes a semantic analogy knowledge conversion operator, which adopts a dual-channel fusion structure; The first channel is the structure mapping channel, which is used to generate structure mapping results based on the structural correspondence between entities, attributes and relationships between entities in the task to be processed. The second channel is a pluggable semantic channel, which is used to call the semantic model through a preset calling interface and generate semantic processing results based on domain semantic information; both the structure mapping result and the semantic processing result are attached with a channel source flag and a channel confidence level. When both channels pass the corresponding verification and the outputs of both channels meet the preset consistency conditions, the dual-channel enhancement mode is adopted; when the structure mapping channel passes the corresponding verification but the pluggable semantic channel fails the corresponding verification or is unavailable, the structure channel priority mode is adopted; when the structure mapping channel fails the corresponding verification or is unavailable, but the pluggable semantic channel passes the corresponding verification, the semantic channel independent fallback mode is adopted; when both channels fail the corresponding verification, or the outputs of both channels do not meet the preset consistency conditions and cannot be resolved according to the preset conflict resolution rules, a verification failure status is generated and the process is transferred to the manual review channel.
7. The failure escalation scheduling method based on knowledge conversion hierarchy criteria according to claim 1, characterized in that, When the output of the knowledge transformation operator corresponding to the current capability node is the minimum difference set, the corresponding independent verifier generates a verification result and a minimumity verification credential corresponding to the minimum difference set. The verification result and the minimumity verification credential are used as components of the execution credential for this execution and are appended to the execution credential chain.
8. The failure escalation scheduling method based on knowledge conversion hierarchy criteria according to claim 1, characterized in that, The resource budget includes at least the execution time budget, accelerator usage budget, model call budget, operator call count budget, and peak memory limit; The preset termination conditions include at least one of the following: there is no target capability node that satisfies the capability gap set; or the remaining resource budget is insufficient to execute the next capability node. The actual peak memory usage exceeds the peak memory limit; the number of upgrades reaches the preset upgrade limit; Alternatively, the current capability node belongs to the highest capability level and the verification requirement satisfies the preset undecidability criterion; The preset undecidability criterion includes a semantic attribute criterion based on Rice's theorem. The semantic attribute criterion is used to identify verification requirements where the verification object is a non-trivial semantic attribute of a function computed by a general computable program, and a general complete determination of the non-trivial semantic attribute is required. The non-trivial semantic attribute is an output semantic attribute that is not satisfied by all general computable programs and is not satisfied by all general computable programs. The execution credential chain includes multiple execution credentials arranged in execution order. Each execution credential includes at least a task digest value, a capability node identifier, a capability level identifier, a knowledge conversion operator version, an independent verifier version, an input digest value, an output digest value, a verification result, actual resource consumption, and a digest value of the previous execution credential.
9. A failure escalation scheduling system based on knowledge conversion hierarchy criteria, characterized in that, include: The capability hierarchy graph management module is used to establish and maintain the capability hierarchy graph of knowledge transformation operators. The capability hierarchy graph is established based on the Chomsky formal language hierarchy, including Type 3 regular capability hierarchy, Type 2 context-independent capability hierarchy, Type 1 context-dependent capability hierarchy, and Type 0 unrestricted capability hierarchy. The four preset categories of knowledge transformation operators form a one-to-one bijective relationship with the four capability levels. The capability hierarchy graph includes multiple capability nodes and directed upgrade edges from capability nodes with lower capability level values to capability nodes with higher capability level values. Each capability node is configured with a knowledge transformation operator, an independent verifier, and an operator capability descriptor. The task constraint parsing module is used to parse the tasks to be processed and generate task constraint vectors and a set of verification requirements. The node determination module is used to determine the minimum sufficient capability level of the task to be processed based on the task constraint coverage relationship, the verification requirement inclusion relationship, the capability level value and the expected resource consumption, determine the initial capability node from the candidate capability nodes corresponding to the minimum sufficient capability level, and set the initial capability node as the current capability node. The operator execution module is used to call the knowledge conversion operator corresponding to the current capability node to process the task to be processed, and generate operator execution data and actual resource consumption; The independent verification module is used to verify the operator output according to the verification predicate corresponding to the current knowledge transformation operator, and generate verification results including verification status, failure type, counterexample information and unverified constraint set; The failure escalation module is used to generate a structured failure signature based on the verification result, determine a set of capability gaps based on the preset correspondence between failure types and capability gaps, and determine a target capability node based on the set of capability gaps, directed upgrade edges, and remaining resource budget, and update the target capability node to the current capability node; the failure escalation module is also used to stop automatic escalation and transfer the pending task to the manual review channel when the current capability node belongs to the highest capability level and the verification requirement meets the preset undecidability criterion; The resource budget module is used to record actual resource consumption, update the remaining resource budget, and determine whether the preset termination conditions are met. The credential chain generation module is used to generate the execution credential for this execution based on the corresponding operator execution data and verification data after each knowledge transformation operator execution and independent verification is completed, and to append the execution credential for this execution to the execution credential chain in the execution order.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method described in any one of claims 1 to 8.
