A dynamic workflow generation and execution method based on boundary constraint and expert experience filtering

Abstract

The application provides a dynamic workflow generation and execution method based on boundary constraint and expert experience filtering, comprising the following steps: firstly, receiving candidate and boundary constraint, compiling predicate to exclude feasible set F and residual set R; then, marking the candidate in R as a gray area candidate, which meets the boundary feasibility and satisfies the gray area condition; then, calling an expert experience filtering module, executing field rule filtering, and making cost-sensitive decision in conflict, and outputting reservation / reduced weight / pending result; then, constructing a DSL abstract syntax tree based on F, filtering result, tool capability graph, resource budget and stop / rollback condition, and completing type and reachability verification; compiling the above process into an executable DSL workflow, and outputting a reply, evidence reference and trajectory field and constructing an evidence graph. The application takes boundary compilation, gray area hard decision, experience filtering and DSL abstract syntax tree compilation and verification as strong constraint processes, so that the expert experience is converted from text suggestion into executable, replayable and auditable decision link, and engineering usability is improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent orchestration and executable workflow generation technology in the material research and development process, and belongs to a dynamic workflow generation and execution method based on boundary constraints and expert experience filtering. Background Technology

[0002] Materials research and development tasks typically involve high-dimensional variables, strongly coupled mechanisms, and resource-constrained experimental / simulation processes. The generation, screening, and validation of candidate solutions require multiple tools across quantum chemistry, molecular dynamics, machine learning, high-throughput screening, and experimental platforms. Fixed standard operating procedures (SOPs) are difficult to dynamically schedule based on differences in material systems, budget variations, and risk appetite.

[0003] Current constraint screening methods largely rely on physical, chemical, or mathematical boundaries to eliminate infeasible candidates. However, in engineering practice, there are numerous gray-zone candidates that are "boundary feasible but still have a high risk of failure": boundary constraints do not trigger elimination, but failure occurs due to factors such as narrow process windows, impurity sensitivity, interface side reactions, scale magnification, or measurement uncertainties. Relying solely on boundary constraints makes it difficult to make reproducible engineering decisions for these candidates. On the other hand, while relying solely on expert experience can effectively avoid engineering pitfalls, it often lacks computable field structures, evidence citations, conflict resolution, and replay mechanisms, making it easily perceived as subjective and difficult to audit. Furthermore, if the workflow only outputs suggestions in natural language, it is difficult to directly drive toolchain execution and form closed-loop evidence.

[0004] Therefore, an engineering solution is needed: to compile boundary constraints into executable predicates and output margins; to perform auditable expert experience filtering on gray zone candidates that cannot be eliminated from the boundary; and to compile the decision-making process into a DSL workflow with type validation and reachability validation, so as to achieve executable, replayable and auditable results. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a dynamic workflow generation and execution method based on boundary constraints and expert experience filtering.

[0006] To achieve the above objectives, the present invention provides the following technical solution: This application provides a dynamic workflow generation and execution method based on boundary constraints and expert experience filtering, including the following steps: S1: Receive the candidate set and the boundary constraint set, compile the boundary constraints in the boundary constraint set into executable predicates and perform joint screening on the candidate set to obtain the feasible set F and the residual set R; S2 marks candidates in the residual set R whose boundaries are feasible but satisfy the gray zone determination criteria as gray zone candidates. The gray zone determination criteria include at least one of margin, prediction uncertainty, evidence conflict degree or extrapolation degree exceeding its corresponding threshold. S3: Invoke the expert experience filtering module for gray area candidates, perform field-based experience rule filtering, and conduct cost-sensitive adjudication in case of rule conflicts, and output the results of retention, demotion, or pending review; S4: Construct a DSL abstract syntax tree based on the feasible set F, the output results of retention, demotion, or pending review, the tool capability graph, the resource budget, and the stop / rollback conditions, and perform type checking and reachability checking; S5: Compile the processes of S1 to S4 into an executable DSL workflow, and output a candidate experiment set including risk labels, evidence pointers, and playback indexes. The candidate experiment set drives the heterogeneous toolchain to execute the toolchain instruction sequence corresponding to the executable DSL workflow, and finally outputs a receipt, evidence reference, and trace field, and constructs an evidence graph based on the receipt, evidence reference, and trace field; S6: Based on the trace field and playback index output in S5, load historical baseline data from the storage. The historical baseline data includes receipts, evidence references, trace fields, and evidence graphs, replay and execute the corresponding DSL workflow, and perform differential comparison between the replay results and the historical baseline to generate a differential report including data, behavior, performance, and status differences for audit analysis and compliance verification. If the difference exceeds the preset threshold, an alarm is triggered.

[0007] Preferably, in step S1, the boundary constraint set includes physical boundaries, chemical boundaries, and mathematical boundaries; The physical boundary at least includes one or more of thermodynamic stability, phase equilibrium reachability, diffusion, or heat transfer upper limit; The chemical boundary at least includes one or more of side reaction window, impurity sensitivity, and interfacial compatibility; the mathematical boundary at least includes one or more of inequality constraints, resource budget constraints, or multi-objective feasible region constraints; The compilation of the executable predicate set includes normalizing the boundary expressions, unit normalization, and parameter binding, and outputting the margin vector of each candidate object.

[0008] Preferably, in step S2, the gray area determination conditions at least satisfy one of the following: 0 < margin < m0, prediction uncertainty > u0, evidence conflict degree of the candidate object in the evidence graph > K0, or extrapolation degree > ood0; where m0 is the upper limit of the allowable margin, u0 is the upper limit of the acceptable uncertainty, K0 is the conflict tolerance threshold, and ood0 is the boundary value for extrapolation discrimination. The conflict degree is calculated based on the weighted count or weighted difference of the support edges and contradiction edges in the evidence graph, and the weights include the evidence quality score and the time decay term. The evidence quality score is preset by an expert or automatically calculated, and the time decay term is calculated through a predefined decay function based on the evidence generation or last update time.

[0009] Preferably, in step S3, the expert experience filtering module rules adopt a conflict resolution strategy for conflicting experience rules. The conflict resolution strategy includes: weighting by confidence level, weighting by evidence freshness or quality, or cost-sensitive adjudication based on either false positive cost or false negative cost. The cost-sensitive adjudication selects the threshold or decision that minimizes the expected misjudgment cost, wherein the expected misjudgment cost includes a combination of false positive cost × false positive cost and false negative cost × false negative cost. The empirical rule record is a field-based structure, including: rule identifier, domain scope, triggering condition, adjudication standard, action template, evidence citation, confidence level, false positive cost, false negative cost, structure version, and rule version.

[0010] Preferably, the specific processing steps of the expert experience filtering module for field-based experience rule filtering and conflict resolution of gray area candidates are as follows: S31. Perform field matching on the gray zone candidates based on the fields in the fieldized empirical rule record; S32. Calculate the comprehensive score for each fielded empirical rule. The calculation formula is as follows: Comprehensive score = w1 × rule confidence + w2 × evidence quality score + w3 × time decay value; Among them, rule confidence is the preset credibility of the rule, evidence quality score is the reliability score of the evidence, time decay value is dynamically calculated according to the newness of the evidence, and w1, w2, and w3 are preset weight coefficients of the system, which are used to adjust the relative importance of the three indicators. S33. Detect whether there are conflicts among the fielded empirical rules. If there is a rule mapped to "pass" and another rule mapped to "reject" in the same field or condition, it is determined to be a conflict. Extract all conflicting rules, generate two candidate decisions: "pass" and "reject," and calculate the sum of the comprehensive scores of the rules supporting each decision. The rule with the higher score is taken as the final decision result for that field, and the conflict resolution status is marked. If the sum of the comprehensive scores is equal, expert rules are introduced for arbitration, and rules with high confidence and more recent evidence are given priority. All conflict handling processes and decision-making basis are recorded in the audit log for subsequent traceability and model optimization. S34. Based on the sum of the comprehensive scores of the conflict rules of the two candidate decisions, calculate the expected misjudgment cost and selection cost of each decision, add the two together to get the total cost, and select the decision with the smaller total cost as the final decision result; if the total costs are equal, decide according to the default security policy.

[0011] Preferably, in step S4, the tool capability diagram includes tool identifier, version, input / output mode, unit system, capability range and prerequisites, and interface mapping is performed based on the tool capability diagram; The DSL abstract syntax tree contains STEP nodes and GATE nodes, where STEP nodes contain tool identifiers, input parameters, and validation fields, and GATE nodes contain predicates, thresholds, costs, and branch pointers. The type verification includes: input parameter matching verification with input-output pattern verification, unit consistency verification, and evidence citation parsing verification. The reachability verification includes: branch pointer existence verification, dead-path-free path verification, and fallback path reachability verification.

[0012] Preferably, in step S5, the DSL workflow includes a stop or rollback field, the stop condition is generated by the budget and risk threshold, and the rollback field is used for rollback actions when the GATE is triggered or the verification fails. The receipt includes: execution status, output fields, evidence list, resource consumption, anomaly diagnosis and receipt hash. The receipt is written into the execution node in the evidence graph and points to the corresponding DSL step and input evidence object through derived edges.

[0013] A system for dynamic workflow generation and execution based on boundary constraints and expert experience filtering includes: The boundary compilation and filtering module is used to compile boundary constraints into executable predicates and perform joint filtering on candidates; The gray zone determination module performs gray zone determination on the residual set R based on margin, uncertainty, degree of evidence conflict and / or extrapolation, and sets the candidate labels of those that are feasible at the boundary but meet the gray zone determination conditions as gray zone candidates; The expert experience filtering module filters and sorts the residual set R based on field-based experience rules and outputs decisions and evidence citations. The tool capability graph module provides information on tool version, input / output modes, and capability range. Dynamic orchestrator and DSL compiler, used to build DSL AST and perform type and reachability checks to generate DSL workflow; The execution engine is used to execute step-by-step according to the DSL and output receipts, evidence, and trajectory fields.

[0014] Preferably, the expert experience filtering module includes a rule matching unit, a conflict resolution unit, and a cost-sensitive adjudication unit, and outputs suggested operations to supplement the characterization or adjust variables; The AST nodes generated by the DSL compiler include stop and fallback nodes, and the fallback path is verified to exist and be reachable in the reachability check; The system also includes a runtime monitor, which checks the verification results against the gating threshold and records the triggering reason and evidence reference when a STOP or FALLBACK is triggered. The system also includes a replay and differential audit module, which is used to load the historical parser based on the tracking identifier and DSL hash, replay the execution, and output the differential for auditing. The system also includes a consistency verification module, which verifies that the pattern version of the empirical rule record is consistent with the DSL field set.

[0015] Compared with existing technologies, this invention provides a dynamic workflow generation and execution method based on boundary constraints and expert experience filtering, which has the following beneficial effects: This invention improves engineering usability and inventiveness by using a strongly constrained process of "boundary compilation + gray zone hard judgment + experience filtering (including cost-sensitive adjudication) + DSL AST compilation and verification". This makes experience filtering no longer just a text suggestion, but an executable, replayable and auditable decision-making process.

[0016] The features and advantages of the present invention will be described in detail through embodiments and in conjunction with the accompanying drawings. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the method steps in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] See Figure 1 This invention provides a dynamic workflow generation and execution method based on boundary constraints and expert experience filtering, comprising the following steps: S1: Receive the candidate set and the boundary constraint set, compile the boundary constraints in the boundary constraint set into executable predicates and perform joint screening on the candidate set to obtain the feasible set F and the residual set R; S2: Candidates in the residual set R that are feasible at the boundary but meet the gray zone determination criteria are labeled as gray zone candidates. The gray zone determination criteria include at least one of margin, prediction uncertainty, evidence conflict degree or extrapolation degree exceeding its corresponding threshold. S3: Call the expert experience filtering module for gray zone candidates, perform field-based experience rule filtering, and make cost-sensitive decisions when rules conflict, outputting the results of retention, downgrading or pending review; S4: Construct a DSL abstract syntax tree based on the feasible set F, the output retention, demotion, or pending review results, the tool capability map, the resource budget, and the stop / rollback conditions, and perform type checking and reachability checking; S5: Compile the processes of S1 to S4 into an executable DSL workflow, and output a candidate experiment set including risk tags, evidence pointers, and playback indices. The candidate experiment set drives the heterogeneous toolchain to execute the toolchain instruction sequence corresponding to the executable DSL workflow, and finally outputs a receipt, evidence reference, and trace field, and constructs an evidence graph based on the receipt, evidence reference, and trace field; S6: Based on the trace field and playback index output by S5, load historical baseline data from the storage. The historical baseline data includes receipts, evidence references, trace fields, and evidence graphs, replay and execute the corresponding DSL workflow, and perform differential comparison between the replay result and the historical baseline to generate a differential report including data, behavior, performance, and status differences for audit analysis and compliance verification. If the difference exceeds the preset threshold, an alarm is triggered.

[0020] Among them, the boundary constraint set includes physical boundaries, chemical boundaries, and mathematical boundaries. The physical boundary includes at least one or more of thermodynamic stability, phase equilibrium reachability, diffusion, or heat transfer upper limit; The chemical boundary includes at least one or more of side reaction window, impurity sensitivity, and interfacial compatibility; the mathematical boundary includes at least one or more of inequality constraints, resource budget constraints, or multi-objective feasible region constraints; The compilation of the executable predicate set includes normalizing the boundary expression, unit normalization, and parameter binding, and outputting a margin vector for each candidate object.

[0021] Among them, the gray area determination condition satisfies at least one of the following: 0 < margin < m0, prediction uncertainty > u0, conflict degree of the candidate object in the evidence graph > K0, or extrapolation degree > ood0; Among them, m0 is the upper limit of the allowable margin, u0 is the upper limit of the acceptable uncertainty, K0 is the conflict tolerance threshold, ood0 is the demarcation value for extrapolation discrimination. The conflict degree is calculated based on the weighted count or weighted difference of the support edges and contradictory edges in the evidence graph. The weight includes at least the evidence quality score and the time decay term. The evidence quality score is preset by an expert or automatically calculated. The time decay term is calculated based on the evidence generation or last update time through a predefined decay function.

[0022] The expert experience filtering module uses conflict resolution strategies for conflicting experience rules. These conflict resolution strategies include: weighting by confidence level, weighting by evidence freshness or quality, or cost-sensitive adjudication based on either false positive or false negative costs. The cost-sensitive adjudication selects the threshold or decision that minimizes the expected cost of misjudgment. The expected cost of misjudgment includes a combination of false positive cost × false positive cost and false negative cost × false negative cost. The empirical rule record is a field-based structure, including: rule identifier, domain scope, triggering condition, adjudication standard, action template, evidence citation, confidence level, false positive cost, false negative cost, structure version, and rule version.

[0023] The specific processing steps of the expert experience filtering module for filtering gray zone candidates using field-based experience rules and resolving conflicts are as follows: S31. Perform field matching on the gray zone candidates based on the fields in the fieldized empirical rule record; S32. Calculate the comprehensive score for each fielded empirical rule. The calculation formula is as follows: Comprehensive score = w1 × rule confidence + w2 × evidence quality score + w3 × time decay value; Among them, rule confidence is the pre-defined level of credibility of the rules, and evidence quality score is the reliability score of the evidence. The time decay value is dynamically calculated based on the age of the evidence. w1, w2, and w3 are preset weighting coefficients used to adjust the relative importance of the three indicators. S33. Detect whether there are conflicts among the fielded empirical rules. If there is a rule mapped to "pass" and another rule mapped to "reject" in the same field or condition, it is determined to be a conflict. Extract all conflicting rules, generate two candidate decisions, "pass" and "reject", and calculate the sum of the comprehensive scores of the rules supporting each decision; S34. Based on the sum of the comprehensive scores of the conflict rules of the two candidate decisions, calculate the expected misjudgment cost and selection cost of each decision, add the two together to get the total cost, and select the decision with the smaller total cost as the final decision result; if the total costs are equal, decide according to the default security policy.

[0024] The tool capability diagram includes tool identifier, version, input / output mode, unit system, capability range and prerequisites, and interface mapping is performed based on the tool capability diagram. The DSL abstract syntax tree contains STEP nodes and GATE nodes, where STEP nodes contain tool identifiers, input parameters, and validation fields, and GATE nodes contain predicates, thresholds, costs, and branch pointers. The type verification includes: input parameter matching verification with input-output pattern verification, unit consistency verification, and evidence citation parsing verification. The reachability verification includes: branch pointer existence verification, dead-path-free path verification, and fallback path reachability verification.

[0025] The DSL workflow includes a stop or rollback field. The stop condition is generated by the budget and risk threshold, and the rollback field is used for rollback actions when the GATE is triggered or the verification fails. The receipt includes: execution status, output fields, evidence list, resource consumption, anomaly diagnosis and receipt hash. The receipt is written into the execution node in the evidence graph and points to the corresponding DSL step and input evidence object through derived edges.

[0026] A system for dynamic workflow generation and execution based on boundary constraints and expert experience filtering includes: The boundary compilation and filtering module is used to compile boundary constraints into executable predicates and perform joint filtering on candidates; The gray zone determination module performs gray zone determination on the residual set R based on margin, uncertainty, degree of evidence conflict and / or extrapolation, and sets the candidate labels of those that are feasible at the boundary but meet the gray zone determination conditions as gray zone candidates; The expert experience filtering module filters and sorts the residual set R based on field-based experience rules and outputs decisions and evidence citations. The tool capability graph module provides information on tool version, input / output modes, and capability range. Dynamic orchestrator and DSL compiler are used to build DSL abstract syntax trees and perform type and reachability checks to generate DSL workflows; The execution engine is used to execute step-by-step according to the DSL and output receipts, evidence, and trajectory fields.

[0027] The expert experience filtering module includes a rule matching unit, a conflict resolution unit, and a cost-sensitive adjudication unit, and outputs suggested operations to supplement the characterization or adjust variables. The method for generating the suggested operation is as follows: S311. Determine the action type based on the cause of failure, wherein the action type includes representation, variable adjustment, and rollback; S312. Instantiate the action template, wherein the action template includes a template identifier, trigger condition, action type, and parameter mapping; S313, Variable Adjustment.

[0028] The specific process of interface mapping is as follows: S41. Tool Capability Retrieval; S42, Parameter Matching; S43, Unit Conversion; S44, API encapsulation; S45, Generation: Tool calls include endpoints, parameters, and versions.

[0029] The abstract syntax tree nodes generated by the DSL compiler include stop and fallback nodes, and the fallback path is verified to exist and be reachable in the reachability check; The system also includes a runtime monitor, which checks the verification results against the gating threshold and records the triggering reason and evidence reference when a stop or rollback is triggered. The system also includes a replay and differential audit module, which is used to load the historical parser based on the tracking identifier and DSL hash, replay the execution, and output the differential for auditing. The system also includes a consistency verification module, which verifies that the pattern version of the empirical rule record is consistent with the DSL field set.

[0030] The runtime monitor includes: verification result acquisition, threshold matching, stop / rollback triggering, and audit logs.

[0031] The steps for replaying execution and outputting the difference include: DSL loading, restoring the execution environment, replaying execution, and difference calculation.

[0032] The verification steps include: obtaining the rule field set, obtaining the DSL field set, and consistency verification.

[0033] The steps for interface mapping based on the tool capability graph include: tool capability retrieval, parameter matching, unit conversion, and API encapsulation.

[0034] An electronic device includes a processor and a memory; the memory stores computer program instructions, and the processor executes the computer program instructions to implement a method for dynamic workflow generation and execution; the electronic device includes interface circuitry and adapter components for invoking at least two heterogeneous tools, and performs interface mapping based on a tool capability graph; the electronic device is configured to generate a DSL hash for a DSL workflow and establish an index association between the DSL hash and a receipt hash to support playback.

[0035] One of the computer-readable storage media stores a computer program thereon, wherein the computer program, when executed by a processor, implements a method for dynamic workflow generation and execution; The DSL data object is also stored on the medium. The DSL data object includes: step identification, tool identification, input, verification, gating, and stop-backoff field sets, and locks the required fields through pattern versioning; the DSL data object includes tool configuration fields and unit system fields for indicating the tool version and parameter unit system; The empirical rule record data object is also stored on the medium. The data object includes a misjudgment cost field for cost-sensitive filtering and adjudication; The marking data object of the residual set R is also stored on the medium. The marking data object includes non-removable candidates, gray-scale reasons, boundaries, or uncertainty fields for recording the basis for gray zone determination and supporting auditing.

[0036] In a preferred embodiment, the dynamic orchestrator takes the state s as input and outputs the action a. The state s at least includes the target, variables, evidence references, budget, and risk preference. The action a at least includes tool selection, sequence, threshold setting, gating branches, and backoff actions; the action sequence is compiled into DSL and the executability is locked through the type system.

[0037] In one embodiment, the system expresses the boundary as a set of computable predicates P = {P1(x),..., Pn(x)}, and calculates the constraint margin for each predicate. The margin is used to measure the distance of the candidate x to the boundary violation. The combined screening uses short-circuit evaluation: if there is Pi(x) = false, it is immediately excluded.

[0038] In one embodiment, the gray zone determination is performed on the premise that P_all(x) = true. The system writes the candidate into the residual set R and sets flag_unremovable = true when any of the following conditions are met: 0 < margin(x) < m0; prediction uncertainty u(x) > u0; evidence conflict degree K(x) > K0; or extrapolation degree ood(x) > ood0. The system also records the gray-scale reasons and reason codes to support auditing.

[0039] In one embodiment, the expert experience filtering module receives R and its context state, and matches based on the scope and triggering conditions of the empirical rule record; performs conflict resolution on the set of matching rules: weights according to the quality score and freshness of the evidence reference in the evidence graph, and combines the misjudgment cost to select conservative or aggressive decisions, and outputs decisions, rankings, reason codes, evidence references, and recommended operations.

[0040] In one embodiment, the dynamic orchestrator and the DSL compiler generate a DSL abstract syntax tree, where AST nodes contain types such as STEP / GATE / STOP / FALLBACK / ANNOTATE. The compiler performs type verification (consistency of input / output structure and units) and reachability verification (no dead ends, no orphaned nodes, and reachable fallback paths) based on the tool capability graph, and generates a DSL hash and version pointer for replay after compilation.

[0041] In one embodiment, the execution engine executes tool calls step-by-step according to the DSL and generates receipts. Each receipt includes output, a list of evidence, and resource consumption, and generates a receipt hash. The runtime monitor checks the verification results against gating thresholds; when a stop or fallback is triggered, the trigger reason is recorded and written to the trajectory field. The system supports loading a history parser by tracking identifier and DSL hash to perform playback and output differentials for review or regulatory auditing.

[0042] The specific table based on the above embodiments is as follows: Table 1. Overview of Input-Processing-Output for Dynamic Workflow Generation and Execution Note: The process of “boundary compilation → gray zone determination → experience filtering → DSL compilation → execution verification” is linked into a strongly constrained chain to highlight the feasibility of the project.

[0043] Table 2 Boundary Constraint Compilation and Joint Screening Note: Write the boundary as an executable predicate and output the margin for subsequent hard determination of "non-removable gray areas".

[0044] Table 3. Criteria for Gray Areas “Cannot Be Removed by Physical / Chemical / Mathematical Boundaries” Note: Assuming the boundary condition is deemed "feasible", the "gray area cannot be removed" condition is written as a computable hard condition using indicators such as margin, uncertainty, conflict, and extrapolation.

[0045] Table 4. Input / Output and Audit Fields for BEF195 Expert Experience Filtering Note: Emphasize "Filtering = Actionable Decisions" and output auditable fields (trigger rules, evidence citations, cost trajectories).

[0046] Table 5 Tool Capability Graph (TCG) and Reachability Verification Note: Write "tool availability / accessibility" as a graph structure and compile-time checks to avoid it being perceived as just if-else orchestration.

[0047] Table 6 DSL AST Node Structure and Type System Note: The "executability" of the DSL is locked through the type system and compile-time verification, and evidence references are included in the semantic checks.

[0048] Table 7 DSL Compilation and Conformance Verification Pipeline Note: By breaking down the compilation process into an AST (Abstract Syntax Tree) and a verification pipeline, the technical depth and provability of the "processing method" are enhanced.

[0049] Table 8. Operational Monitoring, Playback, and Differential Auditing Note: "Replayable / Auditable" is written as runtime mechanism and data object to close the patent technology effect.

Claims

1. A dynamic workflow generation and execution method based on boundary constraints and expert experience filtering, characterized in that: It includes the following steps: S1: Receive a candidate set and a boundary constraint set, compile the boundary constraints in the boundary constraint set into executable predicates, and perform joint screening on the candidate set to obtain a feasible set F and a residual set R; S2: Mark the candidates in the residual set R that are boundary-feasible but meet the gray area determination conditions as gray area candidates, where the gray area determination conditions include at least one of margin, prediction uncertainty, evidence conflict degree, or extrapolation degree exceeding their corresponding thresholds; S3: Call the expert experience filtering module for the gray area candidates, perform field-based experience rule filtering, and perform cost-sensitive adjudication in case of rule conflicts, and output results of retention, weight reduction, or pending review; S4: Build a DSL abstract syntax tree based on the feasible set F, the output results of retention, weight reduction, or pending review, the tool capability map, the resource budget, and the stop / rollback conditions, and perform type checking and reachability checking; S5: Compile the processes of S1 to S4 into an executable DSL workflow, and output a candidate experiment set including risk tags, evidence pointers, and playback indexes. The candidate experiment set drives the heterogeneous toolchain to execute the instruction sequence of the toolchain corresponding to the executable DSL workflow, and finally outputs a receipt, evidence reference, and trace field, and builds an evidence graph based on the receipt, evidence reference, and trace field; S6: Based on the trace field and playback index output in S5, load historical baseline data from the storage. The historical baseline data includes receipts, evidence references, trace fields, and evidence graphs, replay and execute the corresponding DSL workflow, and perform differential comparison between the replay result and the historical baseline to generate a differential report including data, behavior, performance, and status differences for audit analysis and compliance verification. If the difference exceeds the preset threshold, an alarm is triggered.

2. The dynamic workflow generation and execution method based on boundary constraints and expert experience filtering according to claim 1, characterized in that: In the step S1, the boundary constraint set includes physical boundaries, chemical boundaries, and mathematical boundaries; The physical boundaries include at least one or more of thermodynamic stability, phase equilibrium reachability, diffusion, or heat transfer upper limit; The chemical boundaries include at least one or more of side reaction windows, impurity sensitivity, and interfacial compatibility; the mathematical boundaries include at least one or more of inequality constraints, resource budget constraints, or multi-objective feasible region constraints; The compilation of the executable predicate set includes normalizing the boundary expression, unit normalization, and parameter binding, and outputting a margin vector for each candidate object.

3. The dynamic workflow generation and execution method based on boundary constraints and expert experience filtering according to claim 1, characterized in that: In the step S2, the gray area determination conditions at least meet one of the following: 0 < margin < m0, prediction uncertainty > u0, evidence conflict degree of the candidate object in the evidence graph > K0, or extrapolation degree > ood0; Where, m0 is the upper limit of the allowable margin, u0 is the upper limit of the acceptable uncertainty, K0 is the conflict tolerance threshold, ood0 is the demarcation value for extrapolation discrimination. The evidence conflict degree is calculated based on the weighted count or weighted difference of the support edges and contradiction edges in the evidence graph. The weights include the evidence quality score and the time decay term. The evidence quality score is preset by an expert or automatically calculated, and the time decay term is calculated through a predefined decay function based on the evidence generation or last update time.

4. The dynamic workflow generation and execution method based on boundary constraints and expert experience filtering according to claim 1, characterized in that: In step S3, the expert experience filtering module rules adopt a conflict resolution strategy for conflicting experience rules. The conflict resolution strategy includes: weighting by confidence level, weighting by evidence freshness or quality, or cost-sensitive adjudication based on either false positive cost or false negative cost. The cost-sensitive adjudication selects the threshold or decision that minimizes the expected misjudgment cost, wherein the expected misjudgment cost includes a combination of false positive cost × false positive cost and false negative cost × false negative cost. The empirical rule record is a field-based structure, including: rule identifier, domain scope, triggering condition, adjudication standard, action template, evidence citation, confidence level, false positive cost, false negative cost, structure version, and rule version.

5. The dynamic workflow generation and execution method based on boundary constraints and expert experience filtering according to claim 4, characterized in that: The specific processing steps of the expert experience filtering module for field-based experience rule filtering and conflict resolution of gray area candidates are as follows: S31. Perform field matching on the gray zone candidates based on the fields in the fieldized empirical rule record; S32. Calculate the comprehensive score for each fielded empirical rule. The calculation formula is as follows: Comprehensive score = w1 × rule confidence + w2 × evidence quality score + w3 × time decay value; Among them, rule confidence is the pre-defined level of credibility of the rules, and evidence quality score is the reliability score of the evidence. The time decay value is dynamically calculated based on the age of the evidence. w1, w2, and w3 are preset weighting coefficients used to adjust the relative importance of the three indicators. S33. Detect whether there are conflicts among the fielded empirical rules. If there is a rule mapped to "pass" and another rule mapped to "reject" in the same field or condition, it is determined to be a conflict. Extract all conflicting rules, generate two candidate decisions, "pass" and "reject", and calculate the sum of the comprehensive scores of the rules supporting each decision; S34. Based on the sum of the comprehensive scores of the conflict rules of the two candidate decisions, calculate the expected misjudgment cost and selection cost of each decision, add the two together to get the total cost, and select the decision with the smaller total cost as the final decision result; if the total costs are equal, decide according to the default security policy.

6. The dynamic workflow generation and execution method based on boundary constraints and expert experience filtering according to claim 1, characterized in that: In step S4, the tool capability diagram includes tool identifier, version, input / output mode, unit system, capability range and prerequisites, and interface mapping is performed based on the tool capability diagram. The DSL abstract syntax tree contains STEP nodes and GATE nodes, where STEP nodes contain tool identifiers, input parameters, and validation fields, and GATE nodes contain predicates, thresholds, costs, and branch pointers. The type verification includes: input parameter matching verification with input-output pattern verification, unit consistency verification, and evidence citation parsing verification. The reachability verification includes: branch pointer existence verification, dead-path-free path verification, and fallback path reachability verification.

7. The dynamic workflow generation and execution method based on boundary constraints and expert experience filtering according to claim 1, characterized in that: In step S5, the DSL workflow includes a stop or rollback field. The stop condition is generated by the budget and risk threshold, and the rollback field is used for rollback actions when the GATE is triggered or the verification fails. The receipt includes: execution status, output fields, evidence list, resource consumption, anomaly diagnosis and receipt hash value. The receipt is written into the execution node in the evidence graph and points to the corresponding DSL step and input evidence object through derived edges.

8. A system for implementing the dynamic workflow generation and execution method based on boundary constraints and expert experience filtering as described in any one of claims 1 to 7, characterized in that, include: The boundary compilation and filtering module is used to compile boundary constraints into executable predicates and perform joint filtering on candidates; The gray zone determination module performs gray zone determination on the residual set R based on margin, uncertainty, degree of evidence conflict and / or extrapolation, and sets the candidate labels of those that are feasible at the boundary but meet the gray zone determination conditions as gray zone candidates; The expert experience filtering module filters and sorts the residual set R based on field-based experience rules and outputs decisions and evidence citations. The tool capability graph module provides information on tool version, input / output modes, and capability range. Dynamic orchestrator and DSL compiler are used to build DSL abstract syntax trees and perform type and reachability checks to generate DSL workflows; The execution engine is used to execute step-by-step according to the DSL and output receipts, evidence, and trajectory fields.

9. The system of a dynamic workflow generation and execution method based on boundary constraints and expert experience filtering according to claim 8, characterized in that: The expert experience filtering module includes a rule matching unit, a conflict resolution unit, and a cost-sensitive adjudication unit, and outputs suggested operations to supplement the characterization or adjust variables; The abstract syntax tree nodes generated by the DSL compiler include stop and fallback nodes, and the fallback path is verified to exist and be reachable in the reachability check.

10. The system for a dynamic workflow generation and execution method based on boundary constraints and expert experience filtering according to claim 8, characterized in that: The system also includes a runtime monitor, which checks the verification results against the gating threshold and records the triggering reason and evidence reference when a stop or rollback is triggered. The system also includes a replay and differential audit module, which loads a historical parser based on the tracking identifier and DSL hash value to replay the execution and output the differential for auditing. The system also includes a consistency verification module, which verifies that the pattern version of the empirical rule record is consistent with the DSL field set.

Images