A method for detecting original record trustworthy AI compliance auditing

Abstract

The present application relates to the field of artificial intelligence and detection data processing technology, and discloses a kind of detection original record credible AI compliance auditing method, including establishing the association of auditing field in structured detection data and the corresponding content of detection original record;Parse auditing clause, form semantic state clause data and rule state clause data;Determine applicable clause set according to detection task information, the applicable range and version state of auditing clause;Based on applicable clause set, rule auditing and semantic auditing are carried out respectively, and rule auditing result and semantic auditing result are obtained;The auditing clause and original record content on which the semantic auditing result is based are checked;According to rule auditing result and the combined determination of the checked semantic auditing result, form auditing evidence record.The present application can make the auditing result correspond to the auditing field, original record content and auditing clause, improve the determinability of auditing result and the traceability of auditing process.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and detection data processing technology, specifically to a method for credible AI compliance auditing of original records. Background Technology

[0002] Original test records are important data carriers reflecting the actual execution process of test tasks. They typically include test task information, sample information, instrument output data, environmental conditions, and test results. Existing automatic review technologies generally extract field data from the original test records first, and then judge the field values, filling status, and content relationships based on the review clauses in the test standards, test method documents, or quality control documents.

[0003] Because of differences in the format and data source of the original records being inspected, the same audit field may be distributed across different pages or table locations, leading to unclear correspondence between structured fields, original record content, and audit clauses. At the same time, some audit content requires judgment based on multiple fields and contextual semantics, which is difficult to cover using fixed rules alone. When rule judgment and semantic judgment are executed separately, inconsistent results may occur due to differences in the version of the clauses or the audit fields targeted, making it difficult to directly determine the actual audit problem.

[0004] In response to the above, existing technologies still need to address the problems of difficulty in accurately linking audit criteria, difficulty in uniformly judging different audit results, and difficulty in tracing the audit process. Summary of the Invention

[0005] The purpose of this invention is to provide a reliable AI compliance audit method for detecting original records, in order to solve the problems mentioned in the background art, such as the difficulty in accurately associating existing audit criteria, the difficulty in uniformly judging different audit results, and the difficulty in tracing the audit process.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for detecting credible AI compliance audits of original records, comprising: S1. Obtain structured detection data formed from the original detection records, and establish the association between the audit fields in the structured detection data and the corresponding content in the original detection records; S2. Parse the review clauses used for task review, and form semantic clause data for semantic review from the parsing results. Then, form rule-state clause data for rule review from the parsing results of clauses with deterministic judgment conditions. Establish the association between the two types of clause data and the corresponding review clauses. S3. Determine the set of applicable clauses based on the testing task information, the scope of application of the audit clauses, and the version status; S4. Call the rule-state clause data associated with the applicable clause set to perform rule review on the structured detection data, and constrain the language model based on the semantic-state clause data associated with the applicable clause set to perform semantic review on the structured detection data, and obtain the rule review result and semantic review result respectively. S5. Based on the set of applicable terms and the original records to be detected, verify the content of the audit terms and original records on which the semantic audit results are based. S6. Based on the rule review results and the verified semantic review results, make a combined judgment to determine the review issues and their handling paths, and form a review evidence record for the review issues.

[0007] Furthermore, the semantic clause data retains the semantic constraints in the clause parsing results, and the rule-based clause data includes rules formed by the transformation of deterministic judgment conditions; The two types of clause data formed by the same review clause are semantic clause data and rule-based clause data, which correspond to each other through clause association identifiers that represent the review clause and its version.

[0008] Furthermore, the association between the audit field and the corresponding content in the original detection record is established through field location identifiers; The field location identifier is associated with the structured field value of the audit field and its position in the original record being inspected. Both the rule audit result and the semantic audit result record the corresponding field location identifier.

[0009] Furthermore, the set of applicable terms is determined by matching the audit context of the detection task with the applicable conditions and version validity range of the audit terms; When audit clauses with inconsistent constraints on the same audit field are matched, the applicable clauses are filtered according to the preset clause priority relationship, and a conflict review mark is generated for audit clauses that cannot be filtered according to the preset clause priority relationship.

[0010] Furthermore, the rule review includes establishing a mapping between review variables and field positioning identifiers in the rule-state clause data, obtaining the field value of the corresponding review field according to the mapping, and performing deterministic judgment according to the rule-state clause data; The rule review results record the relationship between the review clauses, field location identifiers, and rule judgment process.

[0011] Furthermore, the semantic audit uses the semantic clause data associated with the applicable clause set and the original record content associated with the audit field as the input boundary of the language model, so that the language model generates a semantic audit result within the input boundary, and the semantic audit result contains reference information pointing to the audit clause and original record content on which it is based.

[0012] Furthermore, the verification includes: Based on the cited information, verify whether the audit clauses on which the semantic audit results are based belong to the set of applicable clauses, whether the audit clauses are in a valid version state, and whether there is a supporting relationship between the cited original record content and the corresponding semantic audit results; When all three checks pass, the corresponding semantic audit results will be used for combined judgment. When at least one validation fails, a review identifier is generated for the corresponding semantic audit result.

[0013] Furthermore, the combination determination includes: The rule judgment status is determined based on the triggering of deterministic judgment conditions in the rule review results, and the semantic judgment status is determined based on the identification of non-compliant content in the verified semantic review results. Determine whether the review clauses on which the rule review result and the validated semantic review result are based are consistent, and whether the review fields they target are consistent, to obtain the clause consistency judgment result and the field consistency judgment result; The rule judgment status, semantic judgment status, clause consistency judgment result, and field consistency judgment result are combined into a judgment status combination. The audit issue status corresponding to the judgment status combination is determined in the preset conflict status matrix, and the handling path is determined according to the audit issue status.

[0014] Furthermore, the audit evidence record includes configuring evidence identifiers for audit issues, and associating the audit basis, combined judgment process, and handling results when the audit issue was formed with the evidence identifiers; After the review issue is resolved, the associated content is stored in a way that cannot be overwritten, and subsequent review information is appended to the same evidence identifier to restore the formation process of the review issue based on the evidence identifier.

[0015] Furthermore, when the processing path indicates manual review, manual review is performed on the corresponding review issues to obtain the manual review results; Extract review correction information that can be converted into deterministic judgment conditions from the manual review results, and generate candidate rules based on the review correction information; The candidate rules are verified based on the audit evidence records, and the verified candidate rules are used to update the rule state clause data. When at least one of the semantic clause data, rule-based clause data, and language model configuration is updated, a regression audit is performed on the updated audit version based on the audit evidence record, and the release of the audit version is controlled according to the regression audit results.

[0016] Compared with existing technologies, the beneficial effects of this invention are as follows: This method for credible AI compliance auditing of original detection records establishes a correlation between the original detection records, audit fields, and audit clauses, performs rule-based auditing and semantic auditing respectively, verifies the basis for semantic auditing, and combines the two types of audit results for judgment, so that audit issues can be mapped to specific record content, audit fields, and audit clauses; at the same time, by recording audit evidence and controlling the release of audit versions, the determineability and traceability of compliance auditing of original detection records are improved, as detailed below: 1. By using field location identifiers to associate audit fields, structured field values, and original record content, and by using clause association identifiers to associate audit clauses and their versions, it is easier to accurately locate the audit results to the corresponding audit fields and original record locations, reducing the deviation in audit basis caused by inconsistencies in field location, clause scope of application, or clause version.

[0017] 2. By performing applicability verification, version validity verification, and support relationship verification on the semantic audit results, and determining whether the rule audit results and the verified semantic audit results are consistent in terms of audit clauses and audit fields, a combination of judgment states is formed. The conflict state matrix is ​​then used to determine the audit issue status, which can distinguish between actual non-compliance, inconsistent audit basis, and inconsistent audit objects, so that the dual-path audit results have clear judgment basis.

[0018] 3. By storing and adding audit evidence records that cannot be overwritten, and using the results of manual review for candidate rule verification, regression audit, and audit version release control, the system can retain the correspondence between audit basis, judgment process, and handling results, and verify its impact on existing audit results before the audit version is updated. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the process for the AI ​​compliance audit method for detecting original records according to the present invention. Figure 2 This is a schematic diagram illustrating the data association relationships between field location identifiers, clause association identifiers, and the set of applicable clauses in this invention. Figure 3 This is a schematic diagram of the dual-path review data flow for rule review and semantic review in this invention; Figure 4 This is a schematic diagram of the judgment logic for verifying the semantic audit results of this invention; Figure 5 This is a schematic diagram illustrating the combined determination of the rule review results and semantic review results of this invention; Figure 6 This is a schematic diagram illustrating the control relationship between the review evidence record and the release of the review version in this invention. Detailed Implementation

[0020] 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.

[0021] Example 1: Please refer to Figures 1-6 This invention provides a technical solution: a method for detecting the credibility of original records in AI compliance audits, comprising the following steps: S1. Obtain structured detection data formed from the original detection records, and establish the association between the audit fields in the structured detection data and the corresponding content in the original detection records; In this embodiment, the original detection record is a data carrier formed during the execution of the detection task and used to reflect the actual detection process. It includes basic information of the detection task, sample information, detection items, instrument output data, environmental conditions, operator records, detection results, and signature and time information. The original detection record can come from electronic forms in the detection business system, record files exported by the detection instrument, and electronic files of paper records that have been recognized by layout. When obtaining the original detection record, the detection task identifier, original record identifier, record page number, and layout position information of the record content are obtained at the same time so that different detection tasks and different versions of the original record can be distinguished from each other.

[0022] During structured processing, field names, values, units of measurement, and data types are extracted from the original records based on their titles, table row and column relationships, field labels, and corresponding content, forming structured inspection data. For electronic forms, field values ​​and corresponding positions are read according to the form's field paths. For record files containing text and tables, field values ​​are extracted based on text paragraph positions and table cell positions. For image-based record pages, the coordinate range of the corresponding text area in the original page is preserved while recognizing the text content. Structured processing retains the original content before extraction, ensuring that field values, even after standardizing date, numerical, and unit of measurement formats, still point to their source content. Audit fields are those in the structured inspection data that require compliance judgment based on audit clauses. For example, when inspection task identifier, sample identifier, inspection item, inspection equipment identifier, original measurement value, inspection result, inspection time, and personnel signature are used as audit fields, each audit field has a corresponding field code. This field code is used to determine the processing object in subsequent rule and semantic audits. If the same audit field appears multiple times in the original record, retain the structured field value and original record content corresponding to each occurrence to prevent duplicate fields from losing their original positional relationship during the structuring process.

[0023] In this implementation case, the association between the audit field and the corresponding content in the original detection record is established through field location identifiers; the field location identifiers are associated with the structured field values ​​of the audit field and their positions in the original detection record, and both the rule audit results and the semantic audit results record the corresponding field location identifiers; Specifically, a field location identifier is generated for each audit field. This identifier consists of the original record identifier, record page number, record content object number, and location parameters. When the field label of the original record indicates that an audit field should be recorded, but no field value has been generated, the field code and empty field value of that audit field are retained. A field location identifier is then generated based on the field label and the location parameters of its corresponding filling area, pointing to the corresponding empty position in the original record. The original record identifier is used to determine the original record from which the audit field originates; the record content object number distinguishes between text paragraphs, tables, and image areas within the record; and the location parameters represent the character range of the audit field's corresponding content in a text paragraph, its row and column position in a table, or its area coordinates in a record image. The field location identifier is stored together with the audit field code, structured field value, and corresponding original record content, thus forming a queryable association between the audit field and the original record content. For example, if the original record identifier is JC-2026-0012, and the result value of the detection item is located in the 5th row and 4th column of the 1st table on page 2, then a field location identifier containing JC-2026-0012, page 2, 1st table, 5th row, and 4th column is generated for this result value, and this identifier is associated with the structured field value of the corresponding detection item. Subsequently, based on this field location identifier, the structured field value can be returned to the original table cell, and the original record content on which the structured field value was based can be obtained.

[0024] After S1 is executed, it outputs structured detection data containing audit field codes, structured field values, field location identifiers, and the relationship between the original record content. The field location identifiers are passed to the subsequent audit process along with the corresponding audit fields, so that the rule audit results and semantic audit results can record the audit fields they target and their original positions, and provide the original content location basis for semantic audit basis verification, combination judgment, and audit evidence recording.

[0025] S2. Parse the review clauses used for task review, and form semantic clause data for semantic review from the parsing results. Then, form rule-state clause data for rule review from the parsing results of clauses with deterministic judgment conditions. Establish the association between the two types of clause data and the corresponding review clauses. Specifically, the audit clauses used for reviewing testing tasks originate from testing standards, testing method documents, laboratory management documents, and quality control documents. When obtaining audit clauses, the document to which the clause belongs, the clause number, the clause text, and version information are simultaneously acquired, enabling the parsed clause data to be linked to the corresponding audit clause and its specific version. When parsing audit clauses, the audit object, constraint content, and judgment conditions in the clause text are identified. The audit object is used to determine the audit field targeted by the clause; the constraint content indicates the content requirements that the audit field should meet; and the judgment conditions indicate the judgment relationship between the field value and the limit value, limit range, enumerated value, format requirements, or completion status. For audit clauses involving multiple audit fields, the correspondence between each audit field and its corresponding constraint content is recorded separately, while the combination relationship when multiple audit fields jointly satisfy the clause requirements is also retained.

[0026] In this embodiment, the semantic clause data retains the semantic constraints from the clause parsing results. Semantic constraints refer to the constraints in the audit clause whose meaning needs to be determined by combining the clause context and the original record content, including the audit object, applicable situation, required execution matters, and the relationship between the audit objects. When forming semantic clause data, the audit clause text and its context are retained, and the audit fields, constraint content, and judgment targets obtained from parsing are recorded, so that subsequent semantic audits can identify and detect non-compliant content in the original record within the scope of the audit clause's content. For example, if the audit clause states that when the detection result is abnormal, the retest situation and handling conclusion should be recorded, then the detection result, retest situation, and handling conclusion are determined as relevant audit fields, the abnormal detection result is recorded as an applicable situation, and the retest situation and handling conclusion are recorded as semantic constraints. This semantic clause data is used to determine whether the abnormal results, retest records, and handling conclusions in the original record form a corresponding content relationship. The rule-based clause data includes rules transformed from deterministic judgment conditions. Deterministic judgment conditions refer to conditions that allow for the deriving of an audit conclusion through explicit field values ​​and pre-defined judgment relationships. For clause parsing results containing deterministic judgment conditions, the corresponding audit fields, comparison relationships, limit values, and rule output states are extracted and converted into callable rules. For example, when an audit clause limits the ambient temperature to 20℃±2℃, the ambient temperature is determined as the audit variable, and 18℃ to 22℃ is determined as the allowed value range, forming a deterministic judgment condition for determining whether the ambient temperature field value is within the allowed value range. When the field value is within the allowed value range, a condition-not-triggered state is formed; when the field value exceeds the allowed value range, a condition-triggered state is formed. For mandatory requirements, date format requirements, and preset enumeration value requirements, these are respectively converted into deterministic judgment conditions for whether the field value is empty, whether the field value conforms to the specified format, and whether the field value belongs to the preset enumeration range. When the same review clause contains both deterministic judgment conditions and semantic constraints that require contextual understanding, rule-based clause data and semantic-based clause data are generated based on the review clause. The rule-based clause data carries the deterministic judgment conditions that can be directly executed, while the semantic-based clause data retains the semantic constraints and contextual content in the review clause, so that they can be called by subsequent rule reviews and semantic reviews respectively.

[0027] Among them, the two types of clause data formed by the same review clause are semantic clause data and rule-based clause data, which correspond to each other through clause association identifiers that represent the review clause and its version. Specifically, a clause association identifier is generated based on the file identifier, clause number and version information of the document to which the review clause belongs, and the clause association identifier is written into the semantic clause data and rule-based clause data formed by the review clause respectively. For review clauses that only form semantic clause data, the clause association identifier is written into the corresponding semantic clause data. For review clauses that form both types of clause data, the same clause association identifier is recorded in both types of clause data.

[0028] After S2 is executed, it outputs semantic clause data, rule-based clause data for clauses containing deterministic judgment conditions, and the relationship between each clause data and the corresponding audited clause. Subsequently, based on the clause association identifier, it can be determined whether the semantic clause data and the rule-based clause data come from the same audited clause and the same version, and the audit results can be returned to the corresponding audited clause.

[0029] S3. Determine the set of applicable clauses based on the testing task information, the scope of application of the audit clauses, and the version status. In this embodiment, the testing task information is obtained from the structured testing data formed in S1 and the task configuration data in the testing business system, including the testing object category, testing item, testing method, sample type, task execution agency, task execution time, and testing task identifier. The obtained testing task information is used to determine the audit business scope of the current testing task.

[0030] In this embodiment, the set of applicable clauses is determined by matching the audit context of the testing task with the applicable conditions and version validity range of the audit clauses. Specifically, an audit context is formed based on the testing task information, matched with the applicable conditions of the audit clauses, and audit clauses with valid version status are selected from the matched audit clauses to determine the set of applicable clauses. The audit context refers to a structured set of conditions composed of data items in the testing task information that can define the applicable boundaries of the audit clauses. When forming the audit context, the testing object category, testing item, testing method, sample type, task execution agency, and task execution time are written into the corresponding context fields. Context fields with missing values ​​are recorded as pending supplementation, and audit clauses that rely on these context fields for matching are recorded as pending supplementation matching clauses. These pending supplementation matching clauses are not written into the set of applicable clauses until supplementary values ​​are obtained. When a context field with missing values ​​originates from the original testing record, supplementary values ​​are obtained from the restructured original testing record. When a context field with missing values ​​originates from the testing task configuration, supplementary values ​​are obtained from the updated task configuration data of the testing business system. After obtaining the supplementary value, the applicable condition matching of the audit clause is re-executed, and the determination of whether to include it in the applicable clause set is based on the re-obtained applicable condition matching result. The applicable conditions of the audit clause are formed by transforming the scope of application of the audit clause, including the conditions of the test object, the test item, the test method, the sample type, and the execution agency. During matching, the corresponding fields in the audit context are read item by item according to the applicable conditions actually contained in the audit clause. When the audit context meets all the applicable conditions contained in an audit clause, the audit clause is determined as a matched audit clause. When any applicable condition is inconsistent with the audit context, the audit clause is excluded from the matching result. For audit clauses whose scope of application includes superior and subordinate categories, a hierarchical correspondence between the test object category, the test item, and the sample type is pre-established. When the specific category in the audit context belongs to the superior category defined by the audit clause, the corresponding applicable condition matching is determined. When the audit clause is limited to a specific category, the corresponding applicable condition matching is determined only when the category in the audit context is consistent with the specific category. This allows general clauses applicable to a class of testing tasks and special clauses applicable to specific testing items to participate in the screening according to their respective scopes of application. The version status is determined based on the effective time, expiration time, and controlled status of the version corresponding to the audit clause. The execution time of the detection task is used as the time benchmark for version matching. When the execution time of the detection task is within the range of the effective time to the expiration time of the version corresponding to the audit clause and the version is in a controlled and effective state, the version status of the audit clause is determined to be valid. When the audit clause does not set an expiration time, the time range after its effective time and in a controlled and effective state is taken as the valid range.

[0031] When multiple versions of the same review clause are released consecutively, the version status of each version is matched with the execution time of the detection task, and the review clauses whose version status is valid under that time reference are retained. This ensures that the set of applicable clauses corresponds to the review basis that should be used when the detection task is actually executed, and prevents subsequent reviews from directly using clause versions that have been released at the review time but have not yet taken effect at the time of task execution. When determining the set of applicable clauses, a candidate clause set is first formed based on the matching results of the review context and applicable conditions. Then, invalid versions are filtered out from the candidate clause set based on the version status. The detection task identifier, the clause association identifier of the remaining review clauses, the applicable condition matching results, and the version status are all written into the set of applicable clauses. The detection task identifier is then associated with the written clause association identifier, so that the semantic clause data and rule-based clause data formed by S2 can be called by the subsequent review process within the same applicable boundary based on the clause association identifier.

[0032] In this implementation, when audit clauses with inconsistent constraints on the same audit field are matched, applicable clauses are selected based on a preset clause priority relationship. Inconsistent constraints refer to multiple audit clauses specifying different value ranges, filling requirements, format requirements, or content relationships for the same audit field that cannot be simultaneously satisfied. When identifying inconsistent constraints, the audit field targeted by each audit clause is obtained based on the clause association identifier, and the deterministic judgment conditions in the corresponding rule-based clause data and the semantic constraints in the semantic-based clause data are compared. When the constraint results corresponding to the same field cannot be simultaneously satisfied, the corresponding audit clause is identified as a conflict candidate clause. The preset clause priority relationship is determined when audit clauses are entered into the database based on the document validity level, the specificity of the applicable scope, and the version effective time already determined in the controlled documents. The established clause ranking relationship prioritizes clauses with higher document validity levels during screening. When document validity levels are the same, clauses with more matching conditions to the audit context and more specific application scope are selected. If the aforementioned relationships are still the same, clauses that are valid during the execution time of the testing task and have a later effective time are selected. The clause association identifiers and the priority relationships adopted before and after screening are recorded. For example, for the environmental temperature field of the same testing item, the general audit clause limits the first temperature range, and the audit clause specifically for the current testing method limits the second temperature range. When the two temperature ranges cannot be satisfied simultaneously and the special audit clause fully matches the testing method in the audit context, the special audit clause is selected based on the specificity of its application scope, and its clause association identifier is written into the set of applicable clauses.

[0033] A conflict review identifier is generated for audit clauses that cannot be filtered based on the preset clause priority relationship. That is, when the applicable clause cannot be determined based on the preset clause priority relationship, a conflict review identifier is generated for the corresponding audit clause. Specifically, the conflict review identifier is used to indicate that multiple audit clauses have unresolved constraint conflicts for the same audit field. When generating the conflict review identifier, the detection task identifier, audit field, clause association identifier of the corresponding audit clause, inconsistent constraint content, and priority relationship comparison result are associated, and the corresponding audit clause is recorded as a clause to be reviewed. Before the review is completed and the applicable clause is determined, the clause to be reviewed is not used as the basis for determining automatic audit invocation.

[0034] After S3 is executed, it outputs the set of applicable clauses corresponding to the current detection task and the conflict review identifier formed when the conflict cannot be eliminated by the preset clause priority relationship. The set of applicable clauses provides a common clause application boundary for subsequent rule review and semantic review, and enables each review result to be traced back to the applicable conditions, version status and screening process of the corresponding review clause.

[0035] S4. The rule-state clause data associated with the applicable clause set is invoked to perform rule review on the structured detection data, and the semantic-state clause data associated with the applicable clause set is constrained to perform semantic review on the structured detection data, so as to obtain the rule review result and the semantic review result respectively. In this embodiment, the clause association identifier is read from the applicable clause set output by S3, and the rule-state clause data and semantic-state clause data formed by S2 are obtained through the clause association identifier, and the structured detection data formed by S1 is used as the review object, so that the rule review and the semantic review adopt the same review clause application boundary.

[0036] In this embodiment, rule review includes establishing a mapping between review variables and field location identifiers in the rule-state clause data, obtaining the field value of the corresponding review field according to the mapping, and performing deterministic judgment according to the rule-state clause data. In this embodiment, the review field code to be called is determined according to the review variables recorded in the rule-state clause data, and the corresponding field value is obtained from the structured detection data according to the mapping between the review field code and the field location identifier. The field value is input into the corresponding deterministic judgment condition, and the clause association identifier, review field code, field location identifier, deterministic judgment condition and condition triggering status are recorded to form the rule review result. Specifically, audit variables are data items in the rule-based clause data used to refer to the audit fields to be judged. When establishing the mapping, based on the audit field codes recorded in the rule-based clause data, the same audit field codes are searched in the structured inspection data, and the field location identifier corresponding to the found audit field is written into the value source of the audit variable to determine the structured field value actually called by each deterministic judgment condition and its original record location. When an audit variable corresponds to multiple field location identifiers in the same inspection original record, the field location identifiers participating in the judgment are determined based on the audit objects and field combination relationships recorded in the rule-based clause data. When the deterministic judgment condition is true for each inspection record, The system sequentially retrieves the structured field values ​​associated with each field location identifier for judgment. When a deterministic judgment condition is met for the relationship between multiple audit fields, the corresponding field values ​​are retrieved and judged simultaneously according to the field combination relationship recorded in the rule-state clause data. When retrieving field values ​​according to the mapping, the structured field values ​​associated with the field location identifier in S1 are read, and the data type and unit of measurement of the field value are retained. When the rule-state clause data includes one of the following: value range, enumerated value, filling status, and format requirements, the field value is judged according to the data type required by the deterministic judgment condition, so that the audit result can still be returned to the original record content before conversion through the field location identifier. For example, if a rule-state clause in the applicable clause set limits the ambient temperature to 18℃ to 22℃, the ambient temperature field value is retrieved according to the mapping between the audit variable and the field location identifier. When the retrieved field value is 23℃, it is determined that the deterministic judgment condition has been triggered; when the retrieved field value is within the range of 18℃ to 22℃, it is determined that the deterministic judgment condition has not been triggered.

[0037] For mandatory audit fields, the mandatory condition is triggered when the field value is empty and the corresponding original record position does not contain any content. For audit fields with format requirements, the format condition is triggered when the field value cannot be parsed according to the data format specified by the rule-state clause data. For audit fields with enumeration range requirements, the enumeration condition is triggered when the field value does not belong to the allowed value set of the rule-state clause data record.

[0038] The rule review results record the association between the review clauses, field location identifiers, and the rule judgment process; that is, the rule review results record the association between the invoked review clauses, field location identifiers, obtained field values, and the triggering status of deterministic judgment conditions; specifically, for each deterministic judgment condition judgment completed, a corresponding rule review result is generated, and this result is recorded together with the clause association identifier of the review clause, the field location identifiers involved in the judgment, the actual obtained structured field values, and the condition triggering status, so that subsequent combined judgments can determine the review clauses and review fields targeted by the rule judgment, and the review evidence record can reconstruct the execution basis of this rule judgment.

[0039] In this embodiment, semantic auditing includes using the semantic clause data associated with the set of applicable clauses and the original record content associated with the audit field as the input boundary of the language model, so that the language model generates semantic auditing results within the input boundary. Specifically, the semantic audit context is a data set for the language model to perform single semantic judgments. For each semantic clause data, the relevant audit fields are determined based on the audit object and semantic constraints recorded within it. Then, the structured field values ​​and corresponding original record content are obtained based on the field location identifiers of the relevant audit fields. The semantic clause data, audit field codes, structured field values, and original record content are combined according to their correspondence to form the semantic audit context. When a semantic constraint involves the content relationship between multiple audit fields, each audit field and its original record content are placed in the same semantic audit context, and the field location identifiers corresponding to each audit field are retained. When a semantic constraint only targets a single audit field, only that audit field and its corresponding original record content are combined to reduce the impact of irrelevant detection content on semantic judgments. The impact of semantic review is as follows: When the language model generates semantic review results, the semantic constraints retained in the semantic clause data are set as the basis for judgment, and the relevant review fields and their original record content are set as the content to be judged. The language model is required to determine whether there is content in the structured detection data that does not comply with the corresponding review clauses based only on the semantic review context, and output the review clauses and original record content on which it is based. For example, the semantic clause data requires that the retest situation and processing conclusion be recorded when the detection result is abnormal. When the detection result in the semantic review context is abnormal, and the original record content contains the retest situation but does not record the processing conclusion, the language model identifies the missing processing conclusion based on the semantic constraint, and uses the original record position corresponding to the review clause and the processing conclusion review field as the basis for the semantic review result.

[0040] The semantic audit result includes reference information pointing to the audit clauses and original record content on which it is based. The reference information includes clause association identifiers for pointing to audit clauses and field location identifiers for pointing to original record content. The clause association identifier is used to determine the audit clauses and their versions actually used when the language model forms the semantic audit result, and the field location identifier is used to determine the original record location actually used when the language model forms the semantic audit result. The identified non-compliant content, clause association identifiers, and field location identifiers are established to establish a correspondence, so that S5 can respectively verify the applicability of the audit clauses, the validity of the version, and the support relationship between the original record content and the semantic audit result.

[0041] After S4 is executed, it outputs the rule review result and the semantic review result respectively. The rule review result retains the execution basis and triggering status of the deterministic judgment conditions, while the semantic review result retains the non-compliant content identified based on the semantic review context and its reference information. The two types of review results are then passed to the subsequent verification and combination judgment process.

[0042] S5. Based on the set of applicable clauses and the original records to be detected, verify the audit clauses and original record content on which the semantic audit result is based. In this embodiment, the reference information contained in the semantic audit result output by S4 is read, the audit clause on which the semantic audit result is based is obtained based on the clause association identifier in the reference information, the original record content on which the semantic audit result is based is obtained based on the field positioning identifier in the reference information, and the corresponding semantic audit result, audit clause and original record content are combined into data to be verified, so that each semantic audit result has a clear clause verification object and record content verification object.

[0043] In this embodiment, the verification includes, based on the reference information, verifying whether the audit clause upon which the semantic audit result is based belongs to the applicable clause set, whether the audit clause is in a valid version state, and whether there is a supporting relationship between the referenced original record content and the corresponding semantic audit result. Specifically, when verifying whether the audit clause upon which the semantic audit result is based belongs to the applicable clause set, the clause association identifier in the reference information is compared with the clause association identifier recorded in the applicable clause set formed in S3. If the same clause association identifier exists in the applicable clause set, and the detection task identifier corresponding to that clause association identifier is consistent with the current detection task identifier, the verification is determined to pass. If the same clause association identifier does not exist in the applicable clause set, or the corresponding clause association identifier corresponds to another detection task, the verification is determined to fail. When verifying whether the audit clause is in a valid version state, the version information, effective time, expiration time, and controlled status of the audit clause are obtained based on the clause association identifier, and the detection task is executed accordingly. Time is used as the version verification benchmark. When the execution time of the detection task is within the valid range of the version corresponding to the audit clause, and the version is in a controlled and valid state, the verification is determined to pass. When the execution time of the detection task exceeds the valid range, or the version is in at least one of the invalid, obsolete, and uncontrolled states, the verification is determined to fail, and the actual referenced version and the clause version applicable to the execution time of the detection task are recorded. When verifying whether there is a supporting relationship between the referenced original record content and the corresponding semantic audit result, the corresponding position in the detection original record is returned according to the field positioning identifier, the original record content saved at that position is obtained, and it is checked whether the obtained original record content is consistent with the content used when S4 forms the semantic audit context. On the basis of content consistency, it is further determined whether the original record content can directly indicate the non-compliant content identified by the semantic audit result. The supporting relationship indicates that the referenced original record content can provide the factual content and its record position required to form the corresponding semantic audit result.

[0044] When the semantic audit result indicates that there is content in the original record that does not meet the requirements of the clause, if the referenced original record content includes the identified audit object, the actual recorded content, and the content relationship corresponding to the semantic constraint, then it is determined that there is a supporting relationship between the original record content and the semantic audit result. When the semantic audit result indicates that the required recorded content is missing, if the field positioning identifier points to the missing position of the corresponding audit field, and the reference information is simultaneously associated with the original record content that can trigger the recording requirement, then it is determined that there is a supporting relationship between the original record content and the semantic audit result. For example, if the semantic audit result identifies that the processing conclusion corresponding to the anomaly detection result is missing, and the reference information points to the audit clause that requires the anomaly detection result to record the retest situation and processing conclusion, and points to the field position recording the anomaly detection result, the field position recording the retest situation, and the missing position of the processing conclusion respectively, when the audit clause belongs to the applicable clause set and is in a valid version state, and the aforementioned original record content can indicate that the anomaly detection result has occurred, the retest situation has been recorded, and the processing conclusion has not been recorded, then it is determined that all three checks have passed. When the field location identifier cannot return the corresponding position in the original record, the returned original record content is inconsistent with the content used when forming a semantic audit context with S4, or the returned original record content cannot indicate at least one of the non-compliant content identified by the semantic audit result, the support relationship verification is determined to have failed, and the corresponding verification reasons for failure to locate, inconsistency of content, and lack of support of content are recorded.

[0045] When all three checks pass, the corresponding semantic audit result is used for combined judgment; that is, the semantic audit result is marked as a pass status, and the clause association identifier, field location identifier and the association between the three check results are retained, so that S6 can combine it with the rule audit result for judgment.

[0046] When at least one verification fails, a review identifier is generated for the corresponding semantic audit result. The review identifier is used to indicate that there is a problem with the reference basis of the corresponding semantic audit result that needs to be manually confirmed. When generating the review identifier, the detection task identifier, semantic audit result, failed verification item, verification reason, clause association identifier and field location identifier are associated, and the corresponding semantic audit result is not directly entered into the automatic combination judgment.

[0047] After S5 is executed, it outputs a semantic audit result that passes all three checks, as well as a review flag generated when at least one check fails. This allows the semantic audit result entering the combined judgment to be traced back to the valid audit clauses applicable to the current detection task and the original record content with supporting relationships.

[0048] S6. Based on the rule review results and the verified semantic review results, a combined judgment is made to determine the review issues and their handling paths, and a review evidence record is formed for the review issues. In this embodiment, the rule review results formed in S4 and the semantic review results in S5, which have passed all three verifications, are read. The clause association identifier, review field code, field location identifier, and condition trigger status recorded in the rule review results are combined with the clause association identifier, review field code, field location identifier, and non-compliant content identification results recorded in the semantic review results to form data to be combined. The semantic review results with the review identifier generated in S5 are not included in the automatic combination judgment. Within the same detection task, for each verified semantic review result, its clause association identifier and all the review field codes associated with it are read to form a set of review field codes. The rule review results are read item by item. When the rule review result and the semantic review result meet at least one of the following conditions, they are determined as a candidate combination pair: the clause association identifier of the rule review result is the same as the clause association identifier of the semantic review result; the review field code of the rule review result belongs to the set of review field codes. When neither of the two conditions is met, the rule review result and the semantic review result are not combined into a judgment state combination.

[0049] In this embodiment, the combined judgment includes determining the rule judgment state based on the triggering status of deterministic judgment conditions in the rule review results, and determining the semantic judgment state based on the identification of non-compliant content in the verified semantic review results. Specifically, when a deterministic judgment condition is triggered, the rule judgment state is determined to be rule non-compliant; when a deterministic judgment condition is not triggered, the rule judgment state is determined to be rule compliant. When the verified semantic review results identify non-compliant content, the semantic judgment state is determined to be semantically non-compliant; when the verified semantic review results do not identify non-compliant content, the semantic judgment state is determined to be semantically compliant. The rule judgment state is determined separately for each rule review result. When the same review field corresponds to multiple deterministic judgment conditions, the triggering status of each deterministic judgment condition is retained, and the review field where at least one deterministic judgment condition is triggered is determined to be rule non-compliant. The combination judgment can locate the actual triggered deterministic judgment condition and its field positioning identifier. The semantic judgment state is determined according to the semantic review result after each verification. When the semantic review result involves the content relationship between multiple review fields, the review fields involved in the content relationship judgment are jointly associated with the semantic judgment state. When the semantic review result involves only a single review field, the review field and its field positioning identifier are associated with the semantic judgment state. When the verified semantic review result does not have a rule review result that forms a candidate combination with it, it is recorded as a semantic single-item review result. Among them, if non-compliant content is identified, the review issue state is determined as a semantic single-item pending review state and a manual review handling path is configured. If no non-compliant content is identified, the review issue state is determined as a semantic single-item review passed state. The semantic single-item review result does not enter the preset conflict state matrix.

[0050] The system determines whether the review clauses upon which the rule review result and the validated semantic review result are based are consistent, and whether the review fields they target are consistent, to obtain clause consistency determination results and field consistency determination results. When obtaining the clause consistency determination result, the system compares the clause association identifiers in the rule review result and the semantic review result. If the two clause association identifiers are the same and correspond to the same version of the same review clause, the clause consistency determination result is determined to be consistent; otherwise, the clause consistency determination result is determined to be inconsistent. When obtaining the field consistency determination result, the system reads the review field code and field location identifier of the rule review result, and reads the review field code set of the semantic review result and the field location identifier corresponding to each review field code. If the review field code of the rule review result belongs to the review field code set, and its field location identifier and the field location identifier corresponding to that review field code in the semantic review result point to the same review field in the same detection original record, the field consistency determination result is determined to be consistent; otherwise, the field consistency determination result is determined to be inconsistent.

[0051] The rule judgment state, semantic judgment state, clause consistency judgment result, and field consistency judgment result are combined to form a judgment state combination. The audit issue state corresponding to this combination is determined in a preset conflict state matrix, and a handling path is determined based on the audit issue state. The preset conflict state matrix is ​​a pre-established correspondence between judgment state combinations, audit issue states, and handling paths. Each correspondence in the matrix uses the rule judgment state, semantic judgment state, clause consistency judgment result, and field consistency judgment result as input conditions, and the audit issue state and handling path as output results. When both the rule judgment state and semantic judgment state are non-compliant, and both the clause consistency judgment result and field consistency judgment result are consistent, the audit issue state is determined as a consistent non-compliant state, and the corresponding audit issue confirmation handling path is determined. When both the rule judgment state and semantic judgment state are non-compliant, the audit issue state is determined as a consistent non-compliant state, and a handling path is determined based on the corresponding audit issue confirmation handling path. If both consistency judgments are consistent, the audit issue status is set to "audit passed," and no audit issue is created pending. If one of the rule judgment status and the semantic judgment status is consistent while the other is not, and both the clause consistency judgment result and the field consistency judgment result are consistent, the audit issue status is set to "audit conclusion conflict," and a manual review process is determined. This allows reviewers to verify the differences between the deterministic judgment conditions and the semantic audit results for the same audit clause and the same audit field. If either the clause consistency judgment result or the field consistency judgment result is inconsistent, the audit issue status is set to "audit object pending review," and a manual review process is determined. The associated clause identifier, audit field code, and field location identifier of the inconsistent clause are recorded, enabling reviewers to determine whether the two types of audit results correspond to different audit requirements or different original record content.

[0052] In this embodiment, the audit evidence record includes configuring evidence identifiers for audit issues, and associating the audit basis, combined judgment process, and handling result at the time the audit issue was formed with the evidence identifier. The evidence identifier is a data identifier used to uniquely identify an audit issue and its subsequent handling record. When forming an evidence identifier, a combination relationship is established between the detection task identifier and the audit issue sequence number, so that different audit issues in the same detection task correspond to different evidence identifiers. The evidence record version is associated with the evidence identifier as an additional version field, so that subsequent information of the same audit issue is continuously associated with the same evidence identifier. The audit basis at the time the audit issue was formed includes the clause association identifier and version information of the corresponding audit clause, and the deterministic judgment called by the rule audit result. The conditions and their triggering states, the audit clauses and original record content referenced by the semantic audit results, and the audit field codes and field location identifiers involved in the audit, are combined in the judgment process, which includes rule judgment state, semantic judgment state, clause consistency judgment result, field consistency judgment result, judgment state combination and corresponding audit issue state. When forming audit evidence records, the audit basis, combined judgment process and determined disposal path are first written into the evidence record corresponding to the evidence identifier. After the corresponding disposal is completed, the actual disposal result is associated with the same evidence identifier. Among them, the disposal result corresponding to the manual review disposal path includes the manual review conclusion and review basis, and the disposal result corresponding to the audit issue confirmation disposal path includes the issue confirmation status and completed correction information.

[0053] After the audit issue is resolved, the associated content is stored in an unalterable manner, and subsequent review information is appended to the same evidence identifier to restore the formation process of the audit issue based on the evidence identifier. Unalterable storage means that after the audit issue is resolved, the audit basis, combined judgment process, and handling result already written into the audit evidence record will not be replaced by subsequent content. When supplementary explanations, re-review results, or changes in handling status occur, the new content is written as an append record, and the append time, information source, and corresponding evidence record version are recorded. When restoring the formation process of the audit issue based on the evidence identifier, the audit basis, rule audit result, semantic audit result, combined judgment process, handling path, handling result, and subsequent review information are read sequentially according to the evidence record version and append time, thereby reconstructing the content association of the audit issue from its formation, judgment, handling, and review.

[0054] In this embodiment, when the processing path indicates manual review, manual review is performed on the corresponding review issue to obtain the manual review result. When performing manual review, the reviewer is provided with the applicable review clauses, the verified semantic review result, the corresponding original detection record content, and the corresponding review issue status based on the evidence identifier. When the evidence identifier is associated with a rule review result, the corresponding rule review result is also provided to the reviewer. The manual review result includes the confirmation result of whether the review issue is valid, the review basis used, the judgment content that needs to be corrected, and the corresponding field positioning identifier, and the manual review result is appended to the same evidence identifier.

[0055] The review correction information that can be converted into deterministic judgment conditions is extracted from the manual review results, and candidate rules are generated based on the review correction information. The review correction information that can be converted into deterministic judgment conditions includes explicit review fields, comparison relationships, limit values, value ranges, enumeration values, format requirements, and applicable conditions. When generating candidate rules, the review fields are determined as review variables, the comparison relationships and limit contents are determined as execution conditions, and the clause association identifier, field positioning identifier, and applicable scope corresponding to the manual review results are recorded as the source information of the candidate rules. For review content that cannot form clear judgment boundaries, only manual review information is retained.

[0056] The candidate rules are verified based on the audit evidence records. The verified candidate rules are used to update the rule-state clause data. When verifying the candidate rules, relevant records that can restore the original record content and manual review results are selected from the audit evidence records according to the audit clauses, audit fields and applicable scope associated with the candidate rules. The candidate rules are called to re-execute the judgment on the relevant audit fields, and the execution results are compared with the corresponding manual review results. When the candidate rules can reproduce the corresponding manual review results on all the selected relevant audit evidence records, and do not change the original execution results of other deterministic judgment conditions on the same batch of audit evidence records, the candidate rules are determined to have passed the verification. The verified candidate rules, together with the source evidence identifier and verification results, are written into the unreleased rule-state clause data isolated from the current version, and an unreleased audit version is formed based on the unreleased rule-state clause data.

[0057] When at least one of the semantic clause data, rule-based clause data, and language model configuration is updated, a regression audit is performed on the updated audit version based on the audit evidence record, and the release of the audit version is controlled according to the regression audit result. The audit version is an audit configuration combination jointly identified by the semantic clause data version, rule-based clause data version, and language model configuration version. When at least one of them changes, a new audit version to be released is generated. Before the audit version to be released passes the regression audit, the current audit version continues to be used to perform the original record audit. In the above, the language model configuration is the configuration data used to control the semantic audit execution process, including the model version identifier, audit prompt template version, output field format constraints, and inference parameter configuration. When any of the configuration items changes, a new language model configuration version is recorded. The regression audit refers to using existing audit evidence records to restore the audit clauses, original record content, and confirmed disposal results, calling the updated audit version to re-execute the rule audit, semantic audit, reference verification, and combination judgment, and comparing the re-formed audit issue status with the confirmed results in the original audit evidence record. During the regression audit, the audit version to be released is applied to the selected audit evidence records, and the regression audit result is obtained. For version updates that include verified review corrections, the regression audit results should also reproduce the verified review corrections. For all version updates, the regression audit results should be consistent with the audit results already confirmed as correct in the audit evidence record, and no new audit conclusion conflict status or audit object pending review status should be added. When the corresponding conditions are met, the updated audit version can be released; when the corresponding conditions are not met, the release of the updated audit version should be stopped, and the corresponding evidence identifier, difference results, and version information should be formed into a version review record.

[0058] After S6 is executed, it outputs the status of the audit issue, the handling path, and the corresponding audit evidence record. It also continuously links the handling results, manual review results, candidate rule verification results, and regression audit results through the same evidence identifier, so that the basis for the formation of the audit issue, the combined judgment process, the handling process, and the impact of version updates can be restored and verified.

[0059] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting the credibility of original records in AI compliance audits, characterized in that, include: S1. Obtain structured detection data formed from the original detection records, and establish the association between the audit fields in the structured detection data and the corresponding content in the original detection records; S2. Parse the review clauses used for task review, and form semantic clause data for semantic review from the parsing results. Then, form rule-state clause data for rule review from the parsing results of clauses with deterministic judgment conditions. Establish the association between the two types of clause data and the corresponding review clauses. S3. Determine the set of applicable clauses based on the testing task information, the scope of application of the audit clauses, and the version status; S4. Call the rule-state clause data associated with the applicable clause set to perform rule review on the structured detection data, and constrain the language model based on the semantic-state clause data associated with the applicable clause set to perform semantic review on the structured detection data, and obtain the rule review result and semantic review result respectively. S5. Based on the set of applicable terms and the original records to be detected, verify the content of the audit terms and original records on which the semantic audit results are based. S6. Based on the rule review results and the verified semantic review results, make a combined judgment to determine the review issues and their handling paths, and form a review evidence record for the review issues.

2. The method for detecting the credibility of original records in AI compliance auditing according to claim 1, characterized in that, The semantic clause data retains the semantic constraints in the clause parsing results, and the rule-based clause data includes rules formed by the transformation of deterministic judgment conditions; The two types of clause data formed by the same review clause are semantic clause data and rule-based clause data, which correspond to each other through clause association identifiers that represent the review clause and its version.

3. The method for detecting the credibility of original records in AI compliance auditing according to claim 1, characterized in that, The association between the audit field and the corresponding content in the original detection record is established through field location identifiers; The field location identifier is associated with the structured field value of the audit field and its position in the original record being inspected. Both the rule audit result and the semantic audit result record the corresponding field location identifier.

4. The method for detecting the credibility of original records in AI compliance auditing according to claim 1, characterized in that, The set of applicable terms is determined by matching the audit context of the detection task with the applicable conditions and version validity range of the audit terms; When audit clauses with inconsistent constraints on the same audit field are matched, the applicable clauses are filtered according to the preset clause priority relationship, and a conflict review mark is generated for audit clauses that cannot be filtered according to the preset clause priority relationship.

5. The method for detecting the credibility of original records in AI compliance auditing according to claim 3, characterized in that, The rule review includes establishing a mapping between review variables and field location identifiers in the rule-state clause data, obtaining the field value of the corresponding review field according to the mapping, and performing deterministic judgment according to the rule-state clause data; The rule review results record the relationship between the review clauses, field location identifiers, and rule judgment process.

6. The method for detecting the credibility of original records in AI compliance auditing according to claim 3, characterized in that, The semantic audit uses the semantic clause data associated with the applicable clause set and the original record content associated with the audit field as the input boundary of the language model, so that the language model generates a semantic audit result within the input boundary, and the semantic audit result contains reference information pointing to the audit clause and original record content on which it is based.

7. The method for detecting the credibility of original records in AI compliance auditing according to claim 6, characterized in that, The verification includes: Based on the cited information, verify whether the audit clauses on which the semantic audit results are based belong to the set of applicable clauses, whether the audit clauses are in a valid version state, and whether there is a supporting relationship between the cited original record content and the corresponding semantic audit results; When all three checks pass, the corresponding semantic audit results will be used for combined judgment. When at least one validation fails, a review identifier is generated for the corresponding semantic audit result.

8. The method for detecting the credibility of original records in AI compliance auditing according to claim 7, characterized in that, The combination determination includes: The rule judgment status is determined based on the triggering of deterministic judgment conditions in the rule review results, and the semantic judgment status is determined based on the identification of non-compliant content in the verified semantic review results. Determine whether the review clauses on which the rule review result and the validated semantic review result are based are consistent, and whether the review fields they target are consistent, to obtain the clause consistency judgment result and the field consistency judgment result; The rule judgment status, semantic judgment status, clause consistency judgment result, and field consistency judgment result are combined into a judgment status combination. The audit issue status corresponding to the judgment status combination is determined in the preset conflict status matrix, and the handling path is determined according to the audit issue status.

9. The method for detecting the credibility of original records in AI compliance auditing according to claim 1, characterized in that, The audit evidence record includes configuring evidence identifiers for audit issues and associating the audit basis, combined judgment process, and handling results when the audit issue was formed with the evidence identifiers. After the review issue is resolved, the associated content is stored in a way that cannot be overwritten, and subsequent review information is appended to the same evidence identifier to restore the formation process of the review issue based on the evidence identifier.

10. The method for detecting the credibility of original records in AI compliance auditing according to claim 9, characterized in that, When the processing path indicates manual review, manual review is performed on the corresponding review issues to obtain the manual review results; Extract review correction information that can be converted into deterministic judgment conditions from the manual review results, and generate candidate rules based on the review correction information; The candidate rules are verified based on the audit evidence records, and the verified candidate rules are used to update the rule state clause data. When at least one of the semantic clause data, rule-based clause data, and language model configuration is updated, a regression audit is performed on the updated audit version based on the audit evidence record, and the release of the audit version is controlled according to the regression audit results.

Images