A professional long-text consistency quality inspection method based on multi-source evidence chain

Abstract

The application discloses a professional long text consistency quality inspection method based on a multi-source evidence chain and belongs to the technical field of artificial intelligence document processing. The method comprises the following steps: identifying version changes through structural signatures, field signatures and content signatures, and determining the quality inspection range; in the quality inspection range, uniformly analyzing the text and the table into candidate evidence units, performing information extraction, numerical reasoning or professional knowledge intensive verification on the statements, and outputting interpretable verification labels; performing bidirectional constraint screening on the multi-source documents and rules, separating formula generation and formula execution, and completing deterministic calculation by an external executor; propagating the change influence through a directed dependency graph, performing re-verification only on the affected objects, saving the statement verification labels and rule satisfaction labels respectively, and jointly outputting the content quality conclusion and the process state. The application can cross-verify the facts, fields, rules and conclusions in the professional long text, and improve the localizability, interpretability and traceability of the quality inspection.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence document processing and content quality inspection technology, and more specifically to a professional long-text consistency inspection method based on multi-source evidence chains. Background Technology

[0002] Professional long-form articles refer to documents that are geared towards a specific industry, discipline, professional activity, or business scenario. They revolve around relatively complex professional tasks, drawing upon multi-source professional materials, structured data, professional knowledge, business rules, or documentary evidence. Employing a hierarchical chapter structure and continuous argumentation, they systematically elaborate on the problem background, factual basis, analysis process, methodological solutions, risk assessment, technical effects, or conclusions and recommendations, resulting in lengthy documents with complete semantic structure, clear application purposes, and adherence to professional writing conventions. The definition of a professional long-form article is not solely based on word count. Its "length" is primarily reflected in the complexity of its material sources, multiple chapter levels, lengthy argumentation chains, and high standards for the basis of its conclusions. Its "professionalism" is primarily reflected in the key content's traceable sources, verifiable facts, consistent data collection methods, reviewable logical processes, and well-founded conclusions and opinions. Due diligence reports, industry research reports, academic papers, technical reports, legal opinions, technical solutions, and consulting reports are all specific types of professional long-form articles.

[0003] Professional long articles are not simply longer articles, but professional documents that complete professional tasks within a considerable scope and require factual evidence, structured argumentation, and verifiable quality mechanisms.

[0004] Existing technologies for generating professional long-form texts mainly focus on data retrieval, outline planning, chapter writing, and language polishing, primarily addressing the question of "how to generate long texts." Existing document quality inspection technologies, on the other hand, are mostly focused on formatting correction, typo checking, text similarity detection, or single-class data verification, making it difficult to cover the complex supporting relationships between facts, values, fields, rules, statements, and conclusions in professional long-form texts.

[0005] Specifically, the existing technology has the following problems:

[0006] First, there is a lack of a verifiable unit decomposition mechanism for long, technical documents. Existing technologies struggle to abstract fields, tables, indicators, citations, paragraphs, declarations, processed fields, evidence locations, and rule chains within long documents into machine-processable quality control objects.

[0007] Second, there is a lack of a unified mechanism for constructing multi-source evidence chains. Existing technologies struggle to unify original materials, structured data, external evidence, business rules, and historical version information into a retrievable, comparable, and traceable evidentiary basis. Therefore, it is difficult to determine whether different sources are consistent, mutually supportive, or conflicting through multi-source cross-verification.

[0008] Third, there is a lack of verification capabilities for the processed results and conclusive statements. Much of the key content in lengthy professional articles is not directly derived from the original materials, but rather obtained through formula calculations, terminology conversions, rule judgments, or professional reasoning. Current technologies typically struggle to trace back the dependent fields, calculation formulas, applicable conditions, and original evidence for such processed results, lacking the ability to replay the rule chain.

[0009] Fourth, there is a lack of incremental review mechanisms for scenarios involving multiple rounds of modifications. Existing quality inspection methods often require a full review after each modification, making it difficult to distinguish between different situations such as unchanged structure, modified content within the same structure, and structural changes. It is also difficult to reuse existing field tables, source tracing, and evidence chains, which can easily lead to duplicate calculations, low quality inspection efficiency, and insufficient audit continuity.

[0010] Fifth, the quality inspection results lack a deliverable format that is locatable, interpretable, verifiable, and traceable. Many solutions can only output error messages, quality scores, or general evaluations, without simultaneously providing the chapter / section where the problem occurred, relevant fields, sources of evidence, reasons for the judgment, risk level, and revision suggestions. Summary of the Invention

[0011] The present invention aims to solve the problems existing in the prior art and provide a quality inspection method for consistency of professional long articles based on multi-source evidence chains. This method can perform structured identification of factual statements, field data, numerical indicators, rule judgments and conclusions in professional long articles, and perform cross-validation based on multi-source evidence chains to form a quality audit result that is locatable, interpretable, verifiable and traceable.

[0012] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a method for quality inspection of professional long-form documents based on multi-source evidence chains, including: Step 1, Professional Long Text Analysis and Quality Inspection Scope Determination: Analyze the text content and table content in the professional long text to be inspected to form candidate evidence units with original text location identifiers; when a historical baseline version exists, obtain the structure signature, field signature, and content signature of the current version relative to the historical baseline version, determine the quality inspection route type based on the comparison results of the candidate evidence units and the three types of signatures, and determine the current quality inspection scope based on the quality inspection route type; Step 2, Explanatory Declaration Verification: Within the scope of the quality inspection, based on the declaration to be verified, a long context processing method or an evidence retrieval enhancement processing method is used to locate the textual and tabular evidence related to the declaration; based on the information extraction, numerical reasoning, or professional knowledge reasoning requirements involved in the declaration, the consistency relationship between the declaration and the evidence is judged, and an explanation of the reasoning process is generated while outputting the verification label; Step 3, Multi-source Field and Rule Verification: Within the scope of the quality inspection, acquire multiple documents related to the verification target and a pre-configured professional rule base; based on document type, time range, and rule applicability, perform bidirectional constraint filtering on the documents and rules to form mutually applicable document-rule pairs; according to the rule type, assign the verification task to the symbolic reasoning path or the arithmetic processing path; wherein, for the arithmetic processing path, formula generation and formula execution are separated, the generative model generates the calculation formula, and the deterministic external executor executes the calculation formula; Step 4, Verification Result Association and Audit Output: The verification results output in Step 2 and Step 3 are associated with each other through a unified object identifier. The impact of changes is propagated through a directed dependency graph, and re-verification is performed only on the affected objects. The final output is a locatable, interpretable, and verifiable quality audit result.

[0013] As can be seen from the above technical solution, compared with the prior art, the present invention has the following technical advantages: This invention primarily addresses the technical problems encountered during consistency quality inspection of multi-source heterogeneous electronic documents and their historical versions. These problems include the lack of stable and calculable relationships between fields, declarations, rules, and evidence; the potential for erroneous reuse of existing verification results after changes in evidence sources or rules; and the repeated processing of irrelevant data objects or missed detection of indirectly affected objects due to partial modifications. Specifically, the invention aims to establish and maintain the relationships between these objects while preserving the original document location, source version, and rule snapshot. It determines the scope of quality inspection based on document changes, evidence validity, and rule validity, and only re-verifies affected objects and their downstream dependencies.

[0014] Professional documents typically include body paragraphs, tables, structured fields, numerical indicators, cited content, rule-based judgments, and conclusions. The supporting evidence is scattered across original materials, structured data, external evidence, and business rules. Existing quality control methods often focus on typos, formatting errors, or single fields, making it difficult to determine the sufficiency of evidence for a particular fact, value, or conclusion, or to pinpoint the specific location and cause of any issues. This invention transforms the above content into fields, statements, processed fields, and conclusion objects with unified identifiers and their original text locations. It establishes a correspondence between these objects and multi-source evidence through unified evidence objects, ensuring that the scope and basis of quality control are clearly recorded.

[0015] This invention employs two processing paths—declaration verification and field and rule verification—to verify natural language expressions, structured fields, numerical calculations, and professional judgments in lengthy professional documents. For fields with inconsistent values ​​from multiple sources, the system saves candidate sources, candidate values, conflict types, adopted results, and judgment reasons. For processing fields requiring calculation or rule derivation, the system backtracks to dependent fields and original evidence, and re-executes the corresponding formulas or judgment rules, thereby improving the verifiability of processing results and conclusions.

[0016] This invention establishes a grading system for evidence strength and a mechanism for downgrading weak evidence. When an evidence only hits an interface status field, general prompt information, or context lacking value-level support, it does not directly form a supporting conclusion but is recorded as questionable, lacking evidence, or interpretably blocked. This mechanism reduces erroneous supporting results caused by evidence that is formally relevant but lacks sufficient content, and clearly distinguishes between missing external materials, incomplete internal processing chains, and abnormal program execution.

[0017] For scenarios involving multiple revisions of lengthy professional documents, this invention identifies version change types through structural signatures, field signatures, and content signatures. For content that has not undergone substantial changes and whose evidence and rule snapshots are still valid, the existing structured quality inspection foundation is reused. For content where fields, text, chapters, evidence, or rules have changed, the changed object is identified according to logical_object_id, and its downstream processing results, statements, and conclusions are included in the re-verification scope along the dependency relationship, ensuring that unaffected objects do not enter the current round of re-examination, while maintaining the continuity of evidence sources and quality inspection records between different versions.

[0018] Therefore, this invention can generate structured quality inspection results that include the original text location, evidence source, verification conclusion, rule execution result, dependency relationship, blocking reason, risk level and revision suggestion, thereby improving the location, interpretability, traceability and reproducibility of consistency quality inspection of professional long documents, and providing unified technical support for the generation, acceptance, manual review, business approval and quality audit of professional long documents. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0020] Figure 1 This is an overall flowchart of the professional long-text consistency quality inspection method based on multi-source evidence chain provided in the embodiments of the present invention; Figure 2This is a flowchart of the interpretable declaration verification process provided in the embodiments of the present invention; Figure 3 This is a flowchart of the multi-source field and rule verification process provided in this embodiment of the invention; Figure 4 This is a flowchart illustrating the differentiated quality inspection routing provided in this embodiment of the invention. Figure 5 This is an example diagram of multi-source field aggregation and source tracing provided in this embodiment of the invention; Figure 6 This is a diagram of Category D missing evidence records in the field evidence classification provided in this embodiment of the invention; Figure 7 This is a diagram showing the multi-source conflict types and adjudication records provided in an embodiment of the present invention. Detailed Implementation

[0021] 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.

[0022] This invention focuses on long professional documents and takes consistency quality inspection as its core task. It involves the structured identification of factual statements, field data, numerical indicators, rule judgments, logical deductions, and conclusions in long professional documents. Based on original materials, structured data, external evidence, business rules, and historical version information, a multi-source evidence chain is constructed. Through multi-source cross-validation, the authenticity, consistency, traceability, and verifiability of long professional documents are tested.

[0023] This invention discloses a method for quality inspection of professional long-form documents based on multi-source evidence chains, including: Step 1, Professional Long Text Analysis and Quality Inspection Scope Determination: Analyze the text content and table content in the professional long text to be inspected to form candidate evidence units with original text location identifiers; when a historical baseline version exists, obtain the structure signature, field signature, and content signature of the current version relative to the historical baseline version, determine the quality inspection route type based on the comparison results of the candidate evidence units and the three types of signatures, and determine the current quality inspection scope based on the quality inspection route type; Step 2, Explanatory Declaration Verification: Within the scope of the quality inspection, based on the declaration to be verified, a long context processing method or an evidence retrieval enhancement processing method is used to locate the textual and tabular evidence related to the declaration; based on the information extraction, numerical reasoning, or professional knowledge reasoning requirements involved in the declaration, the consistency relationship between the declaration and the evidence is judged, and an explanation of the reasoning process is generated while outputting the verification label; Step 3, Multi-source Field and Rule Verification: Within the scope of the quality inspection, acquire multiple documents related to the verification target and a pre-configured professional rule base; based on document type, time range, and rule applicability, perform bidirectional constraint filtering on the documents and rules to form mutually applicable document-rule pairs; according to the rule type, assign the verification task to the symbolic reasoning path or the arithmetic processing path; wherein, for the arithmetic processing path, formula generation and formula execution are separated, the generative model generates the calculation formula, and the deterministic external executor executes the calculation formula; Step 4, Verification Result Association and Audit Output: The verification results output in Step 2 and Step 3 are associated with each other through a unified object identifier. The impact of changes is propagated through a directed dependency graph, and re-verification is performed only on the affected objects. The final output is a locatable, interpretable, and verifiable quality audit result.

[0024] Overall technical solution as follows Figure 1 As shown: (1) Professional long article analysis and quality inspection scope determination; First, the chapters, paragraphs, tables, fields, values, statistical periods, units of measurement, and footnotes in the professional long document are analyzed, and document identifiers, chapter paths, paragraph positions, table positions, and field identifiers are configured for each content unit.

[0025] When no valid historical baseline exists, establish a complete baseline for the structure, fields, declarations, evidence, and rules of the current professional long text. When a historical baseline exists, determine routine quality checks, structural modification quality checks, or structural change quality checks through structure signatures, field signatures, and content signatures. The scope of quality checks includes not only directly changed objects but also declarations referencing changed fields, processing results that depend on changed fields, and related conclusions.

[0026] (2) Explanatory declaration verification process; Declaration verification process as follows Figure 2 As shown: Factual statements, numerical judgments, risk opinions, and conclusive statements are converted into statement objects, and either full context verification or evidence retrieval-enhanced verification is selected based on document length. Supporting or refuting conclusions are only output when the evidence is relevant and complete, and the candidate labels simultaneously meet the minimum acceptance threshold and the minimum probability difference. If no valid evidence is obtained, it is recorded as "incomplete evidence"; if the evidence is incomplete or the label confidence threshold is not met, it is recorded as "questionable," the search is expanded, or it is transferred to manual review. The output results simultaneously save the statement location, evidence location, threshold version, judgment process, and verification reasons.

[0027] (3) Verification process for multiple source fields and rules; Multi-source field and rule verification process as follows Figure 3As shown: Based on the verification objective, document type, and business fact baseline time t ref The rules and applicable conditions implement a two-way screening process for multi-source documents and professional rules. Basic fields are matched for evidence based on field value, subject, period, unit, field name, chapter, and source type; weak evidence cannot directly form a supporting conclusion. In multi-source conflicts, inapplicable and non-substantive candidates are first excluded, and then adjudication is made only within the set of strong evidence candidates according to the frozen source hierarchy. If strong evidence is missing or the highest candidate is tied, the case is transferred to manual review. Processed fields are traced back to basic fields and original evidence according to necessary and optional dependencies; rules are not executed if necessary dependencies are not closed.

[0028] The symbolic reasoning path is used for content extraction, condition judgment, and multi-stage dependency processing; the arithmetic processing path separates formula generation, formula verification, and formula execution, with deterministic calculations performed by an external executor. Polynomial formulas are verified through symbolic simplification or polynomial identity checks; formulas containing division, branching, or piecewise operations are also verified against the expression tree, domain, and boundary conditions. Pass or Fail is only output when the rule path is clear, necessary dependencies are complete, and rule execution is successful; in cases of insufficient evidence, rule conflicts, no applicable rules, formula verification failure, or execution anomalies, only the corresponding object status and process status are recorded.

[0029] (4) Differentiated quality inspection process for three types of signatures Differentiated quality inspection routes, such as Figure 4 As shown: When all three types of signatures remain unchanged, existing structured quality control foundations can be reused. However, weak evidence, missing evidence, open dependency chains, and manually reviewed objects cannot directly reuse the final conclusion. When fields or content change, the quality control scope expands to the affected processing fields, statements, and conclusions; when the chapter structure changes, the relationships between fields, statements, evidence, and rules in the changed chapters are re-established. The current version can only be registered as a new valid baseline after the quality control process is fully executed and the document-level quality conclusion is compliant, or after manual confirmation.

[0030] After determining the scope of differentiated quality inspection, the system constructs a directed quality inspection dependency graph for the current version using `object_version_id`, and establishes cross-version object correspondences using `logical_object_id`. When fields, paragraphs, evidence, rules, or manual decisions change, the corresponding nodes are used as change seeds, propagating downstream along field input dependencies, rule bindings, declaration references, evidence support, and conclusion references, forming a set of affected objects. Downstream results are recalculated when necessary dependencies change, and the completeness of results is re-evaluated when optional dependencies change; unaffected completed objects continue to be reused.

[0031] Before implementing impact propagation, dependency cycles are detected. Cycles without registered convergence conditions or those that have not converged after reaching the maximum number of iterations are recorded as `dependency_cycle` and become unacceptable blocks. Manual review results are written back as independent adjudication records, without overwriting the original machine records; the written-back objects serve as new change seeds, triggering downstream re-examinations. During task execution, checkpoints and idempotent execution keys are saved per object. In case of execution anomalies, retrying is performed from the most recent checkpoint. After the blockage is cleared, documents, evidence, and rule hashes are verified first, and then recovery is performed from the unfinished stage.

[0032] (5) Verification results correlation and audit output: The statement verification label indicates the supporting or refuting relationship between the statement and the evidence; the rule fulfillment label indicates whether the document content meets the applicable rules. These two labels are stored separately and do not substitute for each other. When the two types of results are inconsistent, a cross-level verification conflict is recorded and the process is transferred to a review phase.

[0033] Based on execution anomalies, internal link integrity, and the sufficiency of evidence, a status is determined: execution completed, explainable blockage, unacceptable blockage, or execution failed. Within the scope of this round of quality inspection, a document-level quality conclusion is generated, indicating compliance, pending review, or non-compliance. Finally, the content quality conclusion and process execution status are jointly output as Ω(t)=(Q(t),Γ(t)), along with logical_object_id, object_version_id, original document location, source of evidence, verification conclusion, rule execution result, dependencies, blockage reason, checkpoint, manual adjudication record, recovery conditions, risk level, and revision suggestions. No single result component may overwrite another component.

[0034] In summary, this invention integrates professional long-text analysis, multi-source evidence chain construction, declaration verification, field and rule verification, differentiated review, and audit output into a complete technical solution.

[0035] The following section provides explanations of the interpretable declaration verification mechanism, multi-source consistency verification mechanism, and quality inspection routing: (i) An interpretable declaration verification mechanism based on the context of long-form mixed content; In one embodiment of the present invention, addressing the issues of lengthy professional documents, mixed distribution of text and tables, scattered declaration-related information in different locations, and lack of verifiable evidence for verification conclusions, the text and table content in the professional documents are uniformly parsed to form candidate evidence units with original text location identifiers; based on the declaration to be verified, a long context processing method or an evidence retrieval enhancement processing method is selected to locate the textual and tabular evidence related to the declaration; based on the information extraction, numerical reasoning, or professional knowledge reasoning requirements involved in the declaration, the consistency relationship between the declaration and the evidence is judged; while outputting verification tags, an explanation of the reasoning process is generated, thereby forming an interpretable verification result composed of the declaration to be verified, relevant evidence, reasoning process, and verification conclusion.

[0036] The long professional documents to be processed are represented as follows:

[0037] in, This indicates a long, technical document that needs to be processed. This indicates a sequence of text content within a long, technical article. Indicates the first Each text content unit; Indicates the sequence number of the text content unit; Indicates the total number of text content units; This represents a sequence of tables within a long, technical article. Indicates the first A table; Indicates the table number; This indicates the total number of tables.

[0038] Text content units include titles, body paragraphs, table titles, table descriptions, footnotes, and other text content with independent semantics. When parsing text content, a unique location identifier is assigned to each text content unit, and its chapter, hierarchical path, paragraph number, and its order in the original professional long text are recorded, so that subsequent verification results can be traced back to the corresponding original text location.

[0039] For the table content, the table title, column headers, row headers, cell values, units of measurement, statistical periods, footnotes, and table descriptions are extracted. Following the original row and column relationships, column headers or cell values ​​within the same row are separated by vertical bars, and different rows are separated by line breaks, thus converting the two-dimensional table into serialized table content. During serialization, table identifiers, row and column positions, and the correspondence between table headers and cell values ​​are preserved to avoid mismatches in indicator names, statistical periods, units of measurement, or data values ​​after conversion. The parsed text content units and the serialized table together constitute the candidate evidence unit set. .

[0040] Pending verification statement This indicates. (Statement) Statements can be factual statements, field statements, numerical statements, comparative statements, trend statements, calculation results, logical judgments, professional judgments, or conclusive opinions. Statements can be identified from lengthy professional documents through a preceding document parsing process or can be manually specified. For each statement to be verified, at least the subject, object, time, indicator, numerical value, unit, comparative relationship, and conclusion must be identified; for elements not present in the statement, the corresponding fields can be empty.

[0041] This invention employs either a long context processing method or an evidence retrieval enhancement method, depending on the length of the professional text and the context processing capability of the verification model. The long context processing method is suitable when the serialized complete professional text does not exceed the effective context length of the verification model. Under this method, the text content is sequenced... Table sequence and Declarations to be verified The common input verification model enables the verification model to locate evidence and make judgments within the entire document.

[0042] Evidence retrieval enhancement processing is suitable for situations where complete professional texts exceed the effective context length of the verification model, or where it is necessary to reduce interference from irrelevant content. Under this processing method, claims are calculated separately. The relevance score between each candidate evidence unit and the candidate evidence unit is calculated, and the units with higher relevance scores are selected. There are 10 candidate evidence units. The evidence retrieval results are represented as follows:

[0043] in, In response to the statement Before formation A sequence of evidence; This indicates a long professional article The set of candidate evidence units obtained in the process; Represents any candidate unit of evidence; Statement With candidate evidence units The correlation score between them; Indicates the number of candidate evidence units selected; This indicates that the top score is selected based on relevance score. One candidate unit of evidence; Indicates candidate evidence unit Its position within the original long-form academic paper; This indicates that the selected candidate evidence units are reordered from front to back according to their original positions in the text.

[0044] Correlation score Candidate evidence units can be obtained through sparse retrieval methods based on term matching, dense retrieval methods based on contrastive learning, or embedding retrieval methods based on semantic vectors. Term matching-based retrieval is suitable when the statement and evidence contain the same subject name, technical terms, field names, periods, or numerical expressions; semantic vector-based retrieval is suitable when the statement and evidence use different wording but have the same or similar semantics. (Number of candidate evidence units) The number of candidate evidence units can be set to 3, 5, or 10; when it is necessary to improve the coverage of scattered evidence, the top 10 candidate evidence units can be selected. After the candidate evidence units are selected, they are rearranged according to their order of appearance in the original professional text to preserve the contextual relationships between paragraphs, tables, and text and tables.

[0045] After forming a sequence of candidate evidence, the relevance and completeness of the evidence are checked. Evidence relevance refers to whether the candidate evidence units contain facts, data, or judgment bases that can support or refute the statement; evidence completeness refers to whether the existing evidence covers the necessary subjects, objects, periods, indicators, values, units, and logical relationships to form a verification conclusion. Both textual and tabular evidence can be used as the basis for statement verification; when a statement involves both textual descriptions and tabular data, the corresponding textual and tabular content units should be extracted simultaneously, and a verification conclusion should not be drawn based on only one type of content. If the evidence completeness check fails, the statement is marked as "insufficient evidence," and the number of candidate evidence units is increased or the long context processing mode is switched for re-verification; no support or refutation labels are output before the evidence completeness check passes.

[0046] Regarding the statement The resulting verification context is represented as follows:

[0047] in, Indicates the use of verification statements The context; This refers to a complete, lengthy professional document containing both text and tables. This refers to the sequence of evidence formed through evidence retrieval enhancement processing.

[0048] Based on the main reasoning methods required for the claim, claim verification can be divided into information extraction verification, numerical reasoning verification, and professional knowledge intensive verification.

[0049] Information extraction verification is used to process statements for which verification evidence can be directly obtained from the text or tables of lengthy professional documents. The verification process first identifies the subject, event, time, field, value, and limiting conditions in the statement; then, it locates the corresponding content within the verification context and records the chapter, paragraph, table, and cell location of the evidence; finally, it compares the content recorded in the evidence with the content of the statement item by item. If the evidence is consistent with all the decisive elements of the statement, the statement is deemed supported by the evidence; if any decisive element contradicts the evidence, the statement is deemed refuted by the evidence.

[0050] Numerical reasoning verification is used to handle claims that require calculations or statistical analysis based on data from a lengthy professional text before a judgment can be made. The verification process begins by locating all original values ​​related to the claim and recording the corresponding subject, indicator, period, unit, symbol, and evidence location for each value. Then, based on the semantic relationships in the claim, it determines summation, difference, ratio, period variation, percentage calculation, or other appropriate operational relationships. Calculations are performed step-by-step according to the determined operational relationships, and intermediate results are recorded. Finally, the calculated results are compared with those stated in the claim. The claim is considered supported by evidence only when the original data, statistical period, unit of measurement, data scope, operational relationships, and calculation results all match the claim. If the numerical results in the claim are inconsistent with the recalculated results, or if the data subject, period, unit, or statistical scope used in the claim is inconsistent with the evidence, the claim is considered refuted by the evidence.

[0051] Specialized knowledge-intensive verification is used to handle statements that require the application of specialized concepts, industry knowledge, or legal rules to interpret factual information in a document. The verification process first extracts relevant facts and data from the detailed text, then identifies the specialized concepts, industry rules, or legal rules required to make a judgment, and explains how this specialized knowledge applies to the extracted factual information and whether the conclusions of the statement can be drawn from it. Specialized knowledge or legal rules are used to interpret the factual information and its logical meaning, but do not replace the factual evidence in the detailed text; in the absence of corresponding factual evidence, the specific facts in the detailed text cannot be established solely based on general specialized knowledge.

[0052] For statements that are evidence-relevant and complete, set a valid set of verification labels:

[0053] in, Represents the set of valid verification tags; This indicates that all the decisive elements of the statement are supported by verified context; This indicates that the statement contradicts the verification context, or that some of the key elements of the statement contradict the verification context.

[0054] The declaration verification label is determined according to the following formula:

[0055] in, Indicates the final verification label; Represents a set of tags Any candidate label in; The parameter is The conditional probability determined by the verification model; Indicates the verification model parameters; Indicates the context of the declaration verification; Indicates a statement to be verified; condition symbol " "In the context of verification" and statement Under the given conditions; This means selecting the label with the highest conditional probability from the candidate labels.

[0056] The aforementioned highest probability label is only used as a candidate label. The highest candidate probability and the second highest candidate probability are denoted as p_l^(1) and p_l^(2), respectively, and the pre-frozen lowest acceptance threshold and minimum probability difference are denoted as θ. l and δ l Only if p_l^(1)≥θ l And p_l^(1)-p_l^(2)≥δ l Furthermore, only when the evidence integrity check passes is the candidate label accepted as the final verification label; otherwise, an entailed or refuted label is not generated, and candidate evidence is added, the verification method is switched, or the process is transferred to manual_review. l and δ l It should be saved along with the model version and running configuration, and should not be temporarily adjusted based on a single sample to be verified.

[0057] After determining the verification labels, the verification model generates an explanation of the reasoning process used to illustrate the evidence extraction, data processing, and judgment process:

[0058] in, This represents an explanation of the final generated reasoning process; This indicates an explanation of the candidate reasoning process; This represents the set of candidate reasoning process explanations; The parameter is The conditional probability determined by the verification model; Indicates the context of the declaration verification; This indicates a statement pending verification. Indicates the final verification label that has been determined; condition symbol " "In the context of verification" ,statement and final verification label Under the given conditions; This indicates that the candidate reasoning process explanation with the highest conditional probability is selected from the set of candidate reasoning process explanations.

[0059] To ensure the completeness and consistency of the reasoning process, the verification model first reads the verification context and the statement to be verified. Then, it lists the facts, data, and evidence locations related to the statement. Next, it performs information comparison, numerical calculation, or professional knowledge reasoning according to the statement type, recording each intermediate processing step. Finally, it outputs a conclusion consistent with the final verification label based on the above processing. The reasoning process explanation for information extraction-type statements should at least include the decisive content of the statement, the location of the evidence, the content recorded in the evidence, and the comparison results item by item. The reasoning process explanation for numerical reasoning-type statements should at least include the original numerical values ​​involved in the calculation, the evidence location of each value, the indicator name, the statistical period, the unit of measurement, the operational relationship, the intermediate calculation results, the final calculation result, and its comparison with the statement result. The reasoning process explanation for professional knowledge-intensive statements should at least include the documented facts, the professional concepts or rules used, the interaction between the professional concepts or rules and the documented facts, and the final judgment basis.

[0060] After the verification model outputs its results, the verification labels and reasoning process explanations are obtained through response parsing. Response parsing can employ preset field parsing rules or a separate response parsing model. The parsing result is... or When the model outputs meaningless symbols, incomprehensible text, or cannot extract valid labels, the response parsing result is recorded as invalid output and transferred to re-verification or manual review. Invalid output is not directly converted into supporting or refuting conclusions.

[0061] To enable the aforementioned verification mechanism to be configured, validated, or evaluated as a model, verification samples with evidence location and reasoning process explanations can be constructed based on professional long-form articles. First, textual or tabular context is extracted from different locations within the professional long-form article. Personnel with relevant domain knowledge then formulate statements supported by the original textual evidence based on the extracted content, and locate all textual and tabular evidence supporting this statement within the complete professional long-form article. Subsequently, factual perturbations are applied to the subjects, values, periods, relationships, or conclusions of the supported statements, causing the perturbated statements to directly contradict the located evidence, thus forming statements refuted by the evidence. After the factual perturbation, the reasoning process explanation is simultaneously rewritten, ensuring that the comparison or calculation process recorded in the explanation corresponds to the perturbated statements.

[0062] After the statement and evidence are labeled, an interpretation of the reasoning process is developed by personnel different from those who formed the statement, and verification labels are determined accordingly. The interpretation of the reasoning process is organized according to a standardized format: first, all relevant information for the statement is listed; then, the extraction, comparison, calculation, or professional judgment processes performed on each piece of information are described step by step; finally, verification labels are provided. If the labels determined during the statement formation stage are inconsistent with those determined during the reasoning process interpretation stage, an independent reviewer is brought in to re-examine the statement, evidence, and reasoning process, and determine the final labels.

[0063] The resulting verification samples undergo quality verification, which includes: the accuracy and completeness of the statements and reasoning explanations; whether the statements pertain to the relevant professional field and have practical verification significance; whether the located evidence is relevant to the statements and sufficient to complete the verification; whether the verification labels are supported by the located evidence; and whether the reasoning explanations correctly interpret the evidence and employ appropriate reasoning steps. Samples that fail quality verification should be modified or excluded and should not be directly used for the configuration, validation, or evaluation of the verification model.

[0064] When performing quality checks on verification results, at least the following should be checked: relevance of evidence, completeness of evidence, accuracy of evidence citation, correctness of reasoning process, and consistency between labels and interpretations. Relevance of evidence is used to determine whether the cited text paragraphs or tables are relevant to the statement; completeness of evidence is used to determine whether the cited evidence is sufficient to support the verification conclusion; accuracy of evidence citation is used to determine whether the values ​​and facts cited in the interpretation are consistent with the original text; correctness of reasoning process is used to determine whether the comparison, calculation, or professional judgment process used is correct; and consistency between labels and interpretations is used to determine whether the final conclusion of the reasoning process is consistent with the output labels.

[0065] Errors occurring during the verification process are categorized into information extraction errors, numerical reasoning errors, insufficient professional knowledge, and calculation execution errors. Information extraction errors refer to failure to locate relevant context, or, although the relevant table is located, incorrect reading of the table header, row / column relationships, or cell values. Numerical reasoning errors refer to correct extraction of raw data, but incorrect identification of quantitative relationships in the statement or selection of incorrect operational relationships. Insufficient professional knowledge refers to having obtained relevant documentary facts, but lacking the professional concepts, industry knowledge, or regulatory knowledge required to complete the judgment. Calculation execution errors refer to correct operational relationships, but errors occurring during intermediate or final calculations. Error types and statement verification labels are recorded separately to determine the handling method: re-retrieval, re-reasoning, or manual review.

[0066] The final interpretable verification result includes at least the statement to be verified, statement type, verification context type, location of textual evidence, location of tables and cells, content recorded in the evidence, evidence integrity status, explanation of the reasoning process, verification label, response parsing status, and error type. This verification result establishes a clear link between the statement to be verified in the long professional text and the corresponding textual evidence, tabular evidence, and reasoning process, forming a locatable, interpretable, and verifiable in-text evidence chain, and providing standardized statement verification results for subsequent construction of multi-source evidence chains and multi-source cross-validation.

[0067] (ii) A multi-source consistency verification mechanism based on bidirectional constraints of documents and rules and hybrid reasoning; The aforementioned interpretable statement verification mechanism is used to establish the correspondence between statements, evidence, and verification conclusions within a single professional long-form document. Building upon this, this invention further addresses the issues of multiple heterogeneous documents involved in the same business matter, field dependencies between different documents, verification requirements constrained by professional rules, and numerical calculation results easily affected by generative model errors. It establishes a multi-source consistency verification mechanism based on bidirectional constraints of document sets and rule sets. This mechanism jointly filters relevant documents and applicable rules, performs format-preserving structured extraction and auxiliary information enhancement on document content, and assigns verification tasks to symbolic reasoning paths or arithmetic processing paths according to rule types, ultimately forming a multi-source evidence chain containing source documents, applicable rules, evidence fields, execution formulas, calculation results, verification tags, and processing records.

[0068] The aforementioned statement verification label and Used to indicate the supporting or refuting relationship between a statement and the evidence within the text; used in this section and These tags are used to indicate whether a document meets the applicable rules. The two types of tags have different judgment objects and semantics, and should be recorded separately; they must not be substituted for each other or used interchangeably.

[0069] The collection of documents related to the same verification item is represented as:

[0070] in, This represents the set of documents to be verified. Indicates the first A document; Indicates the document sequence number; This indicates the total number of documents. Documents can be scanned PDFs, document images, or page images converted from corresponding files, and can have different page numbers, layouts, document types, and field organization methods. The verification target specified by the user or generated by the system is represented as... The pre-configured professional rule base is represented as .

[0071] rule base Each rule in the framework includes at least a rule identifier, rule content, applicable document type, input field requirements, field dependencies, applicable time conditions, verification relationships, output field requirements, and exception handling conditions. The verification relationships can be field existence relationships, conditional branching relationships, multi-stage logical relationships, unconstrained arithmetic relationships, or arithmetic relationships with precision, priority, and missing value constraints.

[0072] Each rule should also register its version, source, issuing or maintaining entity, professional scope of application, applicable entity type, effective date, expiration date, authority level, mandatory attributes, previous version, subsequent version, current status, and rule content hash value. Rule status should include at least draft, active, suspended, supervised, and retired.

[0073]

[0074] in, This indicates that the quality inspection object q is at the business fact baseline time. The set of valid rules below; The set of rules is represented by r; candidate rules are represented by r. Indicates that rule r is in The corresponding historical status is determined by the rule status change record and the rule version snapshot; and These represent the effective time and expiration time of rule r, respectively; This indicates whether rule r applies to the quality inspection object q; a value of 1 indicates applicability, and a value of 0 indicates non-applicability. When no fixed failure time is set, Take +∞. The baseline time corresponding to the facts to be verified, the statistical period, or the business activity should be used, not the actual system runtime. The actual system execution time is denoted as t. exec It is only used for task scheduling, checkpoints, and audit documentation.

[0075] When multiple rules are matched for the same quality inspection target, first determine if the rules belong to the same source and are different versions with the same scope of application. If a subsequent effective version has clearly replaced a previous version, the subsequent version is used. If the rules have different scopes of application, they are executed separately and the results are saved separately. If the rules have different authority levels or mandatory attributes, they are handled according to the pre-configured and traceable professional rule hierarchy. For rules with the same scope of application and authority level but incompatible verification requirements, they are recorded as rule_conflict, the rule satisfaction label is set to empty, and the process is transferred to manual_review, without forcing the selection of one of them. Each quality inspection saves a snapshot of the rules actually used, the rule version, and the rule content hash value.

[0076] This invention maps verification targets, document sets, and professional rule bases into a set of verification results:

[0077] in, This indicates the multi-source consistency verification process; Indicates the verification objective; This represents the set of documents to be verified. This represents a professional rule base; Represents the set of verification results; Indicates the number of verification results; Indicates the first Verification results; Expressing support for the first A subset of documents containing the verification results; Indicates the first A subset of rules used in the verification process; This indicates that the rule satisfies the label, where This indicates that under the current conditions, no binary rule satisfies the label; This indicates the processing status, and its values ​​include at least completed, insufficient_evidence, no_applicable_rule, formula_validation_failed, execution_failed, and manual_review, which respectively represent completed rule execution and generated a valid label, insufficient evidence, no applicable rule, formula validation failure, execution failure, and need for manual review; Indicates that it is possible to trace back to the first Explanation and record of the process by which the verification results are formed.

[0078] The aforementioned processing statuses are further subdivided as follows: When the object itself is not within the scope of the rule's applicability, it is recorded as "not_applicable" and is not included in the rule quality evaluation set; when the object should be verified but no applicable rule exists within the professional scope, it is recorded as "no_applicable_rule" and is treated as an interpretable block; when the rule should exist but is missing due to rule base configuration, version, or status errors, it is recorded as "rule_configuration_missing" and is treated as an unacceptable block; when the rule exists but the original material lacks necessary input, it is recorded as "insufficient_evidence" and is treated as an interpretable block. None of the above statuses generate a valid Pass or Fail label.

[0079] Object-level processing status This also includes `conflict_pending_review`, `manual_review`, `rule_conflict`, `rule_configuration_missing`, `rule_dependency_undefined`, `dependency_cycle`, `formula_validation_failed`, and `execution_failed`. `conflict_pending_review` indicates that candidate value conflicts persist after automatic adjudication; `rule_conflict` indicates that valid rules with the same scope and authority level contradict each other; after the system creates a manual review task, the corresponding object is transferred to `manual_review`. After manual adjudication is written back and necessary downstream re-checks are completed, the object leaves `manual_review` and enters the `completed`, `insufficient_evidence`, or corresponding abnormal state. `pending_review`, `failed`, `unacceptable_blocked`, and `explained_blocked` are not used for object-level states; `pending_review` is only used for document-level quality conclusions Q(t), and `failed`, `unacceptable_blocked`, and `explained_blocked` are only used for process-level states Γ(t). Object-level completed and process-level completed are bound to respectively Γ(t) and Γ(t) represent the completion of processing a single object and the completion of the current process, respectively. These two values ​​cannot be substituted for each other across layers. The terms nonconforming and conforming are only used for document-level quality conclusions Q(t).

[0080] Multi-source consistency verification includes interrelated document processing paths and rule processing paths. The document processing path sequentially completes document preprocessing, document type identification, document screening, structured content extraction, and content filtering. The rule processing path sequentially completes verification target parsing, rule recall, rule applicability filtering, rule type identification, and auxiliary information enhancement. The two processing paths are bidirectionally constrained by document type, time range, and rule applicability scope, ensuring that documents and rules entering the subsequent verification process form mutually applicable document-rule pairs.

[0081] During document preprocessing, page orientation correction, skew correction, layout normalization, and image quality processing are performed on each document. Character errors, field breaks, and row / column misalignments in the text recognition results are also corrected. Page numbers, text block coordinates, table coordinates, and visual region coordinates are preserved during preprocessing, allowing subsequent field values ​​to be traced back to the original document page and specific area.

[0082] After preprocessing, document type identification is performed based on page-level visual features. Document type can be determined using a lightweight image classification model, or by combining headings, fixed fields, page layout, and identifier information. Professional documents of the same type typically have similar page structures and field distributions; the document type identification results are used to limit the applicability of the rules.

[0083] Simultaneously, document-level metadata, such as document date, data coverage period, subject identifier, and document type, is extracted from each document. This metadata can be obtained through template-guided named entity recognition, regular expressions, and rule-based heuristics. Verification Target The time range, target fields, target body, and target document type are extracted by the target parsing module.

[0084] The time constraint parsed from the verification objective is represented as follows: , will document The time attribute is represented as The collection of documents filtered by time is represented as:

[0085] in, This represents a collection of documents that meet the verification time requirements. Document The document date or data coverage period; Indicates the verification target Time constraints in; symbols This indicates that the document's time attribute meets the corresponding time constraints. For time requirements such as "recent periods," "specified years," or "a certain statistical period," this formula excludes documents whose time range does not meet the requirements.

[0086] During the rule recall process, a semantic encoding model is used to convert the verification target and rule content into semantic vectors, respectively. The semantic relevance score between the verification target and the rule is represented as follows:

[0087] in, Indicates the verification target With rules Semantic relevance score between them; This represents a semantic encoding function; A semantic vector representing the verification target; A semantic vector representing the content of a rule; the symbol " " represents the vector dot product; This represents the vector norm. The top few rules based on their relevance scores are selected to form an initial set of candidate rules.

[0088] Rules The applicable set of document types is represented as , will document The recognition type is represented as The rule set and document set after bidirectional document type constraints are represented as follows:

[0089]

[0090] in, This represents the set of candidate rules obtained after semantic recall and document type constraints. This indicates a rule that selects the top few items based on their relevance scores; Representation rules The set of applicable document types; Represents a collection of documents The collection of document types included; Represents the empty set; This represents the set of candidate documents obtained after applying time constraints and rule applicability constraints. This represents the union of all document types to which the candidate rules apply.

[0091] The above processing forms a closed-loop filtering relationship of document constraint rules and rule-backed document constraints. The time and semantic conditions in the verification objective are used to filter documents and recall rules; document identification type is used to exclude inapplicable rules; and the applicable document type of candidate rules is used to exclude documents irrelevant to all candidate rules. This reduces the interference of irrelevant documents and rules on subsequent model inputs.

[0092] If no applicable document rule pair is found after two-way filtering, the processing status will be changed. Record it as no_applicable_rule or insufficient_evidence, and make the rule satisfy the label. Valid Pass or Fail tags will not be generated. Knowledge extraction and rule execution will only proceed if the document type, time range, rule scope, and required fields all meet the verification conditions.

[0093] When a field or document does not need to be validated by the rule based on its preset scope of application, it is recorded as not_applicable; when the validation target requires the execution of a rule but none of the candidate rules are applicable, it is recorded as no_applicable_rule; when a rule should be applicable but cannot be executed due to incomplete internal configuration of the rule base, invalid rule version, or missing necessary rule records, it is recorded as rule_configuration_missing. These three types of situations must not be treated interchangeably.

[0094] For the candidate document set The present invention converts the document into a hybrid content representation that maintains its layout structure. The definition of mixed content representation is:

[0095] in, Document Layout-preserving hybrid content representation; Represents text paragraphs and semantic text units obtained from a document; Represents a table and its row and column structure; This indicates form fields and field groups; This indicates the spatial grouping relationships and semantic dependencies between text blocks, fields, and table cells; This indicates the non-text visual area that has been preserved; It represents metadata such as document date, coverage period, language, document type, and subject identifier.

[0096] Hybrid content representation does not use a flat key-value structure containing only field names and values. Instead, it uses HTML-like tags or other hierarchical tags to record page hierarchy, chapter hierarchy, table row and column relationships, form field grouping, positional relationships between text blocks, and dependencies between fields. Each structured content unit can be represented as:

[0097] in, Indicates the first Each structured content unit; Indicates the type of content unit, including text, tables, form fields, or visual areas; This represents the text content obtained through recognition; This indicates the bounding box coordinates of the content unit within the page. Indicates the page number; Indicates the confidence level of text recognition or layout recognition; This indicates the parent chapter, table, form, or group node to which this content unit belongs.

[0098] The structured transformation first obtains text areas, table areas, form areas and other visual areas through layout analysis; then it recovers free text, discrete text and table cell content through character recognition, and merges the split characters, terms or text lines into continuous semantic units based on spatial proximity.

[0099] Merge text blocks that are on the same line, overlap vertically, and have a horizontal spacing less than a preset threshold; reconstruct paragraphs from text lines with the same left margin, similar line spacing, and consecutive reading order; restore the relationships between table headers, row headings, column headings, and data values ​​for cells within the same table area whose row and column coordinates correspond. For tables spanning multiple pages, determine whether tables on consecutive pages belong to the same logical table based on repeated headers, number of columns, column widths, and field names.

[0100] For charts, seals, signatures, and other visual content that cannot be reliably converted into text, the determination of whether to retain the corresponding image region is based on text coverage and text recognition confidence. The set of visual regions can be represented as:

[0101] in, Document The set of visual regions that are preserved in the middle; Represents a candidate visual region; Indicates the effective text coverage of the candidate visual region; The confidence level for text recognition in candidate visual regions; Indicates the text coverage threshold; This represents the confidence threshold for character recognition. This indicates a logical OR. When the text coverage is below a threshold or the character recognition confidence is below a threshold, the corresponding region is retained as an image patch and input into the multimodal verification model, and is not deleted due to an empty character recognition result.

[0102] After completing the structured transformation, content filtering is performed on headers, footers, duplicate template content, empty placeholders, and fixed descriptions irrelevant to the verification target. Filtering criteria include document object structure, field tags, page position, and frequency of repetition. Content filtering must not delete fields required by the rules, field definition descriptions, units of measurement, statistical periods, or footnotes that could change the meaning of fields.

[0103] During the auxiliary information enhancement process, language type, document type, technical terms, abbreviations, and field predicates are extracted from layout-preserving mixed content and candidate rules. For low-frequency terms, technical abbreviations, and technical concepts located on the left side of field labels or rule expressions, corresponding terminology explanations are obtained and converted into concise semantic descriptions.

[0104] For each candidate rule Language type, document type, technical terms and their explanations are collectively represented as auxiliary information. Representing rule-related mixed content as The output field structure required by the rule is represented as follows The dynamic verification template is represented as follows:

[0105] in, Indicates the rules The generated dynamic verification template; Represents the template constructor; Represents structured text, tables, forms, and visual content related to rules; It displays auxiliary information such as language, document type, and terminology explanations; Indicates the current rule; This indicates the required output field structure according to the rules.

[0106] The dynamic verification template adopts a step-by-step zero-sample structure and does not rely on fixed examples. The template includes at least the document type, document language, format-preserving structured content, table content, form content, semantic fields, applicable rules, explanations of technical terms, and output format. Whether each part is included and the granularity of its inclusion are jointly determined by the document structure and rule type to reduce the amount of irrelevant content occupying the model context.

[0107] According to the rules, the verification task is further divided into five types: content extraction, conditional logical reasoning, multi-stage logical reasoning, unconstrained arithmetic consistency checking, and constrained arithmetic consistency checking. Content extraction, conditional logical reasoning, and multi-stage logical reasoning enter the symbolic reasoning path; unconstrained arithmetic consistency checking and constrained arithmetic consistency checking enter the arithmetic processing path.

[0108] The rule type identification result is represented as follows:

[0109] in, Representation rules The corresponding verification path type; Indicates the candidate path type; symbolic indicates a symbolic reasoning path; arithmetic indicates an arithmetic processing path. The parameter is The conditional probabilities determined by the rule-based classification model; This represents the parameters of the rule-based classification model; argmax represents the candidate path type with the highest conditional probability.

[0110] The path with the highest probability output by the rule classification model is only used as a candidate path. The probability of the highest path and the probability of the second highest path are represented as p_λ^(1) and p_λ^(2), respectively, and the pre-frozen minimum acceptance threshold and minimum probability difference are represented as θ_λ and δ_λ, respectively. The model classification result is adopted only when p_λ^(1)≥θ_λ and p_λ^(1)-p_λ^(2)≥δ_λ; if the threshold is not met, the path is determined first based on the rule type already registered in the rule base. If the rule type is not registered or there is still ambiguity, it is transferred to manual_review. Low-confidence paths are not directly handed over to subsequent executors.

[0111] Content extraction is used to check if a target field exists and what its value is. For each target field, output at least its presence status and value. When a field exists, set its presence status to true and return the original value retrieved from the document; when a field does not exist, set its presence status to false and set its value to an empty string. Missing fields must not be filled with values ​​from other fields.

[0112] Conditional logic reasoning is used to handle rules containing branch conditions such as "if," "otherwise," "skip," "valid," or "invalid." During verification, condition judgments are executed sequentially according to the order specified in the rule, and the triggered branches are recorded. When a rule explicitly stipulates that subsequent checks should be skipped under a certain condition, the skip status and corresponding reason are output; the skipped condition must not continue to be executed, nor should the determined intermediate result be changed due to the skip.

[0113] Multi-stage logical reasoning is used to handle rules that require multiple dependent steps to reach a conclusion. Fields, calculated values, or judgment states obtained in the previous stage serve as input for the next stage. Each stage records the input fields, applicable conditions, intermediate results, and output fields. If a necessary field is missing, causing a stage to fail, the corresponding result is set to null, and subsequent stages dependent on that result are terminated; missing intermediate results should not be inferred.

[0114] For example, when a rule needs to determine the expected number of stages based on the start and end times and cycle type, and then compare it with the number of stages recorded in the document, the time field extraction and cycle rule matching should be completed first, then the expected number of stages should be calculated, and finally a consistency comparison should be performed. If any prerequisite field is missing, the comparison result should be recorded as "cannot be determined," rather than directly outputting "consistent" or "inconsistent."

[0115] The symbolic reasoning module outputs only the structured fields specified by the rules. Spaces in field names are converted to underscores; missing field values ​​are set to empty strings; dates are output in a uniform date format; Boolean judgments use true, false, or null values; the output content is in structured object format, without any additional explanations, introductions, or free text unrelated to the rules. The structured output is used for subsequent rule judgments and audit record assembly.

[0116] Unconstrained arithmetic consistency checks are used to derive calculation relationships based on relevant data actually appearing in the document. Examples include summing all detailed items, calculating ending values ​​based on beginning values ​​and period increases / decreases, or calculating differences based on totals and items. These rules do not pre-define priority fields or exception handling methods, but every value in the generated formula must originate from located evidence.

[0117] In addition to performing numerical calculations, the constraint arithmetic consistency check must also meet the precision, field priority, missing value handling, and exception conditions specified by the rules. For example, the rules may stipulate that the result retains a certain number of decimal places; when a total field exists, the total field is used first, and detailed items are not summarized again; calculation formulas are not generated when necessary multipliers or measurement fields are missing or zero; and subtotals in a table must not be included repeatedly with their respective details.

[0118] For arithmetic processing rules, this invention separates formula generation from formula execution. The verification model is only responsible for identifying the target field, document declaration value, evidence fields involved in the calculation, and operational relationships, and generating calculation formulas that can be executed by an external executor; the actual calculation is completed by a deterministic external executor, avoiding numerical instability caused by the generative model directly performing arithmetic operations.

[0119] The generated formula should include at least the target field name, the document declaration value, and the calculation formula. The target field name should be normalized using a uniform character set; the document declaration value must come from the located evidence field and be converted into a valid numerical value; if the declaration value does not exist, it should be recorded as a missing value. The calculation formula should only use the original numerical values ​​and basic arithmetic operators already obtained from the evidence, and should not contain fictitious variables, equal signs, calculation results, or explanatory text.

[0120] Requirements involving rounding, precision, or tolerance are handled uniformly by the external executor according to the rule parameters. Unnecessary functions that might alter the original computational relationships are not added to the generated formula. The formula generation module is only responsible for describing the computational relationships between numerical values; the external executor is responsible for deterministic calculations and result formatting.

[0121] The first The formula for generating the wheel is expressed as follows: The output of the independent formula verification module is represented as follows: The formula generation and verification process can be represented as follows:

[0122]

[0123] in, Indicates the first The calculation formula for wheel generation; The parameter is Formula generation model; This represents relevant structured evidence; Indicates the current rule; Indicates auxiliary information; This indicates a formula generated in the previous round but which failed the validation. This indicates the reason why the formula failed the validation in the previous round; Indicates the first Wheel formula verification results; The parameter is Independent formula verification model; This indicates that the formula has passed validation; This indicates that the formula failed the validation.

[0124] When generating the formula for the first time, and Set to empty. Formula validation includes checking whether all fields involved in the calculation come from the located evidence, whether the field subject and period are consistent, whether the operation relationship conforms to the rules, whether subtotals are used incorrectly or fields are repeated, whether the plus or minus signs are consistent with the original text, and whether any necessary operation items required by the rules are omitted.

[0125] when When this happens, the incorrect formula and the reason for failure are used as negative feedback input into the next round of formula generation. The maximum number of retries is... In one specific implementation, Set to 2. If the verification still fails after reaching the maximum number of retries, the processing status will be changed. Record it as formula_validation_failed, and make the rule satisfy the label. The process will then proceed to manual review, and formulas that have not been verified will not be executed.

[0126] The formula for verification is expressed as follows: The numerical value of the evidence cited in the formula is expressed as follows: External execution and rule judgment are represented as follows:

[0127] in, This indicates the calculation result obtained by the external actuator; This represents a function that executes deterministic formulas. This indicates a calculation formula that has been verified and accepted. This indicates the formula input values ​​obtained from relevant documentation evidence; This indicates a state where the rule is satisfied; Represents a rule-based judgment function; Indicates the current rule; It represents the structured content, field status, unit of measurement, statistical period, and other limiting information related to the current rule.

[0128] The rule evaluation function compares the calculated result with the document declaration value, rule threshold, or conditional relationship based on the rule type. It outputs "Pass" if the rule is met and "Fail" if not, and also indicates the processing status. Record as "completed". Record the corresponding processing status when necessary fields are missing, formulas fail validation, or external execution fails, ensuring the rule meets the tag. Binary verification labels are not mandatory.

[0129] This invention does not directly use the free text reasoning process within the model as the basis for auditing, but instead forms an interpretation record based on the structured records of each module. The explanation record should include at least the verification target, the selected documents and their types, the time-based filtering results, the applicable rules and rule identifiers, the evidence fields and page locations, explanations of technical terms, the verification path type, the conditional branches, the intermediate states, the formula input values ​​and their sources, the generated formulas, the formula validation results, the reasons for negative feedback, the number of retries, the external execution results, the rule satisfaction status, and the abnormal status.

[0130] To detect whether the verification process introduces evidence not found in the original document or formulas that do not conform to the rules, a fact illusion indicator is set during the model validation or runtime sampling phase. For documents... and rules The sample-level factual illusion indicator is represented as:

[0131] in, Document In the rules Did the following occur as a hallucination? This indicates an indicator function; it takes the value 1 if the condition within the parentheses is true, and 0 otherwise. This represents the set of evidence actually used in the verification process; Document In accordance with the rules The corresponding set of reference evidence; This indicates that not all of the evidence actually cited is included in the set of referenced evidence; Indicates logical OR; This represents the formula generated during the verification process; Representation rules The corresponding reference formula; This indicates that the two formulas are not equivalent in terms of algebraic relations.

[0132] Before performing inclusion relationship checks on textual evidence, capitalization, punctuation, numerical formats, and units of measurement are standardized. For polynomial formulas consisting only of addition, subtraction, multiplication, and non-negative integer powers, variable names and term orders are first standardized, and then equivalence is determined through symbolic simplification or multiple rounds of random assignment polynomial identity tests; if the results at any valid assignment point are different, the formulas are deemed unequal. For formulas involving division, absolute values, maximum values, minimum values, conditional branches, piecewise functions, or other non-polynomial operations, random assignment is not used as the sole criterion for equivalence. Instead, standardized expression trees, operator order, variable dependencies, domain restrictions, and branch conditions are compared, and tests are performed on denominators of zero, both sides of thresholds, points of sign change, and boundary values ​​specified by the rules. For symbolic reasoning rules that do not involve formulas, evidence inclusion relationships, conditional order, branch triggering conditions, and conclusion mapping relationships are compared.

[0133] Factual illusions include two situations: one is referencing fields or values ​​that do not exist in the document, or using other field values ​​to replace missing fields; the other is that although the referenced field actually exists, the generated calculation relationship is inconsistent with the applicable rules. The calculation process should not be considered correct simply because the final result is the same as the value declared in the document.

[0134] For cases where the evidence fields and generation formulas are correct, but the external execution results or numerical derivation results are incorrect, set a numerical distortion indicator:

[0135] in, Document In the rules Does numerical distortion occur? Indicates an indicator function; This indicates that the corresponding document rules do not pertain to the illusion that no factual event occurred; This indicates that a deterministic calculation is performed on the generating formula; Indicates the generating formula; This indicates the corresponding reference value result; Indicates the absolute error tolerance; Indicates the relative error tolerance; This indicates that, given absolute and relative error tolerances, the formula execution result is inconsistent with the reference result. (This is achieved by...) If set as a prerequisite, errors in evidence or formulas will not be counted as numerical distortion.

[0136] The numerical distortion index only identifies inconsistencies in calculations that occur when the evidence and formulas are correct. Cases where evidence values ​​are fabricated or formula relationships are incorrect are already classified as factual illusions and are not counted again in the numerical distortion index, thus distinguishing between evidence errors, formula errors, and calculation execution errors.

[0137] For complex scenarios such as missing values, outliers, non-standard field combinations, mixed subtotals and details, negative adjustment items, and rule boundary conditions, set a boundary case error indicator:

[0138] in, Indicates whether a verification error occurred in the boundary case; This indicates the rule-satisfaction state of the model output; This indicates that the reference rule is satisfied. This indicates that the two are not equal.

[0139] Boundary case checks are used to identify whether the system incorrectly counts subtotals as duplicates of details, incorrectly changes the sign of negative adjustment items, makes inferences about missing fields, incorrectly handles null and zero values, and correctly executes priority fields and exception conditions in the rules.

[0140] For factual illusions, numerical distortions, and boundary case errors, error rates can be calculated separately according to rules and rule categories. The unified aggregation form is expressed as:

[0141] in, Representation rules The corresponding error rate; Representation rules A collection of evaluation documents; Indicates the number of documents evaluated; This represents the corresponding sample-level error indication quantity, which can be taken as... , or ; This indicates rule categories that have the same verification type; Indicates rule category Weighted error rate; weights Representation rules The corresponding number of evaluation samples. When calculating the boundary case error rate, Only pre-identified boundary case document rule pairs are included.

[0142] The final multi-source consistency verification result should include at least the verification target, relevant document set, document type and time range, document selection criteria, applicable rules, rule selection criteria, rule type, structured evidence fields, evidence page and area location, explanation of professional terms, symbolic reasoning results or arithmetic formulas, formula input values ​​and sources, formula verification records, negative feedback reasons, number of retries, external execution results, rule satisfaction status, processing status, factual illusion type, numerical distortion status, and boundary case error status.

[0143] The aforementioned interpretable statement verification mechanism establishes a correspondence between statements and evidence within a single professional long-form document. Its output statement type, evidence location, and structured fields can serve as semantic fields in the hybrid content representation described in this section. The multi-source consistency verification mechanism further incorporates multiple related documents, structured fields, and professional rules into a two-way constraint and cross-validation process, and generates repeatable rule verification results through symbolic reasoning, external arithmetic execution, and formula verification loops. Together, these two parts constitute a complete quality inspection chain from in-text evidence location to cross-document, cross-field, and cross-rule consistency verification.

[0144] The aforementioned two verification mechanisms transmit verification results to the differentiated quality inspection routing mechanism through a unified object identifier and structured data interface. The professional long-text version object includes at least article_id, version_id, baseline_version_id, structural signature, field signature, and content signature; the unified evidence object includes at least evidence_id, document_id, source type, source version, source hash value, retrieval time, chapter path, page or paragraph position, body, period, unit, original text, and normalized value.

[0145] The declaration verification mechanism outputs declaration identifier, associated field identifier, cited evidence identifier, support or rebuttal label, evidence integrity status, processing status, and structured interpretation record; the multi-source rule verification mechanism outputs rule identifier, rule version, input field identifier, cited evidence identifier, execution path, rule satisfaction label, formula verification result, external execution result, and processing status; the multi-source conflict adjudication mechanism outputs field object identifier, candidate source and value, conflict type, final adopted value or unadjudicated status, and evidence strength.

[0146] The differentiated quality inspection routing mechanism establishes associations between the above results based on article_id, version_id, logical_object_id, object_version_id, field_id, rule_id, and evidence_id, and sequentially completes version change identification, affected object determination, evidence reuse validity judgment, conflict resolution, dependency chain replay, declaration and rule result association, and document-level quality judgment. logical_object_id is used to identify the same logical object across versions, and object_version_id is used to locate object instances in a specific version; any output object can be traced back to the input document location, candidate field values, evidence source, applicable rules, and intermediate processing status.

[0147] (III) Differentiated quality inspection routing, local reconstruction, and collaborative review mechanism based on three types of signatures; The aforementioned interpretable declaration verification mechanism is used to establish the correspondence between internal declarations, evidence, and verification conclusions in professional long documents, while the aforementioned multi-source consistency verification mechanism is used to cross-validate related documents, structured fields, and professional rules. To make both verification mechanisms applicable to multi-round revision scenarios of professional long documents, this invention further establishes a differentiated quality inspection routing mechanism based on structural signatures, field signatures, and content signatures. This mechanism identifies the type and range of changes in the current document relative to the baseline document, determines the quality inspection objects that need to be reused, re-verified, or partially reconstructed, and forms interpretable state control for weak evidence, unclosed processed field dependency chains, and evidence boundaries.

[0148] The current version of the long professional document to be inspected is represented as ,in, The current version number is indicated; the valid version saved during the generation phase or manually confirmed is indicated as the baseline version. The baseline version must be associated with at least the article identifier, generation task identifier, chapter structure, paragraph structure, field workbook, field source tracking, evidence object, applicable rules, and the previous round of quality inspection records. The current version and the baseline version are linked through the same article identifier; identical file names are not the sole basis for identification.

[0149] The current version is structurally parsed, retrieving the collection of chapter rows, paragraphs, and field bindings in the original document order. Each chapter row records at least the chapter number, chapter ID, chapter title, chapter level, chapter identifier, number of paragraphs, and document identifier; each paragraph records at least the paragraph number, paragraph identifier, chapter, paragraph text, and paragraph text hash; each field binding records at least the field name, field value, standard field identifier, subject role, statistical period, unit of measurement, chapter position, paragraph position, and match confidence.

[0150] Based on the structure parsing results, generate the structure signature, field signature, and content signature respectively:

[0151]

[0152]

[0153] in, This indicates the structure signature of the current version; This indicates the field signature for the current version; This indicates the content signature of the current version; Represents a deterministic hash function; This indicates a deterministic serialization function that is executed in the order of predetermined fields. It represents a sequence of structural features consisting of chapter number, chapter title, chapter level, and number of paragraphs; This represents a sequence of field characteristics consisting of chapter position, field name, standard field identifier, subject role, normalized field value, statistical period, and unit of measurement. Indicates the current version of the The text hash of each paragraph; Indicates paragraph number; This indicates the total number of paragraphs in the current version; the symbol "‖" indicates that they are connected in document order. By incorporating normalized field values, statistical periods, and units of measurement into the field signature, substantial field changes in the field workbook can trigger corresponding quality control routes, even if the main text remains unchanged.

[0154] The structural signature is used to determine whether the document skeleton has changed; the field signature is used to determine whether field binding relationships, normalized field values, statistical periods, or units of measurement have changed; and the content signature is used to determine whether the main text content has changed if the chapter and field structures remain unchanged. Before generating the signature, the headings, field names, field values, and paragraph text are normalized using a unified character encoding, whitespace characters, line breaks, and date formatting, but numbers, units of measurement, plus or minus signs, and punctuation marks that can change the professional meaning are not deleted.

[0155] The baseline presence identifier is represented as When a baseline version exists that can establish a valid correspondence with the current version, When no valid baseline version exists, This invention compares the three types of signatures in the current version with the corresponding signatures in the baseline version, and the quality inspection route is represented as follows:

[0156] in, This indicates the current version of the quality inspection route; Indicates structural change quality inspection; This indicates quality inspection of structural modifications; This indicates routine quality inspection; , and These represent the structural signature, field signature, and content signature of the baseline version, respectively. This indicates a logical AND operation. Indicates logical OR; This indicates that the two signatures are not equal. The three routing conditions are mutually exclusive, and each current version only enters one of the main quality inspection routes.

[0157] When no available baseline version exists, the current version will be processed as a structural change for quality inspection. A complete baseline for the structure, fields, declarations, evidence, and rules will be established for the current professional document, and quality inspection will not be directly suspended due to the lack of a baseline. If a baseline exists but some baseline artifacts are missing, the missing artifacts will be regenerated based on the current document first. Only if a complete positioning relationship cannot be established after regeneration will the corresponding blocking state be entered. The temporary signature of the current version must not automatically overwrite the baseline signature. Only after the quality inspection process is fully executed and the document-level quality conclusion is conforming, or after manual confirmation, can it be registered as a new baseline for subsequent version comparisons.

[0158] When all three types of signatures remain unchanged, the process proceeds to routine quality control. Routine quality control reuses the baseline version's field workbook, field rule bindings, evidence source tracking, and document location index, but does not directly treat all historical verification conclusions as the final conclusions of the current version. Objects in the historical record that fall under the categories of weak evidence hit, contextual evidence hit, missing evidence, unclosed dependency chains, or those requiring manual review are still included in the current quality control scope; only objects whose document content has not changed, whose evidence location is still valid, and whose evidence strength in the previous round met the requirements can have their structured verification basis reused.

[0159] The fact that the document's structure signature, field signature, and content signature remain unchanged does not automatically mean that the external evidence and rule conditions have not changed. Before reuse, the version of the evidence source, source hash value, validity of the evidence location, rule version, and rule validity period should be compared; if any condition changes, the corresponding fields, processed fields, statements, and conclusions should be brought back into the quality inspection scope. Historical verification data should only be reused if the document content remains unchanged, the evidence location is still valid, the evidence source and rule snapshot remain unchanged, and the strength of the previous round of evidence meets the requirements.

[0160] When the structure signature remains unchanged, but the field signature or content signature changes, the system proceeds to the same structure modification quality check. The system uses `logical_object_id`, defined later in this section, as the cross-version field key; this field key simultaneously includes the article identifier, object type, stable semantic key, subject key, statistical period, and unit of measurement. If the field has a statistical scope attribute, the statistical scope is encoded into the stable semantic key. Based on this field key, the system compares the baseline field values ​​with the current field values ​​to generate a set of changed fields.

[0161] in, This represents the set of changed fields in the current version; and These represent the sets of field keys for the current version and the baseline version, respectively. It represents the symmetric difference between two sets of field keys, used to retrieve newly added and deleted fields; Represents the intersection of sets; This indicates a field key that exists in both versions. and These represent field keys respectively. Field values ​​in the current version and baseline version; This represents a field value normalization function; the symbol " The expression indicates a set of elements that satisfy the following conditions. Field value normalization does not change the numerical sign, unit of measurement, or statistical period. This expression includes newly added, deleted, and changed field values ​​in separate sets of changed fields, avoiding comparisons of non-existent field values.

[0162] In addition to changing fields, the system also identifies paragraphs whose content has changed based on paragraph identifiers and paragraph text hashes. The actual scope of structural modification quality inspection includes changed fields, changed paragraphs, declarations referencing changed fields, processing fields that take changed fields as input, intermediate fields belonging to the same rule chain as changed fields, and table rows and chapters containing the above objects. This avoids overlooking the cascading effects of cross-chapter references, processing results, and conclusions by only verifying direct modifications.

[0163] When the structural signature changes, a structural change quality check is initiated. The system obtains the chapter identity sets for both the baseline version and the current version. A chapter identity is determined by the chapter sequence number, chapter number, chapter title, and chapter level. For each chapter identity, a chapter row fingerprint is generated based on the number of paragraphs contained in the chapter and the paragraph text hashes.

[0164] in, Indicates the chapter identity in the current version. Corresponding chapter line fingerprint; This indicates the chapter's identity, which consists of chapter number, chapter ID, chapter title, and chapter level. This indicates the number of paragraphs contained in the chapter; This represents a hash vector of paragraph text arranged in the original order within a chapter. and The meaning is the same as the aforementioned signature generation process.

[0165] The set of changed chapters is represented as follows:

[0166] in, This represents the set of chapters that have changed from the baseline version in the current version; and These represent the set of chapter identities in the current version and the baseline version, respectively. Used to retrieve newly added and deleted chapters; and These represent the chapter row fingerprints for the same chapter identity in the current version and the baseline version, respectively. When a chapter is added, deleted, or its chapter row fingerprint changes, it is included in the set of changed chapters.

[0167] The system re-extracts field bindings, evidence blocks, and verifiable declarations from the changed chapters, and re-establishes the relationships between fields and standard fields, declarations and evidence, and rules and dependent fields. For fields appearing in the changed chapters and referenced by other chapters, the quality inspection scope extends outward along the processing field dependencies and declaration associations; chapters that have not changed and are not affected by dependencies can reuse the original structured quality inspection foundation.

[0168] Within the quality inspection scope defined by various routes, this invention uses a unified evidence object to implement field verification and claim verification. Each evidence object includes at least an evidence identifier, source type, source file, content block type, chapter path, page or paragraph location, evidence text, and metadata. Field verification, rule chain replay, and claim verification reference the same evidence identifier, thereby enabling the results obtained from different verification layers to be correlated in a unified audit record.

[0169] When multiple candidate values ​​are obtained for the same quality inspection object, a grouping key is established for the candidate values ​​according to article identifier, version identifier, subject role, subject identifier, logical field identifier, statistical period, and unit of measurement. The set of candidate values ​​for field f is represented as follows: Each candidate value At least include the original values Normalized value Subject, Statistical Period, Unit of Measurement, Data Scope, Source Identifier, Source Type, Source Version, Source Priority Strength of evidence Confidence level Source location and applicability markings .in, N(·) represents a processing function that normalizes character formats, numerical precision, date formats, units of measurement, and equivalent expressions.

[0170] Each candidate value should also be registered with its source authority level. and the completeness of the evidence content and source location . Determined based on the issuing entity, legal validity, or level of business authority; This indicates the pre-configured source priority for the target field within the same authority level; the two should not be confused. Arranged in descending order of strong, context, weak, and missing; and Use normalized values ​​with the same dimension or pre-configured discrete levels. The above attributes, value ranges, and sorting rules are all saved with the version of the adjudication rules.

[0171]

[0172] in, Represents the set of valid candidate values; This refers to a set of non-substantive values, including null values, undisclosed values, not found values, values ​​with only headers, values ​​with only interface status information, or values ​​that cannot be located to specific evidence content. Options whose sources are not applicable to the target field, only return status fields or general prompts, or do not form substantive field values ​​will not participate in the final decision.

[0173] The system categorizes multi-source discrepancies based on normalized values, subject, period, unit, data caliber, and source applicability into four types: consistency, equivalent expression, format or unit difference, time caliber difference, coexistence of missing and valid values, inapplicable source, source reliability difference, semantic ambiguity, and material conflict. Consistent or equivalent expressions from multiple sources use a common normalized value and merge source records; convertible format or unit differences are reclassified after unifying the units; candidate values ​​with different time calibers are registered to their corresponding periods without mandatory merging; candidates with inapplicable sources are excluded; missing values ​​do not overwrite existing valid values. Only differences in source reliability, semantic ambiguity, or material conflict proceed to subsequent automatic adjudication, with strong evidence gating applied first.

[0174]

[0175] in, Represents the set of valid candidate values medium strength of evidence For the set of candidates for strong; This indicates the final candidate value to be adopted; argmax_lex indicates that the comparison is performed lexicographically according to the listed indicators. Indicates the level of authority of the source; This indicates the pre-configured source priority for the target field within the same authority level; Indicates the confidence level of the candidate value; This indicates the completeness of the evidence's content and source. Evidence strength serves only as a preliminary gate for automatic adjudication and is no longer grouped with authority level or source priority in the same ranking hierarchy.

[0176] when When empty, do not automatically select a value from the context, weak, or missing candidates, leaving the final value empty and recording it as conflict_pending_review. If the highest-ranking candidate is unique and not empty, it is determined as the final candidate value; if multiple highest-ranking candidates have the same normalized value or belong to equivalent expressions, their source records are merged and the common value is adopted; if multiple highest-ranking candidates have different normalized values ​​and the sorted tuples are completely identical, no arbitrary selection is performed, the final adopted value is left empty, the object-level processing status is recorded as conflict_pending_review, and all parallel candidate values, sorted tuples and evidence positions are retained.

[0177] For candidate sets with consistent sources, equivalent expressions, and deterministic unit conversions, a common normalized value is directly formed, retaining all sources. For differences in source reliability, semantic ambiguity, or substantial conflict, a final value is formed only when the strong evidence candidate set is not empty, the applicable rules clearly specify the source level, and the highest ranking result is unique. If the above conditions cannot be met, candidate values ​​are not forcibly selected, the final value is set to empty, and recorded as conflict_pending_review. The model can be used for conflict type identification and candidate relationship judgment, but the model output must be mapped to existing candidate values ​​and evidence identifiers; non-candidate values ​​generated by the model must not be directly used as the final value.

[0178] To prevent quality inspection routing errors from reusing historical results with insufficient evidentiary strength, this invention determines the matching strength between fields and evidence. Fields With candidate evidence The matching score between them is expressed as:

[0179] in, Representation field With candidate evidence Match score; This indicates whether the normalized field value appears precisely in the evidence; Indicates whether the value in the field appears in the evidence; Indicates whether the subject to which the field belongs is consistent with the subject of the evidence; Indicates whether the statistical period of the field is consistent with the evidence period; Indicates whether the unit of measurement of the field is consistent with the unit of evidence; Indicates whether the field name appears in the evidence; Indicates whether the module or chapter to which the field belongs matches the location of the evidence; Indicates whether the field keyword appears in the evidence; This indicates whether the source type of the field registration is consistent with the source of the evidence; This indicates whether the candidate evidence meets the criteria for weak evidence. All functions are binary indicator functions, taking the value 1 when the condition is true and 0 otherwise.

[0180] , , , , , , , and These represent the non-negative weights of each positive matching feature; This represents the penalty weight for weak evidence that is greater than zero. Each weight is pre-configured based on the field type and professional scenario. The weight for a precise match of a field value is no less than the weight of any single contextual matching feature. When the statistical period or unit of measurement is a necessary condition of the rule, the corresponding weight must not be set to zero, thereby preventing the formation of strong evidence conclusions based solely on the relevance of field names, chapters, or keywords.

[0181] field The candidate evidence set is represented as The candidate evidence with the highest score is represented as:

[0182] in, Representation field The set of candidate evidence; argmax represents the candidate evidence with the highest score; argmax indicates selecting the candidate with the highest score. This represents the minimum threshold for evidence. When the candidate evidence set is empty or... In such cases, this field is recorded as unlocated evidence.

[0183] The evidence strength status of field f is represented as follows: Its value set is {strong, context, weak, missing}. Here, strong indicates value-level evidence hit; context indicates contextual evidence hit; weak indicates weak evidence hit; and missing indicates unlocated evidence. A strong evidence threshold is set. ,and ≥ The consistency threshold is A(f,e), calculated below according to the necessary verification dimensions of the fields. The strength of evidence status is determined by the following formula:

[0184] The state conditions are determined in the order of missing, weak, strong, and context. An empty candidate evidence set or a score lower than [a certain threshold] is considered a valid state condition. When the evidence strength status is "missing", the evidence strength status is "weak" when the candidate evidence reaches the minimum threshold but meets the weak evidence condition; the highest score is not lower than When the consistency threshold is 1 and the weak evidence condition is not met, the evidence strength status is "strong". For other candidate evidence that has reached the minimum evidence threshold but does not meet the strong evidence condition, the evidence strength status is "context". "Strong" can form a supporting conclusion; "context" and "weak" form a questionable state; "missing" forms a missing evidence state; when there are clearly opposite field values ​​and the subject, period, unit of measurement and statistical caliber are consistent, a rebuttal conclusion can be formed.

[0185] For consistency gating of field values, subjects, statistical periods, and units of measurement, a universal gating quantity calculated based on necessary dimensions is uniformly adopted. Field values ​​are mandatory dimensions; whether subjects, periods, and units participate in gating is pre-registered based on field type and applicable rules. The strength of evidence is not reduced simply because a field does not have a period or unit attribute, even if that dimension is not applicable. The universal gating quantity is determined according to the following formula:

[0186] Where A(f,e) represents the consistency gating quantity between field f and candidate evidence e; This indicates that the field values ​​are consistent; when Take in sequence , , hour, Corresponding to , and , respectively represent fields With candidate evidence The value is 1 if the subject, statistical period, and unit of measurement are consistent, and 0 otherwise.

[0187] for example This indicates whether dimension x is a necessary validation dimension for field f; it takes the value 1 if necessary and 0 otherwise. When a dimension is not applicable to field f, the corresponding... If the value is 0, the corresponding product term will always be 1. The aforementioned evidence strength state uses only A(f,e) as defined in this formula.

[0188] Weak evidence includes at least the following situations: the field value only matches a status field, code field, request parameter, or general prompt field in an external interface; the field value equals the subject name, but the field to be verified is not a subject name field; the evidence only contains the field name or related context, but not the field value to be verified; the subject, statistical period, or unit of measurement of the evidence is inconsistent with the field; source tracing only points to a file or interface and cannot locate the specific content. Weak evidence cannot directly form a supporting conclusion.

[0189] Weak evidence downgrade results are included in quality control routing. Even if the three types of signatures remain unchanged in the current version, fields that were in the context evidence hit, weak evidence hit, questionable, or missing evidence status in the previous round cannot be directly reused as already completed verification objects. When new independent strong evidence is obtained in this round, the evidence score can be recalculated and the corresponding downgrade status can be removed; if no new strong evidence is obtained, the original status and its reasons are retained, and missing evidence is not supplemented by model speculation.

[0190] For processing fields that cannot be directly obtained from raw materials and need to be generated based on basic fields and professional rules, this invention performs dependency chain replay within the scope determined by differentiated routing. The processing field is represented as g, and its applicable rules are represented as... The complete set of dependency fields for rule registration From the necessary dependency field set and optional dependency field set Composition, satisfaction = ∪ and ∩ = The set of optional dependency fields that have already obtained available values ​​at the current execution time is represented as: Only in the complete dependency chain state. When =1, the processing result is retrieved again based on the rules, necessary dependency fields, and available optional dependency fields:

[0191] in, This represents the result of the processed field obtained by recalculating or deriving according to the rules; Exec represents the deterministic rule execution function; This represents the rules corresponding to the processing field g, including applicable conditions, operational relationships, judgment criteria, accuracy requirements, and exception handling conditions; Indicates the set of necessary dependent fields; This represents the set of optional dependency fields for which usable values ​​can be obtained; j represents the dependency field index. This represents the field value that depends on field j; =1 indicates that all necessary dependencies have been met. Missing optional dependencies do not change the status. However, the system should record any missing information and reduce the result integrity level according to the rules; when the rules explicitly require an optional field to participate in a specific branch, that field should be converted into a necessary dependency within that branch.

[0192] Dependency chain replay not only checks whether the dependent fields have numerical values, but also examines the subject, period, unit of measurement, evidence status, and verification conclusion of each dependent field. Required dependency fields are used to determine whether the processing fields can be executed and produce valid results; optional dependency fields are only used to supplement the completeness evaluation and do not participate in the multiplication determination of the completeness of the dependency chain.

[0193] For processing fields, necessary dependency fields and optional dependency fields should be registered separately. If a necessary dependency field is not declared, the product of an empty dependency set cannot be interpreted as a complete dependency. The set of necessary dependency fields for processing field g is represented as follows: The unique dependency chain complete state is represented as And determine according to the following formula:

[0194] Where g is the processing field to be verified; Its unique dependency chain complete state is 1 when all necessary dependency conditions are met, and 0 otherwise; it is also 0 if the set of necessary dependency fields is not declared. For the necessary dependency field set; This is the i-th necessary dependency field; , , and They are respectively The field values, evidence strength status, dimension fit, and verification availability. The definition of field evidence strength status is the same as that in the previous text. =1 means The principal key must be consistent with the target principal of the processed field, and the statistical period, unit of measurement, and statistical scope declared in the rule must be consistent; when a dimension is registered as inapplicable by a rule, that dimension must not be used. It dropped to 0. =1 means No refuted or failed conclusions have been reached; the system is not in the conflict_pending_review, manual_review, rule_conflict, formula_validation_failed, execution_failed, rule_configuration_missing, rule_dependency_undefined, or dependency_cycle states; and there are no unresolved cross-layer verification conflicts. The fact that a certain type of verification label is not applicable to d_i does not affect... . =strong、 =1 and When both conditions are true (e.g., =1), the necessary dependent field can be included in the rule execution. I(·) is an indicator function; it takes the value 1 if the condition is true, and 0 otherwise. It is an empty set.

[0195] When a processing field lacks a documented target value, its evidence strength is recorded as "missing," and the object status is recorded as "insufficient_evidence" or "not_applicable" depending on whether the field is required. When the verification target requires rule execution but no applicable rule exists, it is recorded as "no_applicable_rule" and an interpretable block is formed. When a rule should exist but is missing due to rule base configuration, version, or status errors, it is recorded as "rule_configuration_missing" and an unacceptable block is formed. When a rule fails to declare necessary dependency fields, resulting in an uncertain inference relationship, it is recorded as "rule_dependency_undefined" and an unacceptable block is formed. When all necessary dependencies are missing, it is recorded as "insufficient_evidence." When some necessary dependencies are missing, only context or weak evidence exists, dimensions are mismatched, or verification is unavailable, the rule is not executed and enters manual_review or insufficient_evidence. Rules are only executed when C_g=1. If the result of re-execution matches the result recorded in the document, it will be considered as support or pass; if the result does not match, it will be considered as rebuttal or fail. The declaration verification label and the rule satisfaction label are still saved separately.

[0196] For processing fields with precision or tolerance requirements, first standardize the units of measurement, numerical scale, and percentage expression, then recalculate the results and compare them with the values ​​recorded in the document according to the absolute and relative error tolerances registered in the applicable rules:

[0197] in, This indicates the result of the processing field recalculated according to the rules; Indicates the value recorded in the document; Indicates the absolute error tolerance; This indicates the relative error tolerance. For conditional processing fields, intermediate states are replayed according to the branching order specified by the rules. If any prerequisite necessary dependency field is missing, subsequent derivations dependent on that field are not executed, and the missing dependency is not replaced with a default value or other field value.

[0198] When a base field changes, the system searches for all processing fields that directly or indirectly use that field along the dependency relationships and adds them to the scope of this round of quality inspection. When the result of a processing field changes, it further searches for declarations and conclusions that reference that processing field. This forms an impact propagation chain of "field change - dependency rule - processing result - associated declaration," preventing the reuse of invalid processing results or conclusion verification results after partial field modifications.

[0199] Within the same quality inspection scope, the aforementioned declaration verification results, multi-source rule verification results, and multi-source conflict adjudication results are associated using article_id, version_id, document_id, logical_object_id, object_version_id, paragraph_id, field_id, rule_id, and evidence_id. To distinguish between logical object identities that remain stable across versions and object instances in a specific version, logical_object_id is generated according to the following formula:

[0200]

[0201] Wherein, H(·) represents the deterministic identifier generation function; article_id represents the professional long article identifier that remains unchanged across versions; object_type represents the object type, such as basic field, processing field, declaration, conclusion, rule, or evidence; semantic_key represents the stable semantic key of the object, with fields identified by logical field identifiers, rules by rule identifiers, declarations and conclusions by declaration type, associated logical field, and chapter semantic key, and evidence by source record identifier and evidence type; entity_key represents the entity key composed of entity role and entity identifier; period and unit represent the statistical period and unit of measurement, respectively. logical_object_id does not include version number and original text position, and is used for matching the same logical object between the current version and the baseline version. version_id represents the specific version; locator represents the chapter, table, page, paragraph, or coordinate position; content_hash represents the normalized content hash of the object in this version. object_version_id is used to record specific version instances, original text positions, and audit results. Cross-version comparisons first establish a correspondence based on logical_object_id, and then determine whether the content or position has changed based on object_version_id.

[0202] The explanatory text generated by the model is only saved as an audit record if it can be parsed into a structured record and its declaration identifier, field identifier, rule identifier, and evidence identifier can all be linked back to a unified object; free text that cannot be linked back is not used as verification basis. Declaration verification labels and rule satisfaction labels are saved separately. If the two are inconsistent, the record is recorded as a cross-level verification conflict and transferred to review. One result is not used to directly override the other result.

[0203] Declaration verification labels and rule fulfillment labels are saved separately. In declaration verification... and Used to indicate the supporting or refuting relationship between a statement and the evidence within the text; in rule verification and This is used to indicate whether a document meets applicable rules. The two are linked through a unified quality inspection object, but are not directly replaced one-to-one. When declaration verification and rule verification yield results in different directions, they are recorded as cross-level verification conflicts and transferred to a review process; one result does not directly override the other.

[0204] To ensure the aforementioned impact transmission chain has an executable data structure, the system constructs a directed quality inspection dependency graph. Nodes in the graph must include at least basic fields, processing fields, declarations, conclusions, rules, and evidence objects; directed edges must include at least field input dependencies, rule bindings, declaration references, evidence support, and conclusion references. The system generates directed edges based on the necessary and optional dependency fields, field-paragraph binding relationships, declaration association fields, and evidence reference records registered in the rule table. For each edge, the system registers the upstream object identifier, downstream object identifier, relationship type, necessity, applicable version, validity period, and source record. Nodes use `object_version_id` as the primary key for the current version and `logical_object_id` as the cross-version mapping key; duplicate edges are merged, and edges with missing endpoints or version mismatches are considered internal link anomalies and are not included in the executable graph.

[0205] When a field value or paragraph content changes, evidence becomes invalid, a rule version is replaced, or a manual ruling is rewritten, the system registers the corresponding node as a change seed and propagates the impact downstream along directed edges. The propagation process uses a pending queue, an visited set, and an affected object set: after retrieving a node from the pending queue, its unvisited downstream nodes that are within the current applicable version and validity period are added to the affected object set and the pending queue; when necessary dependencies change, downstream results are set to pending recalculation; when optional dependencies change, the original results are retained but their integrity is re-evaluated. When a field, evidence, or rule is deleted, an invalidation flag is saved, and results that directly or indirectly depend on that object are set to expired. Objects independent of change seeds are not included in this round of re-examination.

[0206] After the dependency graph is constructed, the system identifies self-loops and multi-node loops through depth-first traversal or strongly connected component detection. The system iterates in a fixed order only when the rules constituting the loop have registered initial values, a fixed execution order, convergence criteria, and a maximum number of iterations. The loop is considered converged when the normalized results of two consecutive iterations no longer change, or when the error falls within the tolerance range specified by the rules. If the above conditions are not registered, convergence is not achieved after exceeding the maximum number of iterations, or contradictory outputs exist, execution within the loop is stopped, `dependency_cycle` is recorded, and the process transitions to an unacceptable blocking state. Loop-free propagation terminates when the processing queue is empty and all affected objects enter the final states such as completed, pending review, blocked, or execution failed. It also stops execution and records the blocking reason when the preset maximum propagation depth or the maximum number of visits to a single node is reached.

[0207] To differentiate between normal evidence boundaries, quality control defects, and execution anomalies, this invention establishes four process states: execution completed, explainable blockage, unacceptable blockage, and execution failed. The current version's process state is represented as follows:

[0208] in, Indicates the current process status; Represents the set of execution failures; This represents the set of events that disrupt the integrity of the internal link. A collection of evidence or matters that obstruct professional boundaries; The four states are: empty set, failed, unacceptable_blocked, explained_blocked, and completed. The order of priority for each state is: failed, unacceptable_blocked, explained_blocked, and completed. Multiple main process states will not be output for the same quality inspection task simultaneously.

[0209] Execution failure refers to situations where the document cannot be read, the structure parsing module malfunctions, the formula executor malfunctions, or the model output cannot be converted into the predetermined structure, and a valid result cannot be formed even after scheduled retries or alternative processing. Unacceptable blockage refers to situations where the quality inspection procedure has been executed, but the internal processing chain is incomplete, such as the inability to locate changed chapters, loss of field-document location mapping, missing evidence source records, failure to save rule dependencies, or the content block that should be processed not entering the quality inspection process.

[0210] Explainable blocking refers to a situation where the internal links of a system are complete, but a valid verification conclusion cannot be formed due to reasons such as the lack of external materials, the absence of necessary fields in the original materials, the lack of applicable rules, the requirement for manual verification by professional rules, or insufficient strength of evidence. Explainable blocking is not equivalent to a program error, nor is it mandatory to output an explanation. , , or .

[0211] Each blocking record includes at least the blocking code, the blocked object, the stage of occurrence, missing input, completed processing, affected fields or statements, the reason for not being able to continue execution, and the required supplementary materials or manual processing methods. Once the blocking reason is eliminated, the system resumes the corresponding quality inspection task based on the article identifier and the quality inspection object identifier, only re-executing the blocked object and its dependent objects, and not re-executing all completed quality inspections.

[0212] After manual review, the system generates an independent manual decision record, which at least saves the reviewer's identifier, review time, article and version identifier, quality inspection object identifier, selected candidate value, evidence identifier, rule version, decision conclusion, and decision reasoning. The manual decision record does not overwrite the original machine verification record; instead, it updates the current valid result pointer and retains the previous and current states. When manually entering field values, evidence, or rules, the system performs format, subject, period, unit, and source integrity checks on the entered content. Upon successful checks, a new audit version is generated, and the corresponding object is used as a change seed to re-execute impact propagation, ensuring that dependent fields, processing results, related declarations, and document-level quality conclusions are updated synchronously.

[0213] Quality inspection tasks save checkpoints based on the inspection object. Each checkpoint includes at least the task identifier, article and version identifier, logical_object_id, object_version_id, current processing stage, input object and its hash value, rule version and evidence version used, completed results, number of retries, and last update time. The system generates an idempotent execution key based on the article identifier, version identifier, object_version_id, and input hash value to avoid duplicate submissions and duplicate results. In the event of a temporary execution exception, the system retryes the current object from the most recent checkpoint; after reaching the preset maximum number of retries, it records `execution_failed`. When the blocking cause is eliminated or the task resumes, the system first verifies whether the document, rule, and evidence hash are still consistent with the checkpoint; if consistent, execution continues from the unfinished stage; if inconsistent, it locates the same logical object based on logical_object_id and re-includes the changed object_version_id and its downstream dependent objects in the quality inspection scope. Completed and unaffected objects are reused without a full re-inspection.

[0214] Weak evidence and the dependency chain status of processed fields are involved in the blocking determination. Weak evidence itself constitutes a questionable state; when a rule or conclusion must depend on this field and there is no other independent strong evidence, it constitutes an interpretable blocking. Partial missing dependencies of processed fields constitute either a questionable or interpretable blocking; when the dependent data already exists but cannot be obtained due to lost internal mapping, it constitutes an unacceptable blocking. This distinction avoids mixing objective missing materials, inapplicable rules, and internal processing anomalies into the same failure result.

[0215] Objects with the `not_applicable` keyword are not included in the set of valid document-level quality inspection objects. Objects with the following keywords are included in the set of pending review: `no_applicable_rule`, `insufficient_evidence`, `weak`, `context`, `conflict_pending_review`, `rule_conflict`, cross-layer verification conflicts, and `manual_review`. Internal link issues such as `rule_configuration_missing`, `rule_dependency_undefined`, `dependency_cycle`, missing field position mappings, and missing evidence source records are included in the set of unacceptable blocking issues. When the set of valid quality inspection objects is empty, `conforming` should not be output; instead, the document-level quality conclusion should be recorded as `pending_review` with an explanation of the reason for `no_evaluable_object`.

[0216] The above process status only indicates whether the quality inspection process has been fully executed; it does not guarantee that the content of the lengthy professional document will necessarily meet quality requirements. The set of quality inspection objects in the current version that have generated clear rebuttals or conclusions that the rules are not met is represented as... Its members include fields, statements, rules, and processing results that have been refuted or failed, or for which a clear rebuttal has been reached; the set of quality inspection objects for which a definitive quality conclusion cannot yet be reached is represented as Its members include objects in the context, weak, missing, cross-layer verification conflict, blocked, or manually reviewed states. Document-level quality conclusions are expressed as:

[0217] Here, Q(t) represents the document-level quality conclusion of the current version; nonconforming indicates that at least one item has been explicitly refuted by evidence or applicable rules; pending_review indicates that there is no explicit refutation, but there are still items with doubts, missing evidence, conflicts, blockages, or requiring manual review; conforming indicates that all items included in the quality inspection scope have formed valid support or the rules have been met, and the quality inspection process status is completed. In this section, completed specifically refers to Γ(t) = completed, which only indicates that this round of the process has been fully executed and is not equivalent to the object-level status. =completed is not equivalent to the rule being met (passed), nor does it alone indicate that the document quality is acceptable. The judgment conditions for the three types of document-level quality conclusions are mutually exclusive. The final result outputs both the document-level quality conclusion and the process status.

[0218] Where Ω(t) represents the joint output of the current version, Q(t) represents the content quality conclusion, and Γ(t) represents the process execution status. Even if Q(t) = nonconforming, process statuses such as failed, unacceptable_blocked, or explained_blocked must not be omitted; no component may overwrite another component.

[0219] For sets Each quality inspection issue object in This invention further proposes risk levels and structured revisions:

[0220] in, Indicates the object of quality inspection The risk level is rated as high, medium, or low. Indicates the target of quality inspection issues The resulting structured revision recommendations; Policy represents a pre-configured risk-revision strategy mapping function; Indicates the importance of the subject matter within a lengthy professional article; Indicate the question type and verification result type; Indicates the fields, processing results, declarations, conclusions, and chapter scope affected by the problem; Indicates the strength of evidence status corresponding to the object in question; This represents a multi-source chain of evidence associated with the object in question; This indicates the set of applicable rules. The risk level is determined to be high when a key field, key conclusion, or mandatory rule is refuted or fails, or when its impact propagates along the dependency chain to the key conclusion. The risk level is determined to be medium when the evidence is insufficient, the dependency chain is incomplete, or there are cross-layer conflicts that require further review and no clear rebuttal has yet been formed. The risk level is low when the impact of the problem is limited to non-key objects, there is strong evidence that can directly correct it, and it does not affect other processing results or conclusions.

[0221] Structured revision suggestions At a minimum, the revision should include the target chapter and location, the field or statement to be processed, the recommended action, the supporting evidence, the rules, and the scope of impact. For refuted or failed objects, the recommended action includes correcting the field or statement according to strong evidence and re-performing the verification of the affected processed fields and related conclusions. For weak or missing objects, the recommended action includes supplementing materials from specified sources or establishing locationable value-level evidence. For objects with incomplete dependency chains, the recommended action includes supplementing the necessary dependency fields and replaying the applicable rules. For cross-layer verification conflicts, the recommended action includes manually confirming the subject, period, unit, statistical scope, and source of the conflicting evidence. Revision recommendations must not generate unfounded replacement content that deviates from the located evidence, applicable rules, and dependencies.

[0222] The final differentiated quality inspection audit record should include at least the current version identifier, baseline version identifier, three types of signatures, signature comparison results, quality inspection route, set of changed fields, set of changed chapters, scope of partial reconstruction, reused objects, re-verified objects, evidence matching score, evidence strength status, weak evidence type, reason for evidence downgrade, processing field rules, dependent fields and their evidence status, rule replay results, declaration verification labels, rule satisfaction labels, cross-layer verification conflicts, process status, document-level quality conclusions, reasons for blocking, recovery conditions, risk level of problem objects, and structured revision suggestions.

[0223] Differentiated quality inspection audit records should also save logical_object_id and object_version_id, so that the same logical object can correspond across versions, and the content, location, evidence and processing results of each version instance can be traced independently.

[0224] Through the above mechanism, this invention organizes version change identification, quality inspection scope determination, declaration verification, multi-source rule verification, weak evidence downgrading, processing field dependency chain replay, process status control, document-level quality judgment, risk classification, and revision suggestions into a continuous quality inspection process. When there are no substantial changes to long professional documents, existing structured quality inspection foundations are reused; incremental reviews are performed when fields or content change partially; partial reconstruction is implemented when chapter structure changes; and the reasons for not reaching a conclusion are explained and recorded, thereby maintaining the continuity of field rules, evidence sources, declaration conclusions, and quality audit records throughout multiple rounds of revisions.

[0225] The invention will be further described in detail through a specific embodiment: I. Experimental subjects and statistical methods: To verify the quality control capabilities of this invention for multi-source fields, evidence status, source conflicts, processed fields, and conclusive content, the field rule table, field summary table, evidence classification table, conflict resolution table, processed field traceability table, conclusion verification table, and process audit records generated from actual project operation were analyzed. The field rule table contains 221 basic field source mappings and 86 processed field rules.

[0226] To avoid mixing data from different quality inspection scopes, as shown in Table 1, Sample A and Sample B were used to validate the multi-source aggregation and missing data detection of 66 target fields, respectively; Sample C represents the specific quality inspection results from Chapters 2 and 4 of another professional long paper, containing a total of 630 fields. The following statistics are calculated independently for each corresponding sample, and process coverage is not interpreted as model accuracy.

[0227] II. Multi-source field aggregation and source tracking: Samples A and B each generated 189 candidate values. The system normalizes and summarizes different subject values ​​for the same logical field, recording the final value, filling status, number of subject values, selected source, conflict classification, and source tracking; when no valid value is generated from any source, the field is marked as missing and not filled with speculative content. For example... Figure 5 As shown, an example of multi-source field aggregation and source tracing is displayed.

[0228] Table 1 Results of Multi-Source Field Processing

[0229] Experimental results show that both sets of records can distinguish between filled and missing fields and retain the source information corresponding to the field values. This result illustrates that the multi-source field processing is traceable, but does not indicate that all filled fields have strong supporting evidence.

[0230] III. Field Evidence Classification and Multi-Source Conflict Adjudication: Sample C examined a total of 630 fields, of which 457 were filled and 384 were effectively filled. Effective filling refers to fields that are not empty and do not fall under the categories of "insufficient information," "unable to determine," "not disclosed," or "not found," and this criterion is not replaced by the level of evidence. This chapter examines both the field filling status and the strength of evidence defined in Chapter 5 for categories A through D: filled fields are classified as Category A (direct evidence support) when strong evidence is obtained, as Category B (contextual evidence pending review) when contextual evidence is obtained, and as Category C (insufficient evidence) when weak evidence is obtained or when the field value exists but effective evidence cannot be located; unfilled fields are classified as Category D, and their evidence status is recorded as missing.

[0231] Table 2 Field Evidence Grading Results

[0232] Based on this calculation, the field fill rate is 457 / 630 = 72.54%, the effective fill rate is 384 / 630 = 60.95%, and the evidence locatability rate is (Class A 220 + Class B 235) / 630 = 72.22%. These three indicators respectively reflect whether a field value is formed, whether the field value excludes gap expressions, and whether value-level evidence or relevant context is located; they are not interchangeable. Of the 457 filled fields, 323 fields have two or more deduplicated source clues, resulting in a cross-validation coverage rate of 323 / 457 = 70.68%. Duplicate interface returns from the same publishing entity, the same upstream record, or with the same content hash are merged into one source and are not counted repeatedly due to different interface names. Figure 6 The record shown belongs to category D fields, which are only used to display the correspondence between fields with missing values, missing sources, judgment reasons, and processing suggestions.

[0233] Sample C identified 55 multi-source value difference fields. The system records the conflict type, candidate source and value, whether the model was used for judgment, the final adopted source, the final adopted value, the reason for judgment, and the information basis. Figure 7 The source display does not apply to two types of adjudication instances: those with substantially identical content from different sources but different forms of expression.

[0234] All 55 multi-source value difference fields resulted in conflict classifications, candidate sources and values, judgment reasons, and processing status, with a conflict processing record completion rate of 100%. Objects that could be adjudicated according to applicability, caliber, and source priority were assigned final adopted sources and values. Objects that could not eliminate semantic ambiguity or substantial conflict were retained as candidates and transferred to manual review. This percentage only indicates the completeness of conflict processing records and does not represent the accuracy rate of manual review of the adjudication results.

[0235] IV. Processing Field Dependency Chain and Conclusion Verification: Sample C contains 141 processing or judgment fields. The system reads the dependent fields according to the field rules, backtracks the values ​​and original information of the dependent fields, and re-executes the corresponding calculations or judgments. The verification results of the processing fields and conclusions are shown in Table 3 below.

[0236] Table 3 Verification Results of Processing Fields and Conclusions

[0237] Of these, 63 processing fields can be traced back to the basic fields and original information from the processing results, with a dependency chain traceability rate of 44.68%. When key dependencies are missing, dependency evidence is weak, or there are conflicts between dependency fields, the system will determine the result as questionable, lacking evidence, or refuted, and save the missing dependencies and the reasons for the judgment. All 182 concluding statements save the conclusion text, chapter position, verification conclusion, judgment reasons, and information basis.

[0238] V. Process Integrity and Explainable Blockage: The process audit results for Sample A and Sample B are shown in the table below. All engineering link checks passed, but the content quality status was explanatory blockage in both cases. The process audit results are shown in Table 4. Table 4

[0239] Sample A and Sample B had 16 and 15 executable statements blocked due to insufficient evidence, respectively, and neither had any unacceptable blocking issues. This distinguishes between internal processing chain defects and missing external materials: the former constitutes an unacceptable blocking, while the latter, when the processing chain is complete, is an interpretable blocking. Interpretable blocking does not indicate program failure, nor does it mean that the relevant content has passed quality verification.

[0240] VI. Verification Plan for Testing Accuracy: Existing operational logs can demonstrate the coverage and traceability of field, evidence, rule, conflict, and process status processing. However, records generated by the system itself cannot simultaneously serve as the gold standard for detection accuracy. Therefore, detection accuracy should not be replaced by field fill rate, evidence location rate, conflict handling record completion rate, or dependency chain traceability rate.

[0241] The accuracy evaluation of the test should be based on the gold standard of independent human assessment. Two professionals, without reviewing the system output, should independently establish a "complete set of objects to be inspected" from the original text, source materials, and applicable rules of Sample C. For each field, processing or judgment field, and conclusive statement, they should register the logical_object_id, original text location, expected status, evidence location, conflict type, and human conclusion; disagreements should be adjudicated by a third professional. Subsequently, the system-output objects and the complete set of human objects should be matched according to logical_object_id and location tolerance: objects not output by the system but confirmed by human assessment as requiring verification are counted as missed detections (FN), and invalid objects additionally output by the system are counted as FP. The matching objects should be re-evaluated for status, evidence location, conflict classification, final value or rejection judgment, processing result, and conclusion label. The 55 multi-source difference fields and 141 processing or judgment fields are used as a subset for specific evaluation, but they must not replace the evaluation of the complete set of objects.

[0242]

[0243] Accuracy, Precision, Recall, and F1 represent precision, accuracy, recall, and the harmonic mean of precision and recall, respectively. In object discovery evaluation, the number of successfully matched system objects with manually identified objects is denoted as TP_obj, the number of invalid objects output by the system is denoted as FP_obj, and the number of objects manually identified as eligible for detection but not output by the system is denoted as FN_obj. Based on these, object discovery precision, recall, and F1 are calculated. Status and conclusion evaluation is performed only on successfully matched objects. and These represent the system conclusion and the artificial gold standard conclusion for the i-th matched object, respectively, where i is the matched object number, N is the total number of matched objects, and unmatched objects are not included in the denominator of Accuracy; I(·) is an indicator function, taking 1 if the condition is true, and 0 otherwise. For binary classification, positive classes are frozen before evaluation: field consistency detection uses errors, inconsistencies, substantial conflicts, or cases that should be blocked as positive classes; conclusion verification uses rebuttals or failures as positive classes; correspondingly, TP, FP, and FN represent the number of correctly detected positive classes, the number of false positives (negative classes), and the number of undetected positive classes, respectively. Multi-class results are calculated and reported using a pairwise method for the remaining categories; high-risk errors are reported separately as the false negative rate FN / (TP+FN). The 95% confidence intervals for the two proportional indicators are calculated using the Wilson method, and the 95% confidence intervals for Macro-F1 are calculated using a bootstrap method stratified by object type and risk level. The number of resampling attempts B is set to 1000, and a random seed is saved. We statistically analyzed the indicators for object discovery, evidence location, conflict classification, final value selection, rejection decision, recalculation of processed fields, and conclusion verification, and used Cohen's Kappa to evaluate the consistency of human annotation.

[0244] To verify the technical effectiveness of the multi-source evidence chain, the verification process compares the complete method with configurations that do not perform weak evidence downgrading, conflict resolution, or dependency chain replay. The comparisons include field error detection rate, critical error missed detection rate, evidence location accuracy, and conclusion verification Macro-F1. Since the existing operational log in this embodiment does not yet include independent human gold standard results, this section is used to define the verification object, annotation rules, control group, and evaluation indicators. Existing process coverage statistics are not interpreted as conclusions about detection accuracy.

[0245] The artificial gold standard should be established after the system model version, threshold, weight, and runtime configuration are frozen, and the annotators should not view the system conclusions. The entire set of quality-checked objects should be generated independently from the original text and source materials, without using objects already identified by the system as the sampling frame. Cohen's Kappa accepts that the threshold be frozen before evaluation; in this implementation, 0.80 is used. If the threshold is lower than this, the annotation specifications should be revised and re-annotated, without adjusting the artificial labels based on the system output. The independent artificial gold standard accuracy verification and version difference review control experiment should be carried out according to Table 5.

[0246] Table 5. Independent Artificial Gold Standard and Version Differentiation Review Control Experiment Design

[0247] In the version differentiation verification experiment, the full verification results and the manually labeled set of affected objects are used as controls. The differentiation verification should cover all affected objects, and the verification conclusions on affected objects should be consistent with the full verification; at the same time, the reduction in the number of unaffected objects and changes in processing resources should be recorded. All statistical results should be associated with and saved with the sample version, manual labeling records, rule and evidence versions, runtime configuration, and random seed; before a complete controlled record is formed, the detection accuracy or the improvement in differentiation quality inspection efficiency should not be inferred based on this verification protocol.

[0248] VII. Data Analysis Conclusions: The above-mentioned actual operational data demonstrates that this invention can generate corresponding records for final field values, source records, evidence levels, conflict resolution, processing field dependencies, conclusion verification results, and process status, and perform downgrading or blocking processing when evidence is insufficient. Experimental results are used to illustrate the traceability, interpretability, and verifiability of the quality inspection process, and do not replace human professional judgment or final business approval. Existing archived data does not include controlled control records of different versions of the same professional long-form document; therefore, this chapter does not provide quantitative conclusions regarding the processing efficiency of differentiated quality inspections.

[0249] This invention is mainly used for content consistency quality inspection during the generation, delivery, review and auditing of professional long documents. Through multi-source cross-validation, evidence chain tracing and interpretable quality audit, it can discover inconsistencies, missing evidence, conflicts or unverifiable issues among factual statements, field data, numerical indicators, citation basis, rule judgments and conclusions in professional long documents, and determine whether the relevant key content is supported by evidence and whether it is consistent with the accessed evidence set.

[0250] This invention can be applied to quality acceptance scenarios after the generation of professional long-form documents. Existing intelligent writing or report generation systems typically require manual verification of data sources, citations, indicator definitions, and supporting conclusions after generating long documents. This invention can perform consistency quality checks on verifiable content such as fields, tables, indicators, statements, and conclusions in the document after the generation result is formed, outputting evidence sources, document locations, verification conclusions, risk levels, and revision suggestions, thereby reducing manual review costs and improving the quality of professional long-form document delivery.

[0251] This invention can be used to verify the consistency of facts and data in long professional documents. For example, it can perform multi-source cross-validation among the main text, tables, attachments, structured data, and external interfaces for the same entity name, the same period, the same amount, the same proportion, the same indicator, or the same fact, to identify problems such as field values ​​without source, inconsistencies between the main text and tables, contradictions in the same indicator, inconsistencies between cited content and the original text, and conflicts between external evidence and report statements.

[0252] This invention can be used in scenarios involving rule derivation and conclusion verification in professional long articles. For content formed through formula calculation, term conversion, business rule judgment, or professional reasoning, this invention is used to verify whether its supporting evidence is complete, whether the rule definitions are consistent, and whether the conclusions are supported by corresponding evidence and rules, thereby identifying problems such as calculation errors, inconsistent definitions, missing rule evidence, or insufficient conclusion support.

[0253] This invention can be used in review scenarios involving multiple rounds of revisions to professional long documents. For documents that have undergone manual modification, model continuation, partial rewriting, or version iteration, this invention is used to identify the scope of changes and perform targeted consistency quality checks on the changed content and its related evidence, fields, and conclusions, thereby reducing repetitive full-scale checks and maintaining the continuity and traceability of the quality inspection process.

[0254] This invention can be applied to professional long-form documents such as financial due diligence reports, investment research reports, credit review reports, regulatory submissions, technology reports, academic papers, experimental reports, legal opinions, compliance review reports, engineering design reports, consulting reports, and technical solution documents. It is especially suitable for business processes with high requirements for authenticity, consistency, traceability, verifiability, and audit documentation.

[0255] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0256] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A professional long-text consistency inspection method based on multi-source evidence chains, characterized in that, include: Step 1, Professional Long Text Analysis and Quality Inspection Scope Determination: Analyze the text content and table content in the professional long text to be inspected to form candidate evidence units with original text location identifiers; when a historical baseline version exists, obtain the structure signature, field signature, and content signature of the current version relative to the historical baseline version, determine the quality inspection route type based on the comparison results of the candidate evidence units and the three types of signatures, and determine the current quality inspection scope based on the quality inspection route type; Step 2, Explanatory Declaration Verification: Within the scope of the quality inspection, based on the declaration to be verified, a long context processing method or an evidence retrieval enhancement processing method is used to locate the textual and tabular evidence related to the declaration; based on the information extraction, numerical reasoning, or professional knowledge reasoning requirements involved in the declaration, the consistency relationship between the declaration and the evidence is judged, and an explanation of the reasoning process is generated while outputting the verification label; Step 3, Multi-source field and rule verification: Within the scope of the quality inspection, obtain multiple documents related to the verification target and a pre-configured professional rule base; based on document type, time range, and rule applicability, perform bidirectional constraint filtering on the documents and rules to form mutually applicable document rule pairs; Based on the rule type, the verification task is assigned to either the symbolic reasoning path or the arithmetic processing path; wherein, for the arithmetic processing path, formula generation and formula execution are separated, with the generative model generating the calculation formula and the deterministic external executor executing the calculation formula. Step 4, Verification Result Association and Audit Output: The verification results output in Step 2 and Step 3 are associated with each other through a unified object identifier. The impact of changes is propagated through a directed dependency graph, and re-verification is performed only on the affected objects. The final output is a locatable, interpretable, and verifiable quality audit result.

2. The method according to claim 1, characterized in that, In step 1, the three types of signatures include: Structural signatures are used to characterize the chapter structure features of a document and are generated from chapter number, chapter title, chapter level, and number of paragraphs. Field signatures, used to characterize field binding relationships, are generated from chapter position, field name, standard field identifier, subject role, normalized field value, statistical period, and unit of measurement. Content signatures, used to represent paragraph text content, are generated by concatenating the text hashes of each paragraph in document order; By comparing the three types of signatures of the current version with the baseline version, the quality inspection route is determined to be routine quality inspection, quality inspection for structural modifications, or quality inspection for structural changes.

3. The method according to claim 2, characterized in that, In step 1: The standard quality inspection reuses the baseline version of the structured quality inspection basis, but fields that were in a state of weak evidence hit, contextual evidence hit, doubt or lack of evidence in the previous round cannot be directly reused for the final conclusion; The same structure modification quality inspection identifies the set of changed fields based on the logical object identifier, and includes the downstream dependent objects of the changed fields in the quality inspection scope; The structural change quality inspection identifies the set of changed chapters and re-extracts fields, evidence, and statements from the changed chapters.

4. The method according to claim 1, characterized in that, In step 2, the declaration verification is divided into information extraction verification, numerical reasoning verification and professional knowledge intensive verification. The information extraction type verification is used to process statements for which verification evidence can be directly obtained from the text or tables of professional documents; The numerical reasoning verification is used to process statements that require calculation or statistical analysis based on data in the professional text before a judgment can be made; The aforementioned expertise-intensive verification is used to process statements that require interpretation of factual information in a document by combining professional concepts, industry knowledge, or regulatory rules.

5. The method according to claim 1, characterized in that, In step 2, the verification labels of the verification statement include entailed and refuted. Entailed indicates that all the decisive content of the statement is supported by the verification context; refuted indicates that the statement contradicts the verification context or that some of the decisive content in the statement contradicts the verification context. The candidate label is accepted as the final verification label only when the conditional probability of the candidate label meets the minimum acceptance threshold and the minimum probability difference, and the evidence integrity check is passed; otherwise, the entailed or refuted label is not generated, and the process is moved to expanded search, switching verification methods, or manual review.

6. The method according to claim 1, characterized in that, In step 3, the bidirectional constraint filtering includes: Documents are filtered by time based on the time constraints of the verification objectives; Recalling rules based on semantic relevance; Based on the matching relationship between document identification type and the document type to which the rule applies, documents and rules are mutually constrained, excluding inapplicable documents and irrelevant rules.

7. The method according to claim 1, characterized in that, In step 3, the process of separating formula generation from formula execution includes: The formula generation model generates the calculation formula; The calculation formula is verified by an independent formula verification model. The verification includes whether the fields involved in the calculation are all from the located evidence, whether the subject and period of the fields are consistent, and whether the operation relationship conforms to the rules. After the verification is passed, the calculation formula is executed by a deterministic external actuator to obtain the calculation result; Based on the comparison between the calculation result and the document recorded value or rule threshold, the rule is output to satisfy the tag Pass or Fail.

8. The method according to claim 1, characterized in that, In step 3, the matching strength between the field and the candidate evidence is determined, and the evidence strength status is divided into four levels: strong, context, weak, and missing. Automatic conflict resolution is performed only when strong-level candidate evidence exists; when there is no strong-level candidate evidence, no value is automatically selected from lower-level evidence, and the conflict is recorded as pending review. Weak evidence includes situations where the field value only matches the interface status field, the field value is equal to the subject name but the field to be verified is not a subject name field, the evidence only contains the field name and not the field value, and the subject or period of the evidence is inconsistent with the field.

9. The method according to claim 1, characterized in that, In step 4, the nodes in the relationship graph include basic fields, processing fields, declarations, conclusions, rules, and evidence objects, and the directed edges include field input dependencies, rule bindings, declaration references, evidence support, and conclusion references. When a field value changes, paragraph content changes, evidence becomes invalid, rule version is replaced, or manual ruling is rewritten, the corresponding node is registered as a change seed, and the impact is propagated from upstream to downstream along the directed edge, so that unaffected objects do not enter the current round of re-inspection.

10. The method according to claim 1, characterized in that, In step 4, the quality audit results include: document-level quality conclusions and process status; The document-level quality conclusions include: "Conforming" indicates that all objects included in the quality inspection scope have formed valid support or rule-compliant conclusions, and the process status is "completed". pending_review indicates that there are no explicit objections, but there are still issues such as doubts, missing evidence, conflicts, blockages, or manual review. nonconforming indicates that at least one of the contents has been clearly refuted by evidence or applicable rules; The process status includes four categories, determined in the order of execution failure, unacceptable blockage, explainable blockage, and execution completion.

Images