A strong supervision report embedded explainable AI quality control method and system

Abstract

The application discloses a kind of strong supervision report embedded explainable AI quality control method and system, it is related to explainable artificial intelligence and strong supervision scientific document compliance quality control field.The application receives strong supervision report generation or auditing task, parses multi-source heterogeneous data to construct standardized evidence package;Load matching certainty workflow, execute data analysis, content generation, quality control verification, manual review or audit record according to node responsibility at key business node;The content obtained by generation or auditing is disassembled into multi-level quality control objects, with the smallest quality control unit as a declaration unit and classification is completed;Rely on declaration type driving plug-in QC routing, embedded quality control verification is carried out in workflow full node;According to QC result intelligent adjudication, realize controllable automatic correction and high-risk content manual review trigger;Build the structured association evidence chain of declaration and evidence, generate three-level explainable explanation based on real quality control process constraint;Generation stage audit log, output report after quality control.

Description

Technical Field

[0001] This invention relates to the fields of interpretable artificial intelligence, natural language processing, generation of highly regulated scientific documents, quality control, evidence tracking, audit trails, and compliance information systems. In particular, it relates to a method and system for embedded interpretable AI quality control in highly regulated industries such as life sciences, pharmaceutical R&D, medical devices, public health, chemical safety evaluation, food safety evaluation, and environmental safety evaluation. Specifically, it relates to a method and system for embedded interpretable AI quality control in highly regulated reports. Background Technology

[0002] With the rapid iteration of technologies such as Large Language Modeling (LLM), Multi-Agent Systems, Retrieval Enhancement Generation (RAG), Automated Document Arrangement, and Automated Workflow, artificial intelligence has been widely applied to business scenarios such as scientific document writing assistance, intelligent organization of R&D materials, automatic summarization of professional data, batch generation of compliance reports, and document compliance review.

[0003] In highly regulated fields such as life sciences, pharmaceutical R&D, and medical devices, scientific reports possess high professional barriers and regulatory constraints. They are not ordinary text documents but rather complex knowledge products supported by a variety of heterogeneous sources, including research protocols, raw experimental data, statistical analysis reports, laboratory records, standard operating procedures (SOPs), industry regulations and guidelines, historical compliance report templates, expert review opinions, and system audit logs. The entire process of non-clinical R&D, clinical trials, and medical device registration typically involves multiple sources of data, including protocols, raw records, Excel spreadsheets, Word documents, PDFs, CSVs, historical templates, laboratory SOPs, and external regulations and audit requirements. The difficulty in compiling scientific reports lies not merely in writing and formatting, but in achieving consistency between statistical analysis results and textual descriptions, ensuring logical coherence across report chapters, tracing key scientific conclusions back to the original underlying data, implementing a closed-loop compliance review process through multiple rounds, and ensuring complete archiving of audit logs.

[0004] Existing AI-powered document generation technology can automatically generate semantically coherent and formatted professional text based on user instructions and knowledge base search results. Traditional document review tools can only perform basic post-processing checks after the document is finalized, such as formatting correction, typo checking, standardization of professional terminology, surface-level comparison of basic numerical data, and annotation of literature citations. However, in the context of stringent compliance scenarios for scientific reports under strong regulatory oversight, the simple post-processing check model is no longer sufficient to meet the multiple stringent requirements of regulatory agency review, corporate QA (quality assurance) audits, customer compliance verification, and internal quality system control.

[0005] The primary compliance requirement for heavily regulated scientific documentation is not simply producing fluent, formatted professional text, but rather a rigorous requirement that every key scientific conclusion in the report be traceable to original underlying evidence, every qualitative judgment be supported by regulations and experiments, and every formally delivered report successfully pass internal QA review, third-party compliance audits, and on-site verification and document review by drug regulatory agencies. Driven by both industry needs and regulatory trends, there is an urgent need to develop a fundamental technical solution that deeply embeds quality control mechanisms into the entire report generation process, simultaneously generating a structured and interpretable chain of evidence, a complete audit trail, and a record of manual review.

[0006] Large language models possess fundamental capabilities such as contextual semantic understanding, fluent natural language generation, intelligent document summarization, text rewriting and reconstruction, professional knowledge summarization, and intelligent question-answering interaction. However, their essence is a probabilistic generation architecture, naturally suited to general text scenarios. In the context of highly regulated professional reports, they have unavoidable inherent defects: the generated content appears professional and rigorous on the surface, but the underlying evidence is hidden and untraceable; they are prone to misinterpretation and distortion of original experimental data, statistical report parameters, and research plan clauses; they habitually express speculative and uncertain professional judgments as definitive conclusions; they are prone to AI illusions and fabrication problems regarding key business values, units of measurement, observation time nodes, research group divisions, and statistical significance results; the model generation process is a black box mechanism, unable to naturally provide an auditable, replayable, and traceable conclusion generation chain, which does not meet the basic requirements of compliance and traceability in highly regulated industries.

[0007] Retrieval-enhanced generation (RAG) obtains relevant references by pre-searching industry knowledge bases, project document libraries, and regulatory databases, and uses the search results as contextual constraints for the large model generation logic. While this can alleviate the model illusion problem to some extent, the traditional RAG technical architecture suffers from the following fundamental shortcomings: There is a lack of structured binding relationships between the retrieved documents and the AI-generated text; it can only generally indicate which documents the report content referenced, failing to accurately prove which specific original evidence supports a key professional conclusion; it lacks the ability to definitively and accurately verify core business values, units of measurement, statistical calculation logic, and consistency of cross-chapter statements in the report; and it focuses solely on text generation optimization, failing to build a quality control record of the entire report generation process, making it difficult to support compliance audits and accountability.

[0008] Traditional, heavily regulated document quality control models rely heavily on manual, word-by-word review by senior industry experts, fixed keyword matching, and mechanical format comparison of old and new documents. This approach has several practical shortcomings: quality control intervention is delayed, with reviews only conducted when the report is nearing completion or finished; logical errors, data discrepancies, and phrasing violations discovered in the early stages cannot be corrected promptly, leading to costly and lengthy rectification efforts in the final draft phase; review quality depends on expert experience and professional competence, resulting in subjective, inconsistent, and difficult-to-scale replication and promotion of review standards; the manual review process lacks a digital, structured data accumulation mechanism, preventing the solidification and reuse of review logic, judgment criteria, and rectification basis; and it lacks the technical capabilities for automated playback of the report generation process and evidence-level professional interpretation, resulting in low efficiency in compliance traceability.

[0009] Existing research and development of explainable artificial intelligence (XAI) technology focuses on explaining the predictive logic within the model. It primarily elucidates why the model outputs a particular classification result, scoring conclusion, or text content. However, it cannot fundamentally prove the scientific validity, rigor, and compliance of the derivation process of scientific conclusions through a complete, structured chain of evidence. This approach is only suitable for general AI algorithm transparency scenarios. In the highly regulated professional scenarios of scientific reports, the industry requires more than just a simple explanation of model output logic. It demands a deep explanation system at the evidence, process, and responsibility levels across the entire compliance process, comprehensively addressing five key questions: From which original underlying evidence do the key scientific conclusions of the report originate? What QC control nodes and verification processes are involved in the entire report generation process? Which content in the report is automatically confirmed through data rules, formula calculations, and document comparison? Which high-risk professional content requires manual expert review? Can the final delivered report fully reconstruct the entire generation process and form a closed-loop audit evidence chain? Existing XAI technology can only generate general summary text descriptions and can only record simple operation behavior logs afterward. It cannot achieve deep structured binding of report declaration units, structured evidence, process quality control results, manual review status, and audit logs, and cannot build a native compliance quality control defense line.

[0010] In recent years, drug regulatory agencies around the world have attached great importance to the application and control of artificial intelligence technology in the entire chain of drug research and development, registration and application, review and approval, post-marketing pharmacovigilance, and production quality management. AI is no longer just regarded as a general tool to improve office efficiency, but has gradually intervened in the key information generation process that directly affects drug safety, efficacy, quality compliance and regulatory approval decisions.

[0011] Global regulatory focus has now been significantly upgraded, extending beyond simply reviewing the final report output by AI to rigorously examine the following seven dimensions:

[0012] Whether the definition of AI application business scenarios is clear and credible;

[0013] Whether the original data source on which the generated content is based is authentic and traceable;

[0014] Whether the output results of the AI ​​model have undergone professional verification and quality control checks;

[0015] Does the entire report generation process have a complete traceability chain?

[0016] Are the responsibilities clearly defined and are the records complete for manual review?

[0017] Does the AI ​​system possess full lifecycle version management and security protection capabilities?

[0018] Does the system support complete playback of the audit process and export of the chain of evidence?

[0019] The U.S. FDA has released draft guidance on AI-supported regulatory decisions for drugs and biologics, proposing a framework for assessing the credibility of AI models based on use cases and risks. China's National Medical Products Administration (NMPA) has also released implementation opinions on "Artificial Intelligence + Drug Regulation," proposing to promote the application of AI in drug regulatory scenarios and emphasizing fundamental capabilities such as high-quality datasets, large-scale vertical models, intelligent agents, algorithm transparency, model validation, and safety protection. Under this regulatory trend, the traditional model of simply generating professional text using large language models or conducting only manual post-review after the report is finalized is no longer sufficient to meet the stringent regulatory requirements of future regulations regarding the credibility of AI results, traceability of evidence, process transparency, audit replayability, and the delineation of responsibility.

[0020] In summary, existing AI-powered report generation, document quality control, and explainability technologies suffer from the following technical deficiencies, hindering the intelligent compliance transformation of heavily regulated industries:

[0021] First, existing report generation technologies only focus on the final text output effect and lack an embedded real-time quality control mechanism for key nodes in the entire report generation process, resulting in serious lag in risk management.

[0022] Second, traditional QC quality control is a post-review centralized review model, which cannot identify risks and correct deviations in real time at key business nodes such as raw data intake, standardized construction of multi-source evidence, generation of report chapter drafts, cross-chapter logical integration, and final draft assembly, resulting in a lag in problem discovery.

[0023] Third, existing RAG and literature citation technologies can only achieve coarse-grained citation annotations at the document and paragraph levels, and cannot accurately and structurally bind the core business fields, key values, units of measurement, statistical table content, independent semantic declarations, paragraph and chapter content of the report to the original evidence source, resulting in insufficient source tracing accuracy.

[0024] Fourth, the existing AI generation process is a black box closed loop, lacking a structured playback mechanism for the entire process. It is impossible to accurately trace which input data, which original evidence, what model configuration, which verification rule, and which manual review step were used to form a certain professional conclusion, leaving no basis for audit tracing.

[0025] Fifth, the existing system lacks the ability to classify and control the quality of statements in a refined manner. It cannot distinguish between scientific statement types such as factual, calculation, process, comparison, interpretation and judgment, and compliance. The evidence requirements, verification rules, and risk levels of different statements vary significantly. The adoption of a uniform quality control strategy leads to the omission of high-risk hidden dangers.

[0026] Sixth, existing technologies treat manual review as an independent external process and do not embed manual review nodes into AI-generated workflows, resulting in blurred boundaries of responsibility for human-machine collaboration, inability to synchronize review status in real time, and difficulty in incorporating review opinions into closed-loop management of the audit chain.

[0027] Seventh, most of the existing AI-explainable texts are summative descriptions generated freely by the model. The explanations are arbitrary and lack constraints, making it difficult to ensure that the explanations are consistent with the actual QC verification process, the source of original evidence, the content correction actions, and the conclusions of human review. As a result, the explanations lack credibility and cannot be used for compliance verification. Summary of the Invention

[0028] Therefore, the purpose of this invention is to propose a strongly regulatory reporting embedded interpretable AI quality control method and system. It constructs a workflow-first, agent-in-node, skill-as-asset, and QC-by-design underlying technical architecture to control all business nodes in the process with deterministic workflow, enhance the business processing capabilities of individual nodes with intelligent agents, and implement plug-in quality control throughout the entire report generation chain. It deeply embeds quality control into the entire report generation process and can simultaneously and automatically generate structured evidence chains, full-process quality control records, manual review logs, standardized interpretable explanations, and audit data packages. This addresses the challenges of existing strongly regulatory science methods. AI-generated reports cannot accurately trace back to the original underlying evidence. Traditional QC intervention is delayed, problem discovery is slow, and the final draft is costly. Key values, units, statistical results, limiting conditions, and professional conclusions in the reports are inconsistent with the original source data. The entire report generation process lacks a structured audit trail and cannot be replayed for tracing. AI-generated explanations are severely disconnected from the actual quality control verification process and lack credibility. Different types of scientific statements use homogeneous quality control strategies, making it easy to overlook high-risk issues. The manual review process is disconnected from the AI ​​generation link, the boundaries of responsibility are blurred, and review records cannot be included in the audit loop. These issues need to be addressed to improve the professional credibility and compliance of AI-generated reports with strong regulatory oversight.

[0029] This invention provides a strongly regulated reporting embedded interpretable AI quality control method, constructing a closed-loop chain that includes deterministic workflow, in-node agent, rule, and tool execution, in-node QC gating, node control, and audit package output, comprising the following steps:

[0030] Step S1, Task Reception: Integrate with the task of generating regulatory scientific reports or the compliance review task of existing reports, and collect task-related parameters, including report type, project unique number, standardized document template, scope of chapters to be generated, list of related materials, operator permission level, quality control strictness level, whether manual review is mandatory, and whether audit package export configuration parameters are supported; among them, regulatory scientific reports cover all categories of compliance documents, including non-clinical research reports, clinical research reports (CSR), new drug IND (Investigational New Drug Application) / NDA (New Drug Marketing Application) / BLA (Biologics License Application) registration application materials, pharmacovigilance reports, medical monitoring reports, real-world research reports, quality system documents, etc.

[0031] Specifically, a strongly regulated scientific report refers to a scientific document that needs to meet the requirements of regulations, standards, standard operating procedures (SOPs), audits, and quality systems in regulated fields such as life sciences, pharmaceutical research and development, medical devices, chemicals, food safety, and environmental safety.

[0032] Step S2, Multi-Source Evidence Analysis and Standardized Evidence Package Construction: Collect multi-source heterogeneous raw data associated with the reporting task, including research protocols, statistical analysis plans, raw experimental data, statistical reports, raw experimental records, case report forms, medical monitoring records, safety databases, pharmacovigilance case reports, standard operating procedures (SOPs), industry regulatory guidelines, historical compliance reports, client-customized templates, and human expert review comments; perform format parsing, normalization, and structure transformation on the multi-source heterogeneous data to construct a unified standardized evidence package, converting each raw data into a unique, locatable, summarizable, and citationable evidence_ref standardized evidence object; each evidence object is pre-defined with a unique evidence identifier, evidence source type, evidence source name, standardized location path, original evidence content, original parameter value, standardized value, unit of measurement, file version information, SHA digest checksum, and access control fields.

[0033] Specifically, evidence_ref refers to a reference to an evidence object, used to uniquely identify a source of evidence and its specific location, such as a table cell, a section of text in a plan, a SOP, a raw data record, or a manual review opinion.

[0034] Step S3, Deterministic Workflow Orchestration and Loading: Based on the specific report type and business compliance scenario, the matching deterministic business workflow is automatically loaded. The workflow is implemented using a deterministic state machine, a dedicated workflow DSL description language, or a directed acyclic graph architecture. Different report types are configured with dedicated customized workflow nodes. The CSR workflow for clinical research reports includes nodes for structuring the research protocol, reading the statistical analysis plan, parsing subject distribution data, parsing efficacy / safety data, generating chapter drafts, data consistency quality control, medical logic quality control, cross-chapter consistency quality control, manual medical review, final report assembly, and exporting the audit package. The IND / NDA registration application workflow includes nodes for module structure identification, loading application document templates, mapping non-clinical / clinical / CMC data, verifying regulatory compliance, ensuring consistency of citations, scanning for missing compliance items, classifying risk levels, manual review by registration experts, and generating the submission package.

[0035] Step S4, Embedded Content Generation in Workflow Nodes: In each business node of the deterministic workflow, the large language model, industry vertical domain model, professional expert agent, rule engine, template rendering engine, code sandbox tool, professional data calculation tool, and evidence intelligent retrieval tool are invoked to generate corresponding business content; the content generation output results are forcibly bound to the input context information, evidence package version number, model configuration parameters, prompt word summary, output content summary, and the unique number of the current workflow node, to achieve full-link traceability and anchoring of the generated content.

[0036] Step S5: Multi-level Quality Control Object Decomposition: The report content generated by the workflow nodes is decomposed into nine hierarchical quality control objects from top to bottom, namely: business field, core value, unit of measurement, table, sentence, business claim, paragraph, section, and complete report. Among them, the claim is the smallest unit of business semantic quality control. The character span anchor point is only used for text positioning, highlighting, reference marking and audit playback coordinates, and is not used as a business quality control object. The claim is divided into six standard types: fact, calculation, process, comparison, interpretation and judgment, and compliance.

[0037] Specifically, a claim refers to the smallest independently verifiable unit of business semantics in a report. Claims can be factual, calculation-based, process-based, comparative, explanatory, or compliance-based.

[0038] A character span anchor (SpanAnchor) is a text range anchor in the text used to locate a claim or key content. It can include character positions, paragraph positions, sentence positions, DOM paths, or token positions.

[0039] Step S6: Declaration Type-Driven Plugin-Based Embedded QC Execution: Embed QC sub-processes within each business node of the workflow, configuring nine types of quality control plugins: format quality control, data quality control, calculation quality control, source quality control, logic quality control, consistency quality control, compliance quality control, expression quality control, and audit integrity quality control. Based on the semantic type of the declaration unit, a dedicated QC plugin combination is matched, with different declaration types triggering differentiated quality control verification logic. Factual declarations match data quality control and source quality control plugins; calculation declarations match calculation quality control and data quality control plugins; process declarations match source quality control and compliance quality control plugins; comparison declarations match data quality control, calculation quality control, and logic quality control plugins; explanation and judgment declarations match source quality control, logic quality control, and expression quality control plugins; compliance declarations match compliance quality control and audit integrity quality control plugins. QC verification results are written to the current business node status and used as gating inputs for subsequent node release, automatic correction, manual review task creation, or generation blocking.

[0040] Step S7, Intelligent Adjudication and Content Correction of QC Results: Based on the verification results of various QC plugins, the type of declaration unit, the risk level, and the certainty of the original evidence, the system automatically executes multi-dimensional adjudication. The adjudication results are divided into six categories: matched (approved), corrected (automatically corrected), unresolved (unresolved), human-required (mandatory manual review), blocked (generated), and not applicable (not applicable). The conditions for automatic correction are strictly limited: the source of evidence is clear and traceable, the verification and inspection method is certain, the risk level is within the range that can be automatically corrected, the differences between the content before and after correction can be fully recorded, the reason for correction can be explained in a standardized way, and the correction action supports full process playback. For explanatory judgment, compliance, and high-risk qualitative declaration units, the system only outputs professional suggestions and cannot replace the final judgment made by human experts.

[0041] Step S8: Generation of Structured Evidence Chains: Establish multi-dimensional structured relationships between report text content, various types of declaration units, QC quality control verification results, and standardized evidence objects; pre-set seven types of relationships: supports, contradictions, partially supported, background, requires manual review, derived from data, and corrected by evidence, generating structured evidence chain data with confidence scores and standardized judgment reasons.

[0042] Step S9, Process-Constrained XAI Interpretable Generation: Based on real QC verification records, declaration unit traceability, standardized evidence citations, content correction logs, unresolved conflict markers, manual review status, and workflow stage audit logs, constrain the generation of interpretable explanatory text; prohibit the model from freely fabricating explanatory content, and strictly follow the generation rules: content without supporting evidence shall not be included in the explanatory text, cited data sources must be marked with a unique evidence identifier, automatically corrected content must clearly state the text before and after correction and the reason for correction, unresolved conflicts must be explicitly marked, and manually reviewed content must indicate the reason for review and the responsible node; output three levels of interpretable explanatory text: declaration level, chapter level, and report level.

[0043] Step S10: End-to-end staged audit tracing stored in the database: AuditStageLog stage audit logs are automatically generated at each business stage of the workflow. Each log entry includes a unique runtime identifier, project number, document number, stage name, stage sequence number, executing agent name, underlying model name, prompt word digest hash, input content digest hash, output content digest hash, evidence package digest hash, configuration parameter digest hash, execution status, start and end time, hash value of the previous stage, and complete hash value of the current stage. Audit logs, evidence chain data, quality control results, correction records, and manual review records are uniformly stored in the audit database, supporting multi-dimensional process playback and audit package export by report, chapter, declaration unit, evidence source, manual review node, and correction difference.

[0044] Specifically, the chain of evidence refers to the structured relationship between the report content, the claim, the object of the evidence, the QC results, the correction record, the manual review, and the audit log.

[0045] Step S11, Compliance Output: Output the final report text, structured QC quality control result report, three-level interpretable explanatory documents, and standardized audit data package, which have undergone full-process embedded quality control verification, to complete the compliance delivery of the strongly regulated report.

[0046] This invention also provides an embedded interpretable AI quality control system for strongly regulated reports, used to implement the aforementioned embedded interpretable AI quality control method for strongly regulated reports. The overall system architecture is divided into four layers: input and evidence layer, workflow and generation layer, embedded quality control and adjudication layer, and interpretation, auditing and output layer, specifically including the following functional modules:

[0047] The task receiving module is used to receive tasks for generating scientific reports under strict supervision and reviewing existing reports. It collects and parses task parameters such as report type, project number, document template, chapter scope, data list, user permissions, quality control level, manual review configuration, and audit export configuration.

[0048] The multi-source evidence parsing module is used to parse multi-source heterogeneous original data associated with the task, complete format adaptation, content extraction, normalization processing, and convert it into a unified standardized evidence_ref object to build a standardized evidence package that can be located, referenced, and traced.

[0049] The workflow orchestration module is used to load deterministic workflows based on report type and business scenario. It supports custom workflow template configuration, node addition and deletion, and business process adaptation, and uses state machines, DSLs, or directed acyclic graphs to achieve process control.

[0050] The content generation module is deployed on each workflow node and is used to call large language models, domain vertical models, expert agents, rule engines, template engines, code sandboxes, data computing tools, and evidence retrieval tools to generate professional content, and to forcibly bind input, evidence, models, prompt words, and node traceability information.

[0051] The quality control object decomposition module is used to break down the generated report content into nine levels of quality control objects from top to bottom: fields, values, units, tables, sentences, declaration units, paragraphs, chapters, and reports. It also completes automatic classification and text positioning anchor point matching for the six major categories of declaration units.

[0052] The embedded QC execution module has nine built-in pluggable quality control plugins, which are used to perform embedded quality control verification in each node of the workflow. It routes and matches exclusive QC plugin combinations according to the declaration unit type and outputs standardized quality control verification results.

[0053] The adjudication and correction module is used to execute intelligent adjudication based on QC verification results, declaration type, risk level, and evidentiary certainty. It enables automatic approval, controllable automatic correction, unresolved marking, manual review triggering, generation of blocking differentiated processing, and complete recording of the content before and after correction, as well as the basis and reasons for correction.

[0054] The evidence chain generation module is used to establish a structured relationship between report content, declaration units, QC results, and standardized evidence objects, define seven types of logical relationships, and generate structured evidence chain data with confidence levels and judgment reasons.

[0055] The XAI explanation generation module is used to generate three levels of interpretable explanations—declaration level, chapter level, and report level—based on real QC records, evidence chains, correction logs, manual review status, and audit logs, ensuring that the explanation content is consistent with the actual quality control process.

[0056] The review and audit module is used to generate stage audit logs at each business stage of the workflow. Specifically, it embeds high-risk, evidence conflict, interpretation and judgment, and compliance declaration units into the workflow to realize automatic distribution of manual review tasks, entry of review opinions, synchronization of review status, and archiving of review records, forming a closed loop of human-machine collaborative responsibility.

[0057] The audit trail module is used to generate stage audit logs with hash digests at each stage of the workflow. It retains inputs, outputs, evidence, model configurations, quality control results, correction records, and manual review records throughout the entire process. It supports multi-dimensional process playback, log tracing, and audit package export.

[0058] The configuration management module is used to configure report types, evidence types, workflow templates, declaration types, QC plugins, audit rules, and quality control levels, and supports general reuse across report types.

[0059] Preferably, the present invention constructs a configurable base suitable for different application scenarios, as detailed below:

[0060] Report type configuration: compatible with clinical reports, statistical reports, pharmacology reports, and compliance reports;

[0061] Evidence type configuration: adapt to tables, raw data, schemes, SOPs, logs, etc.;

[0062] Workflow template: Automatic assignment of review, automatic transfer, and automatic release;

[0063] Declaration type configuration: Fact, Calculation, Comparison, Explanation, and Compliance Declarations;

[0064] QC plugin configuration: numerical verification, out-of-bounds check, consistency comparison.

[0065] Based on configurable workflows, declarative types, and QC plugin architecture, it can quickly adapt to various scenarios of generating and controlling scientific reports under strong supervision without refactoring the underlying technology. It has strong cross-report type reuse and expansion capabilities, reducing the cost of intelligent transformation for enterprises in multiple scenarios.

[0066] The report output module is used to integrate the final quality control report text, structured QC results, interpretable documentation, and audit data packages to complete the standardized output and archiving of compliance results.

[0067] This invention adopts a strongly regulated, end-to-end embedded QC architecture for reporting (quality control is not performed separately after the report is completed, but is embedded in multiple stages such as data analysis, evidence construction, content generation, chapter integration, manual review, and final delivery). It abandons the traditional post-finalization centralized inspection model, deeply embedding the quality control mechanism into all business nodes including data acquisition, evidence standardization, chapter generation, data verification, logic validation, cross-chapter consistency verification, manual review, final draft assembly, and audit export. This enables real-time risk identification and immediate deviation correction, significantly reducing final draft costs. It constructs a nine-level hierarchical quality control object decomposition system for report content, decomposing it from top to bottom into multiple levels of objects: fields, values, units, tables, sentences, declaration units, paragraphs, chapters, and the report itself. This invention employs a customized quality control and verification strategy, enabling refined, layered management. It establishes a business semantic quality control mechanism centered on the claim unit, abandoning the traditional model of using tokens and ordinary sentences as quality control units. Instead, it uses the smallest independently verifiable business semantic claim as the quality control carrier, with text span anchors serving only as location and playback anchors and not participating in business quality control, aligning with the professional business logic of strong regulatory oversight. Furthermore, it features a unique claim type-driven intelligent routing mechanism for QC plugins, classifying six standard claim types and automatically matching each type with a customized QC plugin combination. This achieves differentiated and accurate quality control, avoiding the omission of high-risk vulnerabilities due to homogeneous verification. Finally, it standardizes and objectifies multi-source heterogeneous data evidence packages, transforming disorganized and structurally diverse original data... All materials are uniformly converted into uniquely identifiable, accurately located, summary-verified, and compliantly citeable evidence objects, providing standardized underlying data support for claim tracing and quality control verification. A dual-track adjudication mechanism driven by QC results—automatic correction and manual review—is established to automatically correct, manually review, block risks, or adjudicate QC results. This improves the efficiency of automated quality control while strictly adhering to the bottom line of strong regulatory responsibility, achieving a reasonable division of labor in human-machine collaboration. An innovative design uses XAI interpretable explanations generated from constraints of real quality control processes (QC records, evidence chains, correction records, and manual review status). The explanatory text is no longer freely fabricated by the model but is strictly based on constraints from real QC verification records, evidence chain associations, correction logs, and manual review status. To ensure the accuracy, traceability, and usability of the explanations, a stage-based audit log and chained hash digest mechanism are designed. Each workflow node generates independent input, output, evidence, configuration, and stage hash digests, forming a chained, tamper-proof audit trail that supports accurate playback of reports, chapters, statements, and evidence across multiple dimensions throughout the entire process. A human-machine collaborative responsibility closed-loop architecture is constructed, deeply embedding human review nodes into the AI-generated workflow rather than as an independent external process. This enables automatic distribution of review tasks, opinion tracking, status synchronization, and responsibility definition. Review records are fully incorporated into the audit evidence chain, clarifying the boundaries of human and machine responsibility and strictly adhering to the regulatory principle that AI is executable while humans bear ultimate responsibility. A universal, strongly regulatory document quality control platform (such as...) is created across report types. Figure 6As shown, through configurable report types, workflow templates, evidence types, statement types, QC plugins, and audit rules, it can flexibly adapt to all categories of highly regulated documents, such as CSR (Clinical Research Report), IND (New Drug Clinical Trial Application) / NDA (New Drug Marketing Application), pharmacovigilance, medical monitoring, and non-clinical research, and has strong versatility, scalability, and business reuse capabilities.

[0068] The existing approach involves generating a report and then performing an external check, while this invention embeds QC into every node of data acquisition, evidence construction, and chapter generation, transforming post-review into a fully embedded process of QC (QC-by-design). This is an innovation in underlying paradigms.

[0069] This invention constructs a system for AI-generated scientific reports, enabling evidence tracing, quality control documentation, and regulatory compliance in scenarios with strong regulatory oversight. It uses a Claim as the smallest semantic unit, rather than sentences, paragraphs, or documents; embeds QC gating within the generation node, rather than post-processing checks; and rigidly constrains the interpretation content by QC, evidence chains, and correction logs, rather than allowing for free generation. Stage-based hashing and chain-based auditing enable replayability, rather than relying on ordinary logs. Existing solutions cannot simultaneously satisfy the complete chain of evidence traceability, numerical verification, process auditability, and closed-loop accountability.

[0070] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the strongly regulated reporting embedded interpretable AI quality control method as described above.

[0071] The present invention also provides a computer device, the computer device including a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, it implements the steps of the strongly supervised report embedded interpretable AI quality control method as described above.

[0072] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0073] The embedded interpretable AI quality control method and system for strong regulatory reports provided by this invention realizes the transformation and upgrading from post-process quality control to full-process embedded process quality control. It conducts real-time quality control verification and deviation correction at key business nodes in report generation, proactively identifying risks such as data inconsistencies, logical conflicts, illegal expressions, and missing evidence, avoiding centralized rectification of the final draft. It structurally links each statement unit with the original evidence source, QC verification results, and audit logs, fully demonstrating the data source, verification process, and derivation logic of professional conclusions, matching regulatory compliance traceability requirements, and effectively improving the professionalism of AI-generated strong regulatory reports. Credibility and compliance: Through standardized evidence constraints, multi-dimensional data verification, professional calculation review, cross-chapter logic checks, and manual review of high-risk content, multiple mechanisms are employed to reduce compliance risks such as fabricated data, miscited evidence, unfounded over-inferences, and errors in professional expression. The efficiency and accuracy of manual review are optimized by automatically distinguishing between content that has been confirmed by evidence, automatically corrected, conflicting, and requiring manual review. High-risk quality control points are accurately located, eliminating the need for QA, medical, registration, and pharmacovigilance experts to read the entire document and allowing them to focus on key risk points for review, significantly reducing ineffective workload and shortening the review cycle. It constructs a complete, auditable, and replayable report generation process (capable of reconstructing or displaying the formation process of a report from input to final output based on audit logs, evidence objects, model configurations, input / output summaries, and correction records). Each workflow node retains full input / output, evidence configuration, quality control correction, and manual review logs. It supports full-process replay and traceability of the generation of any report, chapter, or key conclusion, meeting the mandatory audit replay requirements of GLP (Good Laboratory Practice), GCP (Good Clinical Practice), and regulatory agencies; it is compatible with deterministic engineering processes and A... The AI ​​system possesses dual capabilities: a deterministic engineering system efficiently completes standardized processes such as fixed chapter structure, template application, field mapping, unit conversion, statistical calculation, and formatting; and AI focuses on advanced intelligent tasks such as contextual semantic understanding, evidence chain organization, and semantic relationship recognition, achieving optimal adaptation to the technical architecture. Furthermore, it deeply integrates human review responsibility definition, process security control, and end-to-end audit tracking into the system's underlying design, transforming AI from an uncontrollable text generation tool into a manageable, explainable, traceable, and auditable compliance productivity tool, comprehensively enhancing compliance delivery capabilities in heavily regulated industries. Attached Figure Description

[0074] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0075] In the attached diagram:

[0076] Figure 1This is the overall architecture diagram of the embedded interpretable AI quality control system for the strong regulatory report of this invention;

[0077] Figure 2 This is a flowchart of the entire embedded QC process for the mandatory regulatory reporting of this invention;

[0078] Figure 3 This is a schematic diagram illustrating the process of tracking, quality control, correction, and review of claim-level evidence in this invention.

[0079] Figure 4 This is a schematic diagram of routing matching for the type-driven QC plugin of this invention;

[0080] Figure 5 This is a schematic diagram illustrating the generation and playback of audit logs during the stages of this invention.

[0081] Figure 6 This is a schematic diagram illustrating the cross-report type reuse of the universal, strongly supervised document quality control base of this invention;

[0082] Figure 7 This is a schematic diagram of the configuration of a computer device according to an embodiment of the present invention;

[0083] Figure 8 This is a schematic diagram illustrating the structure of a computer-readable storage medium according to an embodiment of the present invention. Detailed Implementation

[0084] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of systems and products consistent with some aspects of this disclosure as detailed in the appended claims.

[0085] The terminology used in this disclosure is for the purpose of describing particular embodiments only and is not intended to be limiting of the disclosure. The singular forms “a,” “the,” and “the” as used in this disclosure and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0086] It should be understood that although the terms first, second, third, etc., may be used in this disclosure to describe various information, such information should not be limited to these terms. These terms are used only to distinguish information of the same type from one another. For example, without departing from the scope of this disclosure, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0087] The embodiments of the present invention will be described in further detail below.

[0088] Example

[0089] This invention provides a strongly oversight-based embedded interpretable AI quality control method for reports. It constructs a closed-loop chain including deterministic workflow, in-node agent, rule, and tool execution, in-node QC gating, node control, and audit package output, comprising the following steps:

[0090] S1. Receive tasks for generating scientific reports under strict supervision or reviewing existing reports, and collect task parameters, including report type, project number, document template, chapter scope, data list, operator permissions, quality control level, manual review enable configuration (whether manual review is enabled), and audit package export configuration (whether audit package export is enabled).

[0091] The aforementioned regulatory scientific reports include non-clinical research reports, clinical research reports (CSR), Investigational New Drug (IND) applications, New Drug Applications (NDA), Biologics License Applications (BLA), Pharmacovigilance (PV) reports, medical monitoring reports, real-world research reports, and quality system documents.

[0092] Report types include CSR, IND, NDA, pharmacovigilance report, medical monitoring report, non-clinical study report, real-world study report, etc.

[0093] S2. Collect multi-source heterogeneous raw data associated with the task, perform format parsing and normalization on the raw data, convert it into standardized evidence_ref evidence objects, and construct a unified evidence package; each evidence object contains a unique evidence identifier, evidence source type, evidence source name, standardized location path, original content, original value, standardized value, unit of measurement, version information, summary verification value, and access control field.

[0094] Specifically, multi-source heterogeneous raw data includes research protocols (protocol documents), statistical analysis plans, raw experimental data, statistical tables, experimental records, case report forms, medical monitoring records, safety databases, pharmacovigilance case reports, standard operating procedures (SOPs), regulatory guidelines, historical reports, customized templates, and human review comments.

[0095] S3. Based on the report type and business scenario, load the matching deterministic workflow. The workflow is implemented using a deterministic state machine, workflow DSL, or directed acyclic graph. Different report types are configured with dedicated business nodes.

[0096] S4. In the business nodes of the workflow, according to the node configuration, call the large language model, domain model, expert agent, rule engine, template engine, code sandbox, data calculation tool or evidence retrieval tool to perform at least one of the following processing: data parsing, structured processing, content generation, quality control verification or audit record;

[0097] S5. The generated report content is broken down into nine hierarchical quality control objects from top to bottom. These nine hierarchical quality control objects include fields, values, units, tables, sentences, claims, paragraphs, sections, and reports. Text span anchors are used for positioning, highlighting, citation marking, and audit playback coordinates, but are not used as business quality control objects. The claims are divided into six types: factual, calculation, process, comparison, explanation / judgment, and compliance.

[0098] The text span anchor point is only used for claim unit positioning, highlighting, reference marking and audit playback coordinates, and is not used as a business quality control object; the claim unit is used as the smallest business semantic quality control unit.

[0099] S6. Embed QC sub-processes in each business node of the workflow, configuring nine types of pluggable quality control plugins: format quality control, data quality control, calculation quality control, source quality control, logic quality control, consistency quality control, compliance quality control, expression quality control, and audit integrity quality control; route and match exclusive QC plugin combinations based on the declaration unit type (e.g., ... Figure 4 As shown), embedded quality control verification is performed; the QC verification result is written to the current business node status and used as the gating input for subsequent node release, automatic correction, manual review task creation or generation of blocking;

[0100] Based on the type of declaration unit, the routing matches a dedicated QC plugin combination, including: factual declarations match data quality control and source quality control plugins; calculation declarations match calculation quality control and data quality control plugins; process declarations match source quality control and compliance quality control plugins; comparison declarations match data quality control, calculation quality control, and logic quality control plugins; explanation and judgment declarations match source quality control, logic quality control, and expression quality control plugins; and compliance declarations match compliance quality control and audit integrity quality control plugins.

[0101] S7. Based on QC verification results, declaration unit type, risk level, and evidentiary certainty, intelligent adjudication is performed. Adjudication results include: approved, automatically corrected, unresolved, manually reviewed, blocked, and inapplicable. Automatic correction must meet the following conditions: clear source of evidence, definite testing method, risk level within allowable range, recordable differences before and after correction, explainable reasons for correction, and replayable correction actions. Explanatory and judgmental declarations, compliance declarations, and other high-risk qualitative declarations cannot be automatically replaced by human judgment. The system only generates processing suggestions, risk reasons, and review tasks, and triggers manual review.

[0102] Figure 3 This paper illustrates the process of claim-level evidence tracking, quality control, correction, and review in an embodiment of the present invention.

[0103] S8. Establish a structured relationship between report content, declaration units, QC verification results, and evidence objects, and define relationship types such as support, contradiction, partial support, insufficient evidence, inapplicable, background basis, requiring manual review, calculated by it, and corrected by evidence;

[0104] S9. Based on real QC records, declarative traceability, evidence chain, correction log, manual review status, and audit log, constrain the generation of three levels of explanatory descriptions: declaration level, chapter level, and report level, and prohibit the model from freely fabricating explanatory content.

[0105] The constraints that govern the generation of explanatory statements include: content without supporting evidence shall not be included in the explanation; data sources cited shall be marked with evidence identifiers; automatically corrected content shall clearly state the preceding and following text and the reasons; unresolved conflicts shall be explicitly marked; and manually reviewed content shall specify the reasons for review and the responsible parties.

[0106] This invention explains that the model is not generated freely, but rather constrained by real QC records, evidence chains, correction records, and manual review status.

[0107] S10. Generate stage audit logs for each business stage of the workflow. The audit logs include the run identifier, project number, document number, stage information, execution agent, model information, various digest hashes, start and end times, and chain hashes of previous and subsequent stages. The full-link logs, evidence chains, quality control results, correction records, and manual review records are uniformly stored in the database.

[0108] Specifically, the stage audit log forms a chain-like, tamper-proof structure, supporting multi-dimensional process playback and audit package export by report, chapter, declaration unit, evidence source, manual review node, and correction difference.

[0109] Specifically, each workflow node generates an input digest, output digest, evidence digest, configuration digest, and stage hash (stage-level audit logs and chained digests), supporting process integrity verification and replay (e.g. Figure 5 (As shown).

[0110] S11. Output the post-quality control report, structured QC result report, explanatory documentation, and audit data package to complete compliant delivery.

[0111] The method of this invention does not set up a separate final QC node, but rather embeds QC into the entire report generation process (e.g., Figure 2 As shown, the approach is to build a workflow-first, agent-in-node, skill-as-asset, QC-by-design framework, which means that the process is managed by a deterministic workflow, the node capabilities are improved by the agent, and the underlying technical solution is to use plug-in QC to run through every node in the entire chain.

[0112] This invention also provides a strongly regulatory reporting embedded interpretable AI quality control system (such as...). Figure 1 As shown), this system is used to implement the embedded interpretable AI quality control method for strongly regulated reports as described above. The system is divided into an input and evidence layer, a workflow and generation layer, an embedded quality control and adjudication layer, and an interpretation, audit, and output layer, including:

[0113] The task receiving module is used to receive tasks for generating scientific reports under strict supervision or reviewing existing reports, and to collect task parameters, including report type, project number, document template, chapter scope, data list, operator permissions, quality control level, manual review enable configuration, and audit package export configuration.

[0114] The multi-source evidence parsing module is used to collect multi-source heterogeneous original data associated with the task, perform format parsing and normalization processing on the original data, convert it into standardized evidence_ref evidence objects, and construct a unified evidence package; each evidence object contains a unique evidence identifier, evidence source type, evidence source name, standardized location path, original content, original value, standardized value, unit of measurement, version information, summary verification value, and access control field;

[0115] The multi-source evidence parsing module converts data in different formats into a unified `evidence_ref` object. An example is shown below:

[0116] {

[0117] "evidence_id":"evd_001",

[0118] "source_type":"statistical_table",

[0119] "source_name":"efficacy_summary.xlsx",

[0120] "source_ref_label":"[Table 3] Main Endpoint Statistics",

[0121] "canonical_locator":"excel: / / file / sheet1 / G18",

[0122] "raw_value":"12.3%",

[0123] "normalized_value": 0.123,

[0124] "unit":"%",

[0125] "digest":"sha256:xxxx"

[0126] }

[0127] The above data structure example implements the parsing and standardization of statistical table-type evidence, assigns a unique evidence identifier to it, marks the source type, source file name and reference tag, accurately locates the table cell position through the normalized locator, extracts the original value and converts it into a normalized value and uniform unit, and generates a hash digest for integrity verification and audit trail, forming a structured evidence object that can be directly used for declaration verification.

[0128] The workflow orchestration module is used to load matching deterministic workflows based on report type and business scenario. The workflows are implemented using deterministic state machines, workflow DSLs, or directed acyclic graphs, and different report types are configured with dedicated business nodes.

[0129] Specifically, for CSR reporting, the workflow includes:

[0130] Research protocol structuring; statistical analysis plan retrieval; subject distribution data analysis; efficacy data analysis; safety data analysis; chapter draft generation; data consistency QC; medical logic QC; cross-chapter consistency QC; manual medical review; final report assembly; audit package export.

[0131] For IND / NDA documents, the workflow includes:

[0132] Module structure identification; loading of application document templates; mapping of non-clinical / clinical / CMC data; verification of regulatory standards; consistency of citation QC; scanning of missing items; risk item classification; manual registration review; and generation of submission packages.

[0133] The content generation module is deployed at each workflow node and is used to call large language models, domain models, expert agents, rule engines, template engines, code sandboxes, data computing tools or evidence retrieval tools in the business nodes of the workflow according to the node configuration to perform at least one of the following processes: data parsing, structured processing, content generation, quality control verification or audit records.

[0134] The quality control object decomposition module is used to break down the generated report content into nine hierarchical quality control objects from top to bottom. The nine hierarchical quality control objects include fields, values, units, tables, sentences, claims, paragraphs, sections, and reports. Text span anchors are used for positioning, highlighting, citation marking, and audit playback coordinates, and are not used as business quality control objects. The claims are divided into six types: factual claims, computational claims, procedural claims, comparative claims, interpretive claims, and compliance claims.

[0135] Example as follows:

[0136] {

[0137] "claim_id":"clm_001",

[0138] "claim_text":"The high-dose group of animals experienced a 12.3% decrease in body weight compared to the control group."

[0139] "claim_semantic_type":"comparative",

[0140] "verification_method":"recalculation",

[0141] "regulatory_role":"result",

[0142] "risk_level":"high",

[0143] "span_anchor_id":"span_001

[0144] }

[0145] The above data structure example is implemented to create a standardized quality control unit for comparative high-risk conclusion statements in scientific reports with strong supervision. The content "the weight of animals in the high-dose group decreased by 12.3% compared with the control group" is identified with a unique ID, and is verified by recalculation. It is clearly identified as a key regulatory information of the result type, and is bound to a text positioning anchor point to achieve precise positioning, automatic quality control, evidence traceability and audit playback.

[0146] The embedded QC execution module is used to embed QC sub-processes into each business node of the workflow. It configures nine types of pluggable quality control plugins: formatQC, dataQC, computationQC, sourceQC, logicQC, consistencyQC, complianceQC, expressionQC, and audit integrityQC. Based on the claim unit type, it routes and matches specific QC plugin combinations. Different claim types trigger different QC combinations to execute embedded quality control verification, as shown in Table 1.

[0147] Table 1

[0148]

[0149] The embedded QC execution module adopts a plug-in architecture, as follows:

[0150] classQCPlugin:

[0151] qc_type:str

[0152] supported_claim_types:list[str]

[0153] defcheck(self,claim,evidence_bundle,section_context):

[0154] returnQCResult

[0155] The data structure example above defines a basic abstract class for quality control plugins in an interpretable AI quality control system, used to standardize the unified interface of all quality control verification plugins. As the base class for quality control plugins, this class declares two core attributes: quality control type and a list of supported declaration types. It also defines a unified verification entry method, check, which receives three types of parameters: the declaration to be verified, the evidence package, and the chapter context. After performing quality control verification, it returns a standardized quality control result object, providing the system with a general and extensible quality control execution specification.

[0156] QCPlugin refers to plug-in modules used to perform specific types of quality control, such as data QC plug-in, computational QC plug-in, logic QC plug-in, compliance QC plug-in, and expression QC plug-in.

[0157] QC verification results are written to the current business node status and used as gating inputs for subsequent node release, automatic correction, manual review task creation, or generation of blocking.

[0158] The adjudication and correction module is used to make intelligent adjudications based on QC verification results, declaration unit type, risk level, and evidence certainty. Adjudication results include: matched, corrected, unresolved, human-required, blocked, and not applicable. Automatic correction must meet the following conditions: clear evidence source, definite testing method, risk level within allowable range, recordable differences before and after correction, explainable reason for correction, and replayable correction action.

[0159] Example as follows:

[0160] {

[0161] "qc_check_id":"qc_002",

[0162] "claim_id":"clm_001",

[0163] "qc_type":"data_qc",

[0164] "status":"corrected",

[0165] "before_text":"Decreased by 12.3%",

[0166] "after_text":"Decreased by 13.2%",

[0167] "evidence_refs":["evd_083"],

[0168] "decision_reason": "The values ​​in the text are inconsistent with those in the statistical table. They have been corrected according to the statistical table."

[0169] "requires_human_review":false

[0170] }

[0171] The above data structure example records the complete processing result of the AI ​​quality control system after performing data verification on a statement. Verification information, modified content, basis, and review requirements are retained in a standardized format. Specifically, for the data quality control check result with ID qc_002, associated with the statement to be verified clm_001, the status is "corrected." The original text "decreased by 12.3%" is corrected to "decreased by 13.2%." The basis for the correction is evidence evd_083, because the text value is inconsistent with the statistical table, and the correction is completed according to the statistical table. This correction does not require manual review.

[0172] For claims that are explanatory, compliance-related, or high-risk, the system can provide suggestions, but it cannot directly replace the final human judgment.

[0173] The evidence chain generation module is used to establish a structured relationship between report content, declaration units, QC verification results, and evidence objects, and to define relationship types such as support, contradiction, partial support, insufficient evidence, inapplicable, background basis, requiring manual review, calculated from it, and corrected by evidence.

[0174] Relationship types include:

[0175] supports; contradictions; partially_supports; background; requires_manual_review; derived_from; corrected_by.

[0176] Example as follows:

[0177] {

[0178] "claim_id":"clm_001",

[0179] "evidence_id":"evd_083",

[0180] "relation_type":"supports",

[0181] "support_status":"matched",

[0182] "confidence": 0.98

[0183] }

[0184] The above data structure example is used to establish and record the standardized association between quality control statements and evidence. Its specific function is to establish a supporting association between statement clm_001 and evidence evd_083, determine that the two information matches, with a matching confidence level of up to 0.98, and form a traceable evidence support relationship.

[0185] The XAI explanation generation module is used to generate three levels of explanatory explanations—declaration level, chapter level, and report level—based on real QC records, declaration traceability, evidence chain, correction logs, manual review status, and audit logs, and prohibits the model from freely fabricating explanation content.

[0186] Specifically, the sources of the explanations include:

[0187] claim_trace_items; qc_check_results; evidence_refs; patch_records; unresolved_conflicts; human_review_items; audit_stage_logs.

[0188] audit_stage_logs refers to the audit logs for a specific stage of the workflow, recording the inputs, outputs, evidence, configurations, execution status, manual review status, and hash digests for that stage.

[0189] The interpreted output includes:

[0190] Chapter-level explanation; report-level explanation; claim-level explanation; data source explanation; explanation of reasons for correction; manual review explanation; audit summary.

[0191] The interpretation of generation rules includes:

[0192] Content that is not supported by evidence must not appear in explanations or descriptions;

[0193] If a data source is cited, the corresponding evidence object should be specified;

[0194] If automatic correction occurs, the before and after correction and the reason should be explained;

[0195] If any unresolved conflicts exist, they should be clearly marked;

[0196] If manual review is required, the reason for review and the responsible party should be output.

[0197] The explanation should be consistent with the QC results.

[0198] The review and audit module is used to generate stage audit logs for each business stage of the workflow;

[0199] Specifically, based on risk decisions, manual review tasks are created, review roles are assigned, review comments and review status are recorded, and review records are archived to the audit chain;

[0200] Specifically, the manual review module is embedded within the AI ​​workflow, rather than as an independent external process, to achieve a closed loop of human-machine collaborative responsibility, and the review records are fully incorporated into the audit evidence chain;

[0201] The report output module is used to output the main text of the quality control report, structured QC result reports, interpretable explanatory documents, and audit data packages to complete compliant delivery.

[0202] Audit trail module: Used to generate stage audit logs at each business stage of the workflow. The audit logs include run identifier, project number, document number, stage information, executing agent, model information, various digest hashes, start and end time, and chain hashes of previous and subsequent stages; the entire chain logs, evidence chain, quality control results, correction records, and manual review records are uniformly stored in the database;

[0203] Specifically, each workflow stage generates an audit_stage_log, which includes at least:

[0204] {

[0205] "run_id":"run_001",

[0206] "project_id":"proj_001",

[0207] "document_id":"doc_001",

[0208] "stage_name":"data_qc",

[0209] "stage_index":5,

[0210] "agent_name":"DataQCAgent",

[0211] "model_name":"model_x",

[0212] "prompt_digest":"sha256:aaa",

[0213] "input_digest":"sha256:bbb",

[0214] "output_digest":"sha256:ccc",

[0215] "evidence_digest":"sha256:ddd",

[0216] "config_digest":"sha256:eee",

[0217] "status":"completed",

[0218] "started_at":"2026-04-29T10:00:00Z",

[0219] "finished_at":"2026-04-29T10:00:08Z",

[0220] "prev_step_hash":"hash_prev",

[0221] "step_hash":"hash_current"

[0222] }

[0223] The data structure example above is a complete AI quality control execution audit record, fully recording the execution information, subject, data, time, and hash verification chain of a data quality control task. Specifically, it associates project proj_001 and document doc_001. The DataQCAgent agent, based on the model_x model, executes the data_qc data quality control task in stage 5, recording the hash digest values ​​of prompts, inputs, outputs, evidence, and configurations, as well as the start and end times and execution status. An immutable audit chain is formed through the hash values ​​of the preceding and following steps, and the task is finally completed.

[0224] Through the audit trail module, the system supports playback by report; playback by chapter; playback by claim; playback by evidence source; playback by manual review node; playback by corrected discrepancies; and export of audit packages.

[0225] Configuration Management Module: Used for configurable management of report types, workflow templates, evidence types, declaration types, QC plugins, and audit rules.

[0226] This invention employs a fully embedded QC design, unlike traditional post-processing QC. Instead of a unified check after report completion, QC is embedded at each key node of report generation. These nodes include: data acquisition, evidence standardization, chapter generation, data verification, logical verification, cross-chapter consistency verification, manual review, final draft assembly, and audit export. The report content is decomposed into multi-level QC objects: Field, value, unit, table, sentence, claim, paragraph, section, and report. Different QC methods are used for different levels of objects. This system does not use tokens as business QC units, but rather Claims as the smallest business semantic verification unit. Tokens or spans are only used for location, highlighting, referencing anchors, and audit playback coordinates. A Claim-type driven QC plugin routing is designed, automatically triggering different QC combinations for different Claim types. Factual claims trigger data and source QC; calculation-based claims trigger recalculation QC; comparison-based claims trigger data, calculation, and logic QC; explanatory claims trigger evidence boundaries and expression QC; and compliance-based claims trigger compliance and audit QC. Multi-source data is converted into a unified `evidence_ref` object, making each data point, table, text fragment, rule, or human opinion uniquely locatorable, summarizable, and citeable, achieving evidence package standardization and evidence objectification. A QC result-driven automatic correction and human review mechanism is designed. The system determines the following quality control measures based on QC type, Claim type, risk level, and evidence certainty: automatic approval, automatic correction, reminder for human review, generation of blocking mechanisms, and marking of unresolved conflicts. The system embeds human review nodes into the AI ​​workflow, rather than as external notes. Human review status, reviewer, review opinion, and review conclusion all become part of the audit chain.

[0227] Preferably, the system of the present invention features a configurable, cross-report type universal strongly regulatory document base, not limited to a single report type, but reusable across report types (e.g., Figure 6 As shown, through report type configuration, evidence type configuration, workflow template configuration, claim type configuration, QC plugin configuration, and audit rule configuration, it can be extended to the generation and quality control scenarios of CSR, IND / NDA, pharmacovigilance, medical monitoring, non-clinical, clinical, registration, and other highly regulated scientific documents.

[0228] Application examples

[0229] The present invention can be applied to a variety of practical scenarios. The following are specific application examples of the present invention in three different scenarios:

[0230] Application Example 1: Embedded interpretable AI quality control in the efficacy section of clinical research reports

[0231] This application example demonstrates embedded, interpretable AI-based quality control throughout the entire process of generating the efficacy section of a clinical research report (CSR). The specific implementation steps are as follows:

[0232] The first step is task reception and parameter configuration: The system receives the task of generating the efficacy chapter of the CSR report, configures the project number, CSR standard document template, efficacy chapter generation scope, enables medical manual review, enables the audit package export function, and locks the quality control strictness level to high compliance control.

[0233] The second step is multi-source evidence analysis and evidence package construction: Import related original data, including clinical trial study protocols, statistical analysis plans (SAP), primary / secondary efficacy endpoint statistical tables, subject baseline data records, historical CSR report templates, GCP regulatory guidelines, and laboratory SOP documents; the system automatically parses various heterogeneous data and converts them into standardized evidence_ref evidence objects. Each piece of evidence is configured with a unique ID, source type, document location path, original value, standardized value, unit, SHA verification summary, and access permissions to construct a complete efficacy chapter evidence package.

[0234] Example as follows:

[0235] {

[0236] "source_type":"protocol",

[0237] "source_ref_label":"[Research Protocol] Definition of Primary Endpoints",

[0238] "locator":"protocol: / / section / 5.1",

[0239] "digest":"sha256:xxx"

[0240] }

[0241] {

[0242] "source_type":"statistical_table",

[0243] "source_ref_label":"[Table 2] Statistical Table of Major Therapeutic Endpoints",

[0244] "locator":"excel: / / sheet / endpoint / G18",

[0245] "digest":"sha256:yyy"

[0246] }

[0247] The data structure example described above is used to establish standardized, locationable, and verifiable traceability metadata records for evidence resources in the quality control system. It performs source type labeling, citation tag description, location path definition, and hash digest verification for two core data sources: research plan text and statistical tables. This enables precise location, unique identification, and tamper-proof verification of evidence, providing a reliable traceability basis for evidence verification in AI quality control.

[0248] The third step is the loading of the deterministic workflow: The system automatically matches the CSR-specific workflow and loads the following nodes in sequence: structured analysis of the research plan, analysis of efficacy data, generation of chapter drafts, data consistency QC, medical logic QC, cross-chapter consistency quality control, manual review by medical experts, chapter assembly, and audit log storage.

[0249] The fourth step is to generate embedded content in the node: In the efficacy chapter generation node, the medical vertical large model, data calculation tools, and evidence retrieval tools are called to automatically generate the efficacy chapter draft text based on the evidence package. The generated content is forcibly bound to the evidence package version, model configuration, prompt word summary, and workflow node number.

[0250] Step 5: Multi-level quality control object decomposition: The draft text is decomposed into quality control objects such as fields, values, tables, sentences, and declaration units. Declaration units are extracted from the text: ① The treatment group showed a 12.3% improvement in the primary endpoint compared to the placebo group at week 12; ② The difference between groups was statistically significant. The system automatically classifies: Declaration ① is a comparative declaration, and Declaration ② is a calculation + interpretation judgment declaration; text positioning anchors are generated simultaneously for subsequent highlighting and playback.

[0251] Step 6, Execution of Plug-in Embedded QC: QC plugins are routed and matched based on the declaration type; comparison-type declarations trigger data quality control, calculation quality control, source quality control, and cross-chapter consistency quality control plugins; calculation and interpretation-type declarations trigger statistical quality control, source quality control, expression quality control, and manual review gate (when the system identifies high-risk, conflicting evidence, out-of-bounds interpretation, compliance judgment, or content that cannot be automatically decided, it forces entry into the manual confirmation process node); system verification found: the standard improvement value of the statistical table is 13.2%, which deviates from the text's 12.3%; the statistical report did not find a p-value or confidence interval to support the statistical significance statement.

[0252] Step 7, QC Decision and Automatic Correction: For numerical deviation statements, since the source of evidence is clear and the risk level is within the allowable range, the system performs automatic correction, records "12.3%" before correction and "13.2%" after correction, and marks the correction basis as the data in the specified cell of the efficacy endpoint statistics table; for statistical significance statements, since there is a lack of statistical evidence to support them, the system determines that the evidence is insufficient and automatically triggers the manual review process by medical statistics experts.

[0253] Specific examples are as follows:

[0254] If the statistics table shows an improvement value of 13.2%, and the system detects a discrepancy between the text and the data, it will generate a correction suggestion:

[0255] {

[0256] "before_text":"Improvement by 12.3%",

[0257] "after_text":"13.2% improvement",

[0258] "decision_reason":"The statistics table shows an improvement of 13.2% for this indicator, which is inconsistent with the original data."

[0259] "status":"corrected"

[0260] }

[0261] If a p-value is missing to support "statistically significant", then the system marks it as:

[0262] {

[0263] "claim_text":"The difference is statistically significant",

[0264] "status":"human_required",

[0265] "decision_reason": "No p-value or confidence interval was found in the statistical table to support statistical significance."

[0266] }

[0267] The data structure example above demonstrates the AI ​​quality control system's automatic verification, correction, and labeling functions for two different types of problems. Specifically, the system automatically verifies the accuracy of data in the text. When the data is inconsistent with the statistical table, it directly corrects the original text "improved by 12.3%" to "improved by 13.2%" and records the reason for the correction. When the text states "the difference is statistically significant" but lacks corresponding p-values ​​or confidence intervals, it automatically marks the content as requiring manual processing and explains that no valid statistical basis has been found, thus achieving the labeling of insufficient statistical evidence and the reminder for reviewing high-risk statistical conclusions.

[0268] Step 8, Evidence Chain Generation: Establish structured associations between the two declaration units and the corresponding statistical tables and research protocol evidence objects. The numerical correction declaration is marked as derived_from and supports association, and the statistical significance declaration is marked as insufficient and requires_manual_review association. Assign support scores and standardized judgment reasons.

[0269] Step 9, Constrained Interpretable Generation: The system generates chapter-level explanations based on real QC verification records. This chapter conducts embedded quality control verification based on the clinical trial study protocol, statistical analysis plan, and primary efficacy endpoint statistical table. It verifies the consistency, calculation logic, and source compliance of data for inter-group comparison groups, observation time points, efficacy improvement values, and statistical significance statements. The original efficacy improvement value of 12.3% is inconsistent with the standard data of 13.2% in the statistical table and has been automatically corrected according to the original evidence. The statistical significance statement did not find supporting evidence of P-value and confidence interval and has been marked as high risk, forcibly triggering manual review by medical statistics experts.

[0270] Step 10, Audit Log and Output: Each node in the workflow automatically generates a stage audit log with chained hashes, fully recording input and output summaries, evidence summaries, model configurations, quality control results, correction records, and manual review status; the final output includes the main text of the post-quality control efficacy chapter, structured QC verification reports, explanatory documents, and audit data packages, supporting subsequent audit playback and regulatory review.

[0271] Application Example 2: Embedded interpretable AI quality control for causal judgment in pharmacovigilance reports:

[0272] This application example focuses on quality control in the section on determining the causal relationship of adverse events in pharmacovigilance reports. The specific implementation process is as follows:

[0273] The system receives pharmacovigilance report generation tasks, imports adverse event occurrence time records, dosing time logs, post-discontinuation symptom follow-up records, re-challenge trial data, subject's past medical history, concomitant medication records, medical assessment logs, and other multi-source data to construct a standardized evidence package; it loads a pharmacovigilance-specific deterministic workflow to generate a draft text for adverse event causal judgment and extracts the statement unit "This serious adverse event may be related to the investigation drug".

[0274] The system automatically categorizes the statement as an explanatory and judgmental statement, and routes it for quality control, logic quality control, expression quality control, and compliance quality control plugins. After evidence matching and verification, the current evidence package only contains basic case information and lacks professional medical causal assessment records and authoritative judgment basis. The system determines that the evidence is insufficient to support the qualitative conclusion, upgrades the risk level to critical risk, prohibits automatic correction, and forcibly pushes the pharmacovigilance expert for manual review.

[0275] The system generated a binding and interpretable statement: This statement belongs to the category of medical interpretation and judgment safety causal qualitative statement. The current relevant evidence only covers the basic information of the case and lacks professional medical causal assessment basis. The evidence is insufficient and it is judged to be at a critical risk level. It has automatically triggered a manual review by pharmacovigilance experts. The review opinion will be included in the audit evidence chain and archived.

[0276] The entire process generates audit logs at each stage, retaining evidence matching records, quality control verification results, and manual review task distribution records, supporting subsequent playback verification by declaration unit.

[0277] Application Example 3: Embedded Explainable AI Quality Control for Compliance Declaration of Registration Application Materials

[0278] This application example demonstrates embedded quality control for the compliance statement content in IND registration application materials. The specific implementation process is as follows:

[0279] Receive IND registration application material preparation tasks, import evidence materials such as data integrity check records, system audit logs, data lock approval documents, document version management records, and QA manual review opinions; load the registration application-specific workflow and generate a compliance statement text "The data in this study is complete, accurate, and fully traceable".

[0280] Extract compliance declaration units and automatically match them with compliance quality control and audit integrity quality control plugins for verification; the system checks and finds that there are audit logs and QA review records, but lacks official data to lock key compliance evidence for approval; it is judged as partially supported and high-risk level, and automatic correction is not allowed, so it is automatically pushed to a registered compliance expert for manual review.

[0281] Explanatory statement generated: This statement is an official statement related to registration and application compliance. Existing evidence can partially support the completeness and traceability of the data, but the lack of data lock approval compliance certificate cannot prove the compliance of the statement. It has been marked as high risk and manual review by registration compliance experts has been initiated. After the review is completed, the quality control status and audit records will be updated.

[0282] This invention addresses the technical pain points of existing technologies, such as lagging post-quality control, coarse evidence tracing, arbitrary fabrication of AI interpretations, omissions of risks in homogeneous quality control, separation of human and machine review, and lack of audit playback. It pioneers a full-process embedded XAI-QC (XAI-QC refers to a technical mechanism that combines interpretable artificial intelligence and quality control to explain how the report content is generated, what evidence supports it, what verifications it has undergone, what corrections have been made, and whether manual review is required) architecture, declaration type plug-in routing, process-constrained interpretable generation, and chain-style audit tracing technology. It can be widely applied to the intelligent generation, process quality control, evidence traceability, manual review, and compliance auditing of various scientific reports in highly regulated fields such as pharmaceutical R&D, medical devices, and public health. It has outstanding advantages such as real-time risk control, accurate traceability, suppression of AI illusions, high review efficiency, audit playback capability, and cross-scenario reusability. It meets the global regulatory requirements for traceability, interpretability, and auditability of AI applications. It is especially suitable for scenarios that require evidence traceability, manual review, audit trails, and compliant delivery of scientific documents, such as non-clinical research reports, clinical research reports (CSR), new drug clinical trial applications (IND), new drug marketing applications (NDA), biologics license applications (BLA), pharmacovigilance reports, medical monitoring reports, real-world research reports, and quality system documents. It has broad prospects for promotion and application.

[0283] This invention also provides a computer device and a computer-readable storage medium. Figure 7 , Figure 8 These are schematic diagrams of the computer device and computer-readable storage medium provided in the embodiments of the present invention; see the accompanying drawings. Figure 7 As shown, the computer device includes: an input device 23, an output device 24, a memory 22, and a processor 21; the memory 22 is used to store one or more programs; when the one or more programs are executed by the one or more processors 21, the one or more processors 21 implement the strongly supervised reporting embedded interpretable AI quality control method provided in the above embodiments; wherein the input device 23, the output device 24, the memory 22, and the processor 21 can be connected via a bus or other means. Figure 7 Taking a bus connection as an example, the computer-readable storage medium includes a storage medium memory 33 and a stored program 32.

[0284] The memory 22, as a read / write storage medium for computing devices, can be used to store software programs and computer-executable programs, such as the program instructions corresponding to the embedded interpretable AI quality control method for the strongly supervised report described in this embodiment of the invention. The memory 22 may primarily include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a function; the data storage area may store data created based on the use of the device. Furthermore, the memory 22 may include high-speed random access memory and non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some instances, the memory 22 may further include memory remotely located relative to the processor 21, and these remote memories can be connected to the device via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0285] Input device 23 can be used to receive input digital or character information, and generate key signal inputs related to user settings and function control of the device; output device 24 may include display devices such as a display screen.

[0286] The processor 21 executes various functional applications and data processing of the device by running software programs, instructions and modules stored in the memory 22, thereby realizing the above-mentioned strongly supervised report embedded interpretable AI quality control method.

[0287] The computer equipment provided above can be used to execute the strongly supervised reporting embedded interpretable AI quality control method provided in the above embodiments, and has corresponding functions and beneficial effects.

[0288] This invention also provides a storage medium containing computer-executable instructions, which, when executed by a computer processor, are used to perform the strongly regulated reporting embedded interpretable AI quality control method provided in the above embodiments. The storage medium can be any type of memory device or storage device, including: mounting media such as CD-ROM, floppy disk, or magnetic tape; computer system memory or random access memory such as DRAM, DDRRAM, SRAM, EDORAM, Rambus RAM, etc.; non-volatile memory such as flash memory, magnetic media (e.g., hard disk or optical storage); registers or other similar types of memory elements; the storage medium may also include other types of memory or combinations thereof; furthermore, the storage medium may reside in a first computer system in which the program is executed, or it may reside in a different second computer system connected to the first computer system via a network (such as the Internet); the second computer system can provide program instructions to the first computer for execution. The storage medium includes two or more storage media that may reside in different locations (e.g., in different computer systems connected via a network). The storage medium may store program instructions (e.g., specifically implemented as a computer program) executable by one or more processors.

[0289] Of course, the computer-executable instructions provided in the embodiments of the present invention are not limited to the strongly supervised report embedded interpretable AI quality control method described in the above embodiments, but can also execute related operations in the strongly supervised report embedded interpretable AI quality control method provided in any embodiment of the present invention.

[0290] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0291] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A strongly regulated reporting embedded interpretable AI quality control method, characterized in that, Constructing a closed-loop workflow that includes deterministic workflow, execution of agents, rules, and tools within nodes, QC gating within nodes, node control, and audit package output, specifically includes the following steps: S1. Receive tasks for generating scientific reports under strict supervision or reviewing existing reports, and collect task parameters. S2. Collect multi-source heterogeneous raw data associated with the task, perform format parsing and normalization on the raw data, convert it into standardized evidence_ref evidence objects, and construct a unified evidence package; S3. Based on the report type and business scenario, load the matching deterministic workflow. The workflow is implemented using a deterministic state machine, workflow DSL, or directed acyclic graph. Different report types are configured with dedicated business nodes. S4. In the business nodes of the workflow, according to the node configuration, call the large language model, domain model, expert agent, rule engine, template engine, code sandbox, data calculation tool or evidence retrieval tool to perform at least one of the following processing: data parsing, structured processing, content generation, quality control verification or audit record; S5. The generated report content is broken down into nine hierarchical quality control objects from top to bottom. These nine hierarchical quality control objects include fields, values, units, tables, sentences, claims, paragraphs, sections, and reports. Text span anchors are used for positioning, highlighting, citation marking, and audit playback coordinates, but are not used as business quality control objects. The claims are divided into six types: factual, calculation, process, comparison, explanation / judgment, and compliance. S6. Embed QC sub-processes in each business node of the workflow, and configure nine types of pluggable quality control plugins: format quality control, data quality control, calculation quality control, source quality control, logic quality control, consistency quality control, compliance quality control, expression quality control, and audit integrity quality control; match exclusive QC plugin combinations according to the declaration unit type and perform embedded quality control verification; write the QC verification results to the current business node status and use them as gating inputs for subsequent node release, automatic correction, manual review task creation, or generation of blocking. S7. Based on QC verification results, declaration unit type, risk level, and evidentiary certainty, intelligent adjudication is performed. Adjudication results include: approved, automatically corrected, unresolved, manually reviewed, blocked, and inapplicable. Automatic correction must meet the following conditions: clear source of evidence, definite testing method, risk level within allowable range, recordable differences before and after correction, explainable reasons for correction, and replayable correction actions. Explanatory and judgmental declarations, compliance declarations, and other high-risk qualitative declarations cannot be automatically replaced by human judgment. The system only generates processing suggestions, risk reasons, and review tasks, and triggers manual review. S8. Establish a structured relationship between the report content, declaration units, QC verification results, and evidence objects, and define the relationship types as supportive, contradictory, partially supportive, insufficient evidence, inapplicable, background basis, requiring manual review, calculated from it, and corrected by evidence. S9. Based on real QC records, declarative traceability, evidence chain, correction log, manual review status, and audit log, constrain the generation of three levels of explanatory descriptions: declaration level, chapter level, and report level, and prohibit the model from freely fabricating explanatory content. S10. Generate stage audit logs for each business stage of the workflow; S11. Output the post-quality control report, structured QC result report, explanatory documentation, and audit data package to complete compliant delivery.

2. The strongly regulated report embedded interpretable AI quality control method according to claim 1, characterized in that, The regulatory-critical scientific reports include non-clinical research reports, clinical research reports (CSRs), Investigational New Drug (IND) applications, New Drug Applications (NDAs), Biologics License Applications (BLAs), pharmacovigilance reports, medical monitoring reports, real-world research reports, and quality system documents. The multi-source heterogeneous raw data include research protocols, statistical analysis plans, raw experimental data, statistical tables, experimental records, case report forms, medical monitoring records, safety databases, pharmacovigilance case reports, standard operating procedures (SOPs), regulatory guidelines, historical reports, customized templates, and human review comments.

3. The embedded interpretable AI quality control method for strong regulatory reports according to claim 1, characterized in that, The text span anchor point Span in step S5 is only used for claim unit positioning, highlighting, reference marking and audit playback coordinates, and is not used as a business quality control object; the claim unit is used as the smallest business semantic quality control unit.

4. The strongly regulated report embedded interpretable AI quality control method according to claim 1, characterized in that, The route matching of dedicated QC plugin combinations based on the declaration unit type in step S6 includes: factual declarations matching data quality control and source quality control plugins; calculation declarations matching calculation quality control and data quality control plugins; process declarations matching source quality control and compliance quality control plugins; comparison declarations matching data quality control, calculation quality control, and logic quality control plugins; explanation and judgment declarations matching source quality control, logic quality control, and expression quality control plugins; and compliance declarations matching compliance quality control and audit integrity quality control plugins.

5. The embedded interpretable AI quality control method for strong regulatory reports according to claim 1, characterized in that, The constraints that can be explained in step S9 include: content without supporting evidence should not be included in the explanation; data sources should be marked with evidence identifiers; automatically corrected content should clearly state the preceding and following text and reasons; unresolved conflicts should be explicitly marked; and manually reviewed content should indicate the reasons for review and the responsible parties.

6. The embedded interpretable AI quality control method for strong regulatory reports according to claim 1, characterized in that, The stage audit log in step S10 forms a chain-like, tamper-proof structure that supports multi-dimensional process playback and audit package export by report, chapter, declaration unit, evidence source, manual review node, and correction difference.

7. A strongly regulated reporting embedded interpretable AI quality control system, characterized in that, To implement the strongly regulated reporting embedded interpretable AI quality control method as described in any one of claims 1-6, the system is divided into an input and evidence layer, a workflow and generation layer, an embedded quality control and adjudication layer, and an interpretive audit and output layer, including: The task receiving module is used to receive tasks for generating scientific reports under strict supervision or reviewing existing reports, and to collect task parameters. The multi-source evidence parsing module is used to collect multi-source heterogeneous raw data associated with the task, perform format parsing and normalization processing on the raw data, convert it into standardized evidence_ref evidence objects, and construct a unified evidence package; The workflow orchestration module is used to load matching deterministic workflows based on report type and business scenario. The workflows are implemented using deterministic state machines, workflow DSLs, or directed acyclic graphs, and different report types are configured with dedicated business nodes. The content generation module is deployed at each workflow node and is used to call large language models, domain models, expert agents, rule engines, template engines, code sandboxes, data computing tools or evidence retrieval tools in the business nodes of the workflow according to the node configuration to perform at least one of the following processes: data parsing, structured processing, content generation, quality control verification or audit records. The quality control object decomposition module is used to break down the generated report content into nine hierarchical quality control objects from top to bottom. These nine hierarchical quality control objects include fields, values, units, tables, sentences, claims, paragraphs, sections, and reports. Text span anchors are used for positioning, highlighting, citation marking, and audit playback coordinates, but are not considered business quality control objects. The claims are divided into six types: factual, calculation, process, comparison, explanation / judgment, and compliance. The embedded QC execution module is used to embed QC sub-processes into each business node of the workflow. It configures nine types of pluggable quality control plugins: format quality control, data quality control, calculation quality control, source quality control, logic quality control, consistency quality control, compliance quality control, expression quality control, and audit integrity quality control. It routes and matches exclusive QC plugin combinations based on the declaration unit type and performs embedded quality control verification. The QC verification results are written to the current business node status and used as gating inputs for subsequent node release, automatic correction, manual review task creation, or generation of blocking. The adjudication and correction module is used to make intelligent adjudications based on QC verification results, declaration unit type, risk level, and evidentiary certainty. Adjudication results include approval, automatic correction, unresolved, manual review, generation of blocking, and inapplicable. Automatic correction must meet the following conditions: clear source of evidence, definite testing method, risk level within the allowable range, recordable differences before and after correction, explainable reasons for correction, and replayable correction actions. Explanatory judgments, compliance statements, and other high-risk qualitative statements cannot be automatically replaced by human judgment. The system only generates processing suggestions, risk reasons, and review tasks, and triggers manual review. The evidence chain generation module is used to establish a structured relationship between report content, declaration units, QC verification results, and evidence objects, and to define relationship types such as support, contradiction, partial support, insufficient evidence, inapplicable, background basis, requiring manual review, calculated from it, and corrected by evidence. The XAI explanation generation module is used to generate three levels of explanatory explanations—declaration level, chapter level, and report level—based on real QC records, declaration traceability, evidence chain, correction logs, manual review status, and audit logs, and prohibits the model from freely fabricating explanation content. The review and audit module is used to generate stage audit logs for each business stage of the workflow; The report output module is used to output the main text of the quality control report, structured QC result reports, interpretable explanatory documents, and audit data packages to complete compliant delivery.

8. The strongly regulated report embedded interpretable AI quality control system according to claim 7, characterized in that, Also includes: Audit trail module: used to generate stage audit logs at each business stage of the workflow. The audit logs include run identifier, project number, document number, stage information, executing agent, model information, various digest hashes, start and end time, and chained hashes of previous and next stages. All end-to-end logs, evidence chains, quality control results, correction records, and manual review records are stored in a unified database; Configuration Management Module: Used for configurable management of report types, workflow templates, evidence types, declaration types, QC plugins, and audit rules.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the strongly regulatory reporting embedded interpretable AI quality control method as described in any one of claims 1-6.

10. A computer device, the computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the strongly regulated reporting embedded interpretable AI quality control method as described in any one of claims 1-6.

Images