A risk identification method for large models based on the GRC context engineering framework

By using structured rule sets and deterministic loading of version identifiers and content fingerprints in GRC scenarios, combined with technologies such as rule application boundary gating and cryptographic signatures, the problems of unreproducible and untraceable conclusions in risk identification of large models are solved, achieving reproducibility, accuracy and auditability of risk identification, and improving compliance completeness and concurrent throughput efficiency.

CN123116019APending Publication Date: 2026-11-17SUZHOU RUIKE DIGITAL INTELLIGENCE TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611415163.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-12
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In the context of strict regulatory GRC, existing risk identification solutions based on large-scale models generated by retrieval enhancement cannot guarantee the reproducibility and consistency of risk conclusions. They suffer from problems such as rule mismatch, unreproducible conclusions, and lack of tamper-proof and auditable features.

Method used

By acquiring and organizing structured rule sets according to risk classification, deterministic loading is performed using version identifiers and content fingerprints, and the rules are routed to the large language model in batches. Combined with technologies such as rule application boundary gating, cryptographic signature anti-tampering, and virtual thread concurrent execution, the consistency of the rule set and the traceability of the conclusions are ensured.

Benefits of technology

It achieves reproducibility and traceability consistency of risk conclusions, improves identification accuracy and tamper-proof auditability, ensures the stability and compliance of conclusions, and improves the throughput efficiency of multi-class concurrent review.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116019A_ABST
    Figure CN123116019A_ABST
Patent Text Reader

Abstract

This invention discloses a large-scale risk identification method based on the GRC context engineering framework, comprising: deterministically loading a pre-extracted and decomposed structured rule set from regulatory documents after consistency verification by version identifier and content fingerprint, replacing retrieval-based recall; selecting rule subsets according to the risk category to which the business data belongs, and deterministically arranging context windows according to priority using category-specific prompt word templates; routing the tasks to be identified to the large language model in batches according to risk category, with each batch only inputting the rule subset corresponding to the category; and aggregating the results of each category to generate a traceable risk conclusion with rule identifiers and content fingerprints. This invention enables consistent and reproducible risk conclusions for the same input across different call scenarios, and improves the accuracy and auditability of GRC risk identification through boundary gating, overall signature anti-tampering, and regulatory clause mapping, meeting the requirements of traceability consistency in highly regulated scenarios.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to large-scale model risk identification technology in the field of corporate governance, risk and compliance (GRC), specifically to a large-scale model GRC risk identification method based on a context engineering framework and with deterministic rule supply guaranteeing traceability and reproducibility. Background Technology

[0002] The large-scale GRC risk identification scheme based on retrieval-enhanced generation (RAG) uses real-time retrieval of knowledge bases and rule bases to assist the large model in generating risk conclusions. It has advantages in improving identification coverage and flexibility and is one of the current mainstream technical approaches.

[0003] However, in highly regulated scenarios such as finance and compliance of central and state-owned enterprises, risk assessment conclusions need to be reproducible and auditable—that is, the same input should yield consistent tracing conclusions and rule bases at different times and under different model calls, so as to facilitate regulatory review and accountability. Real-time retrieval-based solutions suffer from inherent uncertainties in retrieval results (such as the top K candidate sets, vector recall order, and reordered output), making it difficult to guarantee complete consistency between rule sets and conclusions across different calls; the same business material may be loaded into different rule fragments during two reviews due to recall fluctuations, resulting in unreproducible conclusions.

[0004] Furthermore, when organizing rules by risk classification, existing solutions lack definitive gating of the rules' applicable boundaries, which can easily lead to rule mismatches (including rules unrelated to the current business in the context); the conclusion generation process lacks an auditable mechanism to prevent tampering, making it difficult to prove that "the conclusion has not been modified afterward"; and there is a lack of efficient concurrent orchestration methods when conducting concurrent reviews of multiple classifications.

[0005] Therefore, there is an urgent need for a GRC risk identification method that can take into account domain identification capabilities, as well as ensure that the conclusions are reproducible, traceable, tamper-proof, and auditable. Summary of the Invention

[0006] The technical problem this invention aims to solve is: in a strongly regulated GRC scenario, how to replace retrieval-based recall with deterministic rule supply, so that the same input loads a consistent rule set and context under different calls, thereby ensuring the reproducibility and traceability consistency of risk conclusions, while improving the accuracy and auditability of identification through classification context orchestration, boundary gating and tamper-proof signatures.

[0007] This invention provides a method for risk identification in large models based on the GRC context engineering framework, comprising the following steps: S1. Obtain a structured rule set that has been extracted and split from the policy documents in advance. The rule set is organized according to risk classification and associated with version identifiers and content fingerprints. When each review is initiated, the rule set is deterministically loaded after a consistency check based on the version identifiers and content fingerprints, so that the content of the rule set loaded each time is consistent under the same review scenario, and the rules are obtained without retrieval recall. S2, for the GRC business data to be identified, according to the risk classification to which the business data belongs, select the corresponding rule subset from the deterministically loaded rule set, and call the prompt word template specific to the classification to construct the model context, wherein the rule subsets of different risk classifications are deterministically arranged in the context window according to a preset priority; S3 routes the tasks to be identified to the large language model in batches according to risk classification, so that each batch only inputs a subset of rules and context corresponding to the classification, instead of inputting all rules and all business data into the same model at once. S4 receives the risk assessment results returned by the large language model according to each category, aggregates them into a unified risk conclusion, and attaches the rule identifier and content fingerprint on which each category is based, forming a traceable conclusion record; Since S1 replaces retrieval-based recall with deterministic supply of content fingerprint verification, the rule set loaded under multiple independent calls of the same input remains consistent with the generated context, making the risk conclusion and its rule basis reproducible and traceable.

[0008] In further technical solutions, the present invention also includes: gating verification of the rule application boundary, cryptographic signature anti-tampering of the conclusion as a whole, function call with confidence and manual review routing, concurrent execution of virtual threads, anchoring of regulatory version validity and recalculation prompts of historical conclusions, comparison of immutable review snapshots and cross-period deterministic differences, and source mapping of regulatory clauses and proof of compliance completeness.

[0009] The present invention has the following beneficial effects: 1. Reproducibility and traceability: The rule set is provided with version identifiers and content fingerprints for deterministic supply, does not rely on retrieval and recall, and loads the same input across calls in a consistent manner. The risk conclusions and the basis of the rules are reproducible and traceable. 2. Improved recognition accuracy: Context windows are definitively arranged and routed in batches according to risk classification, avoiding noise interference and context dilution caused by inputting all rules and all data at once, thus improving the accuracy of each classification. 3. Prevent rule mismatch: Each rule is associated with a structured applicable condition predicate. Boundary gating is performed before selecting a subset of rules. Only rules that meet the conditions are included in the context, and rules that do not meet the predicate are forcibly excluded and recorded. 4. Tamper-proof and auditable: The input data, rule set, prompt word template and model version involved in the generation of the conclusion are cryptographically signed as a whole. If any element is changed, the verification will fail and be marked in red, thus realizing the tamper-proof and auditable conclusion. 5. Safeguarding the stability of conclusions: The large language model outputs judgment results with confidence levels through structured function calls. Low-confidence results are automatically routed to manual review, further stabilizing the conclusions based on deterministic supply. 6. Improved Concurrency Throughput: Batch routing is executed concurrently by a virtual thread pool, with each category batch completing the invocation and aggregation in an independent thread, improving the throughput efficiency of multi-category concurrent review; 7. Timeliness of regulations: Rules are linked to the version of regulations and the effective date window. When regulations are updated, the list of existing historical conclusions that are affected is marked and the scope that needs to be reviewed is indicated. 8. Cross-period consistency retrospective: Each review generates an immutable snapshot, performs a deterministic difference comparison between any two historical reviews, and outputs the rule changes and the resulting differences in conclusions; 9. Compliance completeness proof: The rules are linked to the source regulatory provisions and traceable. The conclusion is accompanied by the regulatory provisions and outputs proof of provision coverage or omission gaps, which elevates risk identification to a defensible compliance completeness proof. Attached Figure Description

[0010] Figure 1 This is a flowchart of a large model risk identification method based on the GRC context engineering framework of the present invention. Detailed Implementation

[0011] The present invention will be further described in detail below with reference to specific embodiments. Those skilled in the art will understand that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0012] Example 1: Core Method like Figure 1 As shown, the method of this embodiment includes steps S1 to S4, which are described below in conjunction with... Figure 1 The preparation of each step and rule set is explained.

[0013] Preparation of the rule set: Before implementing the method of this invention, a structured rule set needs to be prepared. This rule set is obtained by pre-extracting and splitting institutional documents (such as laws and regulations, regulatory provisions, internal compliance systems, risk management systems, etc.). The extraction and splitting can be achieved with the help of existing document structuring technologies in the field, such as clause extraction based on natural language processing, parsing tools based on rule templates, and semantic parsing based on large models. The pre-extraction and splitting process itself is a known technology in the field and is not an improvement of this invention; this invention directly obtains the pre-extracted structured rule set as the input for the subsequent S1 step.

[0014] The structured rule set is organized by risk category (such as anti-fraud, data security, related-party transactions, compliance, operational risk, etc.), and each rule is associated with a version identifier and a content fingerprint. The version identifier is used to mark the release version of the rule set (e.g., v2026.01); the content fingerprint is a cryptographic hash value calculated based on the full text of the rule set, used to uniquely identify the content of this version of the rule set.

[0015] S1 Deterministic Loading: At each review initiation, the system performs a consistency check on the rule set based on the version identifier and content fingerprint (e.g., comparing the hash value of the currently loaded rule set with the expected fingerprint). If the check passes, the rule set is deterministically loaded. Deterministic loading means that under the same review scenario, the content of the loaded rule set is completely consistent each time, and rules are obtained without retrieval-based recall (such as vector retrieval, top K item recall, or reordering). This is fundamentally different from the RAG solution's "real-time retrieval on each call": this invention replaces retrieval recall with a fixed supply of fingerprint verification.

[0016] S2 Classification Selection and Context Arrangement: For GRC business data to be identified (such as a purchase contract, a cash flow statement, or a personnel appointment / removal), the system first determines the risk category to which the business data belongs. Then, it selects the rule subset corresponding to that category from the deterministically loaded rule set and calls the category-specific prompt word template to construct the model context. Rule subsets of different risk categories are deterministically arranged in the context window according to a preset priority (e.g., compliance red line categories take precedence over general risk categories), ensuring that key rules occupy a stable position within the context window.

[0017] S3 Batch Routing: The system routes tasks to be identified to the large language model in batches according to risk classification. Each batch only inputs a subset of rules for the corresponding category and the context of that category, instead of inputting all rules and all business data into the same model at once. For example, "anti-fraud" and "data security" are called as two independent batches, and the batches do not affect each other.

[0018] S4 Aggregation and Traceable Records: The system receives risk assessment results returned by the large language model according to each category, aggregates them into a unified risk conclusion, and attaches the rule identifiers and content fingerprints on which each category is based, forming a traceable conclusion record. This record includes at least: the conclusion content, the conclusion for each category, the corresponding rule identifier (such as the rule number), and the content fingerprint.

[0019] Reproducibility explanation: Since S1 replaces retrieval-based recall with deterministic supply of content fingerprint verification, the rule set loaded and the generated context remain consistent under multiple independent calls for the same input. Therefore, the risk conclusions and their rule basis have reproducibility and traceability consistency. When the same business material is reviewed at different times, as long as the rule set version and content fingerprint remain unchanged, the conclusions and their rule basis are consistent.

[0020] Example 2: Rule Application Boundary Gating Based on Example 1, each rule in the rule set is associated with a structured applicability predicate, which characterizes the application boundary of the rule (e.g., "applicable object = purchase contract and amount > 1 million yuan"). Before selecting a subset of rules in S2, the system first performs boundary gating verification on the candidate rules based on the applicability predicate: only rules that meet the corresponding business data applicability conditions are included in the context; rules that do not meet the predicate are forcibly excluded, and a record is left in the conclusion record (recording "rule R is excluded because it does not meet predicate P").

[0021] This gating ensures that all rules entering the context are related to the current business data, avoiding misjudgments caused by rule mismatches, while the traceability requirement for "exclusion reasons" is met by the audit.

[0022] Example 3: Conclusion - Overall Signature Anti-Tampering Based on Example 1, after the risk conclusion is generated in S4, the system performs a cryptographic signature on the input data hash, rule set content fingerprint, prompt word template hash and model version number involved in the conclusion generation (such as using SHA-256 or the national cryptographic SM3 to calculate the digest and then sign it), and attaches the signature to the conclusion record.

[0023] In any subsequent review, if any of the above elements change (e.g., the rule set is updated, the prompt word template is modified, or the model version is upgraded), the recalculated signature will be inconsistent with the original signature, the verification will fail and be highlighted in red, thus proving that the conclusion has not been tampered with since its generation, achieving tamper-proof and auditable results.

[0024] Example 4: Function calls with confidence and manually reviewed routes Building upon Example 1, the large language model outputs risk assessment results with confidence levels via structured function calls (e.g., returned as structured key-value pairs: classification = high risk, confidence level = 0.82). The system sets a confidence threshold (e.g., 0.7): classification results with confidence levels below the threshold are routed for manual review; classification results with confidence levels above the threshold are automatically approved. This mechanism, based on the deterministic rule supply, performs another stability filter on the model output, further ensuring the stability of the conclusions.

[0025] Example 5: Virtual Thread Concurrency Building upon Example 1, the process of S3 batch routing to the large language model is executed concurrently by a virtual thread pool: each category batch completes model invocation and result aggregation in an independent virtual thread, without blocking each other. Lightweight virtual threads (such as Java Virtual Threads) have lower resource overhead than traditional operating system threads, which can significantly improve the throughput efficiency of multi-category concurrent review.

[0026] Example 6: Regulatory Version Timeliness Anchoring and Historical Conclusion Recalculation Prompt Based on Example 1, each rule in the rule set is associated with a regulatory version identifier and an effective time window (e.g., "Based on the 2026 revised version of the Basic Norms for Enterprise Internal Control, effective from 2026-01-01 to 2026-12-31"). The risk conclusion record is associated with review time and space coordinates, including the review execution time and the applicable regulatory version.

[0027] When a regulatory version update leads to a change in the rule set version, the system uses the version identifier to definitively mark the list of existing historical conclusions that are affected (i.e., which historical conclusions were based on rules that have been replaced by the new version) to indicate the scope that needs to be reviewed again, thus preventing the continued misuse of historical conclusions under outdated regulations.

[0028] Example 7: Comparison of Immutable Review Snapshots and Intertemporal Deterministic Differences Based on Example 1, the system generates an immutable review snapshot each time a review is performed. The snapshot contains the input data hash, rule fingerprint, prompt word template hash, and model version number for this review, and cannot be changed once generated.

[0029] In response to a retrospective request for any two historical reviews, the system performs a deterministic difference comparison (compares the four elements item by item) on the corresponding two snapshots, outputs the rule change content (e.g., "the rule set has changed from v2026.01 to v2026.02") and the corresponding change in conclusion (e.g., "the original conclusion was low risk, and the revised conclusion is medium risk"), thereby supporting cross-period consistency retrospective and regulatory verification.

[0030] Example 8: Regulatory Clause Source Mapping and Compliance Completeness Proof Building upon Example 1, each rule in the rule set is associated with a source regulatory clause identification (e.g., corresponding to Article XX of the Securities Law or Article XX of the Guidelines for Internal Control of Commercial Banks). When generating risk conclusions in S4, the system includes the regulatory clause identification for each category and, based on the mapping relationship between the rule set and regulatory clauses, outputs a clause coverage certificate (listing all regulatory clauses covered in this review) or a gap / omission alert (listing regulatory clauses that should be covered but are not). This makes the risk identification result not only a risk list but also a defensible proof of compliance completeness, directly serving regulatory reporting and compliance self-inspection.

[0031] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Those skilled in the art can make various equivalent transformations and combinations without departing from the spirit and scope defined by the claims, and these all fall within the protection scope of the present invention.

Claims

1. A method for risk identification in large models based on the GRC context engineering framework, characterized in that, Includes the following steps: S1, Obtain a structured rule set that has been extracted and split from the policy documents in advance. The rule set is organized according to risk classification and is associated with version identifiers and content fingerprints. At each review start, the rule set is deterministically loaded after consistency verification based on the version identifier and content fingerprint, so that the rule set content loaded each time is consistent under the same review scenario, and the rules are not retrieved through retrieval recall. S2, for the GRC business data to be identified, according to the risk classification to which the business data belongs, select the corresponding rule subset from the deterministically loaded rule set, and call the prompt word template specific to the classification to construct the model context, wherein the rule subsets of different risk classifications are deterministically arranged in the context window according to a preset priority; S3 routes the tasks to be identified to the large language model in batches according to risk classification, so that each batch only inputs a subset of rules and context corresponding to the classification, instead of inputting all rules and all business data into the same model at once. S4 receives the risk assessment results returned by the large language model according to each category, aggregates them into a unified risk conclusion, and attaches the rule identifier and content fingerprint on which each category is based, forming a traceable conclusion record; Since S1 replaces retrieval-based recall with deterministic supply of content fingerprint verification, the rule set loaded under multiple independent calls of the same input remains consistent with the generated context, making the risk conclusion and its rule basis reproducible and traceable.

2. The method according to claim 1, characterized in that, Each rule in the rule set is associated with a structured applicability predicate, which is used to characterize the applicability boundary of the rule. Before selecting a subset of rules in S2, the candidate rules are first subjected to boundary gating verification based on the applicability predicate. Only rules that meet the applicability conditions of the corresponding business data are included in the context, while rules that do not meet the predicate are forcibly excluded and recorded.

3. The method according to claim 1, characterized in that, After generating the risk conclusion in S4, the input data hash, rule set content fingerprint, prompt word template hash, and model version number involved in the conclusion generation are cryptographically signed as a whole, and the signature is attached to the conclusion record. In any subsequent review, if any of the above elements are changed, the signature verification fails and is highlighted in red to achieve tamper-proof and auditable conclusions.

4. The method according to claim 1, characterized in that, The large language model outputs risk assessment results with confidence levels using structured function calls. Classification results with confidence levels below a preset threshold are routed to manual review, while classification results with confidence levels above the threshold are automatically approved, thereby further ensuring the stability of the conclusions based on the deterministic supply.

5. The method according to claim 1, characterized in that, The process of routing to the large language model in batches is executed concurrently by a virtual thread pool. Each category batch completes model invocation and result aggregation in an independent thread to improve the throughput efficiency of multi-category concurrent review.

6. The method according to claim 1, characterized in that, Each rule in the rule set is associated with a regulatory version identifier and an effective time window; the risk conclusion record is associated with review time and space coordinates, which include the review execution time and the applicable regulatory version; when a regulatory version update causes a change in the rule set version, the list of affected existing historical conclusions is definitively marked based on the version identifier to indicate the scope that needs to be re-reviewed.

7. The method according to claim 1, characterized in that, Each time a review is performed, an immutable review snapshot is generated. The snapshot contains the input data hash, rule fingerprint, prompt word template hash, and model version number of the current review. In response to a retrospective request for any two historical reviews, a deterministic difference comparison is performed on the corresponding two snapshots, and the rule changes and the resulting differences in conclusions are output to support cross-period consistency retrospective.

8. The method according to claim 1, characterized in that, Each rule in the rule set is associated with a source regulatory clause traceability identifier; when S4 generates a risk conclusion, it attaches the regulatory clause identifier on which each category is based, and outputs clause coverage proof or omission gap prompts based on the mapping relationship between the rule set and regulatory clauses, so as to upgrade risk identification to a defensible proof of compliance completeness.