Gap-covering drive-based trusted knowledge production flywheel method and system

Abstract

The application discloses a kind of gap-driven coverage credible knowledge production flywheel method and system.The method comprises: according to the knowledge package of publishing area existing knowledge, the coverage gap of target knowledge range is calculated, when coverage gap exceeds threshold, data demand is created and formal output is blocked;According to data demand, multi-source material collection, snapshot analysis, deduplication conflict processing are carried out, and candidate knowledge is generated and written into candidate area;Review by the review subject independent of candidate generation subject, write into publishing area after certainty check and hash binding;Failure after application, error correction and coverage gap are converted into next round task.The application avoids overstepping use of candidate material by three-area isolation and independent review mechanism of candidate area, review area, publishing area, reduces repeated collection and false coverage by coverage gap quantization, and realizes credible knowledge production flywheel of auditable, publishable and feedbackable.

Description

Technical Field

[0001] This invention relates to the fields of computer knowledge processing and artificial intelligence technology, specifically to a flywheel method and system for trusted knowledge production driven by coverage gaps. Background Technology

[0002] With the widespread application of Retrieval Augmentation (RAG) technology and knowledge-based intelligent systems, the credibility and auditability of knowledge production processes have become increasingly critical issues. Existing technologies, such as data processing, version control, and workflow traceability systems (e.g., US11042523B2), can record data cleaning, entity parsing, transformation, and publishing processes. However, their objects are typically data tables or datasets, and they do not determine whether a particular knowledge topic is adequately covered, nor do they distinguish the access boundaries of whether candidate knowledge can be used by formal intelligent systems.

[0003] Retrieval enhancement and knowledge graph enhancement technologies (e.g., US20250112878A1, US20250131289A1) enhance the output of the model by retrieving external documents. Their focus is on how to retrieve and utilize knowledge during the response process, but they do not address how to independently review, verify, hash-bind, and officially publish the retrieved information after it has been processed into candidate objects. RAG re-hint and data flywheel feedback technologies utilize knowledge graphs to determine re-hint schemes, improving subsequent hints or retrieval strategies based on knowledge graph relationships. However, they lack a complete trusted knowledge production process encompassing gap quantification, data requirements, candidate processing, review and publication, and application feedback (e.g., US20250371066A1).

[0004] Existing technologies have the following drawbacks: First, retrieved data is often directly used to answer questions, lacking a strong boundary between candidate knowledge and formal knowledge. This leads to insufficiently verified, duplicate, outdated, or low-reliability data being packaged as formal knowledge. Second, existing RAGs focus primarily on recall quality, lacking quantitative judgment on what knowledge is still missing from the target scope, easily resulting in duplicate collection and false coverage. Third, the candidate generation entity and the review entity are prone to having the same origin, leading to model self-review or context-dependent bias, reducing the effectiveness of review. Fourth, there is a lack of consistency binding between candidate content, review reports, verification results, and the release package before release, meaning that content may be replaced after release and cannot be traced. Fifth, application failures and user corrections are difficult to reliably translate into the next round of data production tasks, and knowledge assets cannot be continuously updated. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a flywheel method and system for trusted knowledge production driven by coverage gaps, so that the knowledge-based intelligent system only creates data requirements when the target scope cannot be met in the publishing area; multi-source retrieval, snapshot parsing, deduplication and conflict handling, and candidate knowledge generation are carried out around the data requirements; candidate knowledge can only enter the publishing area after independent review, deterministic verification, and hash binding; failures, error corrections, and coverage gaps after formal application continue to be fed back as data requirements for the next round, thereby avoiding unauthorized use of candidate data and false coverage.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A flywheel method for trusted knowledge production driven by coverage gaps includes: acquiring a target knowledge scope, the target knowledge scope including at least one sub-target; calculating the coverage satisfaction degree of the publishing area for the target knowledge scope based on existing knowledge packages in the publishing area, and determining a coverage gap based on the coverage satisfaction degree; when the coverage gap is greater than or equal to a preset gap threshold, creating a data requirement corresponding to the target knowledge scope, and blocking the formal output of the target knowledge scope based on the publishing area; generating multiple types of query statements based on the data requirement, collecting data from multiple sources based on the query statements, saving the original snapshot of the collected data, and parsing the original snapshot to obtain a parsed document; performing deduplication and conflict recording on the parsed document based on normalized URLs, content hashes, similarity, and version identifiers, and using the deduplicated extraction result as a candidate knowledge base. The candidate knowledge is written into the candidate area, which is only allowed to be read by the review process and not by the formal application module. The review body, independent of the candidate generation body, reviews the candidate knowledge in the candidate area and outputs a review conclusion of pass, rejection, or need for supplementation. For the candidate knowledge with a review conclusion of pass, a deterministic verification is performed, which includes mandatory field checks, boundary declaration checks, duplicate conflict checks, golden example checks, and release scope checks. For the candidate knowledge that passes the deterministic verification, hash binding is performed to bind the candidate content, review report, verification result, and release package in a consistent manner to form a release consistency proof. The candidate knowledge that has completed hash binding is written into the release area as formal knowledge. And the reasons for failure, user corrections, coverage gaps, and expiration events of the formal knowledge during the application process are collected and transformed into the next round of data requirements, revision tasks, or retirement tasks.

[0007] Preferably, the step of calculating the coverage satisfaction of the publishing area for the target knowledge scope based on the existing knowledge packages in the publishing area, and determining the coverage gap based on the coverage satisfaction, includes: assigning an importance weight to each sub-target in the target knowledge scope, wherein the sum of the importance weights of all sub-targets is 1 and each weight is non-negative; calculating the coverage satisfaction of the publishing area for each sub-target, wherein the coverage satisfaction ranges from 0 to 1; multiplying the importance weight of each sub-target by the corresponding coverage satisfaction and summing the results to obtain the coverage satisfaction of the publishing area for the target knowledge scope; and subtracting the coverage satisfaction from 1 to obtain the coverage gap, the calculation formula being as follows: =Σ( × ), =1- Equation (1).

[0008] Preferably, the data requirements include target scope, missing knowledge points, minimum source requirements, acceptable uses, and prohibition statements; wherein, the minimum source requirements specify the number and type of sources required for the candidate knowledge, the acceptable uses specify the application scenarios in which the knowledge can be used, and the prohibition statements specify the content that the knowledge should not contain or the scenarios in which it should not be used.

[0009] Preferably, the various types of query statements include core queries, synonym queries, counterexample queries, file queries, and site queries; the parsed documents are deduplicated and conflict recorded based on normalized URLs, content hashes, similarity, and version identifiers, including: unifying different URL forms of the same webpage through normalized URLs, identifying completely identical documents through content hashes, identifying highly similar but not completely identical documents through similarity, and distinguishing different versions of the same material through version identifiers; for conflicting materials, recording the conflict content and conflict source; calculating the duplication rate and conflict rate, where the duplication rate is the proportion of duplicate sources to the total collected sources, and the conflict rate is the proportion of conflicting candidate knowledge to the total candidate knowledge.

[0010] Preferably, the review of candidate knowledge in the candidate area by a reviewing entity independent of the candidate generating entity includes: confirming that the reviewing entity and the candidate generating entity are different entities; the reviewing entity examines the accuracy of the content, reliability of the source, and completeness of the boundary statements of the candidate knowledge; and outputting a review conclusion of pass, rejection, or need for supplementation, wherein when the review conclusion is that supplementation is needed, a supplementary data requirement is generated and the data collection step is returned; when the review conclusion is rejection, the candidate knowledge does not enter the subsequent process.

[0011] Preferably, after performing hash binding on the candidate knowledge that has passed the deterministic verification, the method further includes: calculating a publication credibility score for the candidate knowledge, wherein the publication credibility score is calculated by weighting an evidence chain health score, a usefulness score, a difference score, and a conflict score, and the calculation formula is as follows: = + + - Equation (3), where, Assess the credibility of the publication of candidate knowledge object e. To score the health of the chain of evidence, For usefulness rating, Score the difference. Rate the level of conflict. , , , In some embodiments, preset weighting coefficients are used. =0.4, =0.3, =0.2, =0.1; when the publication credibility score is greater than or equal to the preset publication threshold, the candidate knowledge is allowed to enter the publication area; and when the publication credibility score is less than the preset publication threshold, the candidate knowledge is rejected from entering the publication area and a revision task is generated.

[0012] Preferably, the evidence chain health score is calculated by detecting the source accessibility, hash consistency, review completeness, authorization timeliness, and dissemination completeness of candidate knowledge, respectively, to obtain the missing rates for the corresponding five dimensions. , , , , ; and subtract the average of the missing rates of the five dimensions from 1 to obtain the evidence chain health score, calculated using the following formula: =1-( + + + + Equation (2) is 5 / 5, where: source accessibility checks whether the source of the original data is still accessible; hash consistency checks whether the hash value of the candidate content is consistent with that of the original snapshot; review integrity checks whether the review report is complete; authorization timeliness checks whether the authorization of the data is within the validity period; and dissemination integrity checks whether the knowledge is complete and has not been tampered with during the dissemination process. The missing rate of each dimension ranges from 0 to 1.

[0013] This invention also provides a flywheel system for trusted knowledge production driven by coverage gaps, comprising: a coverage access control module, used to acquire a target knowledge range, calculate the coverage satisfaction of the publishing area with the target knowledge range based on existing knowledge packages in the publishing area, and determine the coverage gap; when the coverage gap is greater than or equal to a preset gap threshold, a data requirement is created and formal output is blocked; a data requirement module, used to record the data requirement, the data requirement including the target range, missing knowledge points, minimum source requirements, acceptable uses, and prohibition statements; a query combination module, used to generate core queries, synonym queries, counterexample queries, file queries, and site queries based on the data requirement; a snapshot parsing module, used to collect data from multiple sources based on the query, save the original snapshot, and parse the original snapshot to obtain a parsed document; a deduplication and conflict removal module, used to perform deduplication processing and conflict recording on the parsed document based on normalized URLs, content hashes, similarity, and version identifiers; and a candidate knowledge module, used to process the deduplication... The extracted results are written as candidate knowledge into the candidate area, which can only be read by the review process and not by the formal application module. An independent review module, independent of the candidate generation entity, reviews the candidate knowledge in the candidate area and outputs a pass, rejection, or supplementation decision. A deterministic verification module performs deterministic verification on candidate knowledge that has passed the review, including mandatory field checks, boundary declaration checks, duplicate conflict checks, golden example checks, and release scope checks. A hash binding module performs hash binding on candidate knowledge that has passed the deterministic verification, binding the candidate content, review report, verification results, and release package to form a release consistency proof. A release area stores the formal knowledge that has completed hash binding. A feedback reproduction module collects the reasons for failures, user corrections, coverage gaps, and expiration events of the formal knowledge during application, transforming them into the next round of data requirements, revision tasks, or retirement tasks.

[0014] Preferably, the coverage access control module is further configured to: set an importance weight for each sub-target in the target knowledge scope, the sum of the importance weights of all sub-targets being 1 and each weight being non-negative; calculate the coverage satisfaction of the publishing area for each sub-target respectively; multiply the importance weight of each sub-target by the corresponding coverage satisfaction and sum them to obtain the coverage satisfaction of the publishing area for the target knowledge scope; subtract the coverage satisfaction from 1 to obtain the coverage gap; the independent review module is further configured to: confirm that the review subject and the candidate generation subject are different subjects; have the review subject review the accuracy of the candidate knowledge content, the reliability of the source, and the completeness of the boundary declaration; output a review conclusion of pass, rejection, or need for supplementation, wherein when the review conclusion is that supplementation is needed, a supplementary data requirement is generated and the query combination module is notified to re-collect data.

[0015] Preferably, the system further includes a release license judgment module and an evidence chain health detection and repair module; the release license judgment module is used to: calculate the release credibility score of candidate knowledge, the release credibility score being calculated by weighting the evidence chain health score, usefulness score, difference score, and conflict score; when the release credibility score is greater than or equal to a preset release threshold, the candidate knowledge is allowed to enter the release area; when the release credibility score is less than the preset release threshold, the candidate knowledge is rejected from entering the release area and a revision task is generated; wherein, the evidence chain health score is obtained by detecting the source accessibility, hash consistency, review integrity, authorization timeliness, and dissemination integrity of candidate knowledge to obtain the missing rate of the corresponding five dimensions. The average of the missing rates of the five dimensions is subtracted from 1 to obtain the result. The evidence chain health detection and repair module is used to perform evidence chain health detection and repair on the formal knowledge in the publishing area, including: reading the evidence chain of formal knowledge, which includes the publishing package, verified memory, context projection, and published content; sequentially performing source reachability detection, hash consistency detection, review integrity detection, authorization timeliness detection, and propagation integrity detection; when a break in the evidence chain is detected, a repair plan is generated; it is determined whether the break can be re-verified. If it can be re-verified, re-verification collection and re-examination are performed, and hash binding is rebuilt; if it cannot be re-verified but can be replaced, a replacement evidence block is generated; if it is unrecoverable, active retirement is performed; the repair result is recorded and the health score is updated.

[0016] Compared with the prior art, the present invention has the following beneficial effects: First, the process of separating the retrieval of data from the formal use of knowledge is achieved through a three-zone separation of candidate area, review area, and release area. This ensures that candidate knowledge cannot bypass the review process and directly affect the formal output, thereby reducing the risk of unauthorized use of candidate data.

[0017] Second, the missing knowledge is transformed into a computable access control function using the coverage gap function. Data demand is created and collection is initiated only when the coverage gap exceeds a preset threshold. Coverage gap quantification reduces duplicate collection and false coverage.

[0018] Third, by establishing an independent review mechanism that separates the candidate generation entity from the review entity, we can reduce model self-review and homology bias, thereby improving the effectiveness of the review process.

[0019] Fourth, by combining deterministic verification with hash binding in the release licensing mechanism, candidate content, review reports, verification results and release packages are bound together in a consistent manner to form a release consistency certificate, preventing the release content from being replaced and thus untraceable.

[0020] Fifth, the feedback reproduction module transforms the reasons for failures, user corrections, coverage gaps, and expiration events after formal application into the next round of data requirements, revision tasks, or retirement tasks, so that knowledge assets are continuously updated and an auditable, publishable, and feedback-enabled knowledge production flywheel is formed. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the evidence chain object and the three-layer detection and repair architecture of the present invention; Figure 2 This is a schematic diagram of the evidence chain health detection and repair process of the present invention. Detailed Implementation

[0022] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0023] This invention breaks down the knowledge production process into six stages: access control, data collection, candidate processing, independent review, reliable release, and feedback reproduction. It also ensures that candidate knowledge cannot bypass the review process and directly affect the formal output by isolating the candidate area, review area, and release area.

[0024] The system of this invention includes an access control module, a data requirement module, a query combination module, a snapshot parsing module, a deduplication and conflict removal module, a candidate knowledge module, an independent review module, a deterministic verification module, a hash binding module, a publishing area, and a feedback reproduction module. The candidate knowledge module corresponds to the candidate area, the independent review module and the deterministic verification module correspond to the review area, and the publishing area is the formal knowledge storage area. These three areas are isolated from each other, and there is no direct data path between the candidate area and the publishing area.

[0025] The method of the present invention includes the following steps: Step S101: Obtain the target knowledge scope. The target knowledge scope includes at least one sub-target, and each sub-target corresponds to a knowledge topic that needs to be covered by formal knowledge.

[0026] In step S102, the access control module calculates the coverage satisfaction of the publishing area with the target knowledge range based on the existing knowledge packages in the publishing area, and determines the coverage gap based on the coverage satisfaction.

[0027] The calculation process for the coverage gap includes: for each sub-objective in the target knowledge scope G. Set importance weights The sum of the importance weights of all sub-targets is 1, and each weight is non-negative; calculate the importance of the publishing area for each sub-target. Coverage satisfaction The coverage satisfaction value ranges from 0 to 1, where 0 indicates no coverage and 1 indicates complete coverage; the importance weight of each sub-objective is assigned. With corresponding coverage satisfaction After multiplying and summing, we obtain the coverage satisfaction degree of the publishing area with the target knowledge range G. Subtract the coverage satisfaction from 1. To obtain the coverage gap The calculation formula is: Equation (1) Step S103: Determine whether the coverage gap is greater than or equal to a preset gap threshold. In some embodiments, the preset gap threshold is 0.3. When the coverage gap is greater than or equal to the preset gap threshold, proceed to step S104; when the coverage gap is less than the preset gap threshold, it indicates that the publishing area can meet the coverage requirements of the target knowledge scope, and no data creation is required.

[0028] Step S104: Create data requirements corresponding to the target knowledge scope and block the formal output of the target knowledge scope based on the publishing area. The data requirements recorded in the data requirements module include the target scope, missing knowledge points, minimum source requirements, acceptable uses, and prohibitions. Among them, the minimum source requirements specify the number and type of sources required for candidate knowledge; acceptable uses specify the application scenarios in which the knowledge can be used; and prohibitions specify the content that the knowledge should not contain or the scenarios in which it should not be used.

[0029] In step S105, the query combination module generates various types of query statements based on data requirements. These query types include core queries, synonym queries, counterexample queries, document queries, and site queries. Core queries are generated directly targeting missing knowledge points; synonym queries use synonyms or related terms to expand the search scope; counterexample queries are used to retrieve materials that contradict candidate knowledge to verify boundaries; document queries limit file types (such as PDF, DOCX); and site queries limit specific websites or domains.

[0030] Step S106: The snapshot parsing module collects data from multiple sources based on the query statement, saves the original snapshots of the collected data, and parses the original snapshots to obtain parsed documents. Multiple sources include public web pages, user-uploaded data, and internal documents. Saving the original snapshots allows subsequent candidate processing and review to replay the original data, ensuring traceability. The parsed documents convert the original data into a structured format, facilitating subsequent deduplication and extraction.

[0031] In step S107, the deduplication and conflict removal module performs deduplication processing and conflict recording on the parsed documents based on normalized URLs, content hashes, similarity, and version identifiers. Normalized URLs unify different URL formats for the same webpage; content hashes identify identical documents by calculating the hash value of the document content; similarity uses a text similarity algorithm to identify highly similar but not identical documents; and version identifiers distinguish different versions of the same material. For conflicting materials, the conflict content and source are recorded for subsequent review. The deduplication and conflict removal module also calculates the duplication rate and conflict rate. The duplication rate is the proportion of duplicate sources to the total number of collected sources, used to prevent false coverage; the conflict rate is the proportion of conflicting candidate knowledge to the total number of candidate knowledge, used to trigger review or block access.

[0032] In step S108, the candidate knowledge module writes the deduplicated extraction results into the candidate area as candidate knowledge. The candidate area can only be read by the review process and is not allowed to be directly read by the formal application module, thus ensuring that the candidate knowledge cannot bypass the review process and directly affect the formal output.

[0033] Step S109: The independent review module, with a review subject independent of the candidate generation subject, reviews the candidate knowledge in the candidate area. The review process includes: confirming that the review subject and the candidate generation subject are different subjects, generating a review subject separation result S, where S takes the value 0 or 1, S=1 indicating that the review subject and the candidate generation subject are different subjects, and S=0 indicating that they are the same subject; the review subject examines the accuracy of the candidate knowledge content, the reliability of its source, and the completeness of its boundary statements; and outputs a review conclusion of pass, rejection, or need for supplementation. When the review conclusion is "needs supplementation," a supplementary data requirement is generated and the process returns to the data collection step (step S105), where the query combination module generates a new query statement based on the supplementary requirement and collects data. When the review conclusion is "rejection," the candidate knowledge does not proceed to the next step.

[0034] Step S110: The deterministic verification module performs deterministic verification on candidate knowledge whose review conclusion is passed. Deterministic verification includes mandatory field checks, boundary declaration checks, duplication and conflict checks, golden example checks, and publication scope checks. Mandatory field checks confirm that the candidate knowledge contains all required fields; boundary declaration checks confirm that the candidate knowledge clearly indicates its applicable and inapplicable scope; duplication and conflict checks confirm that the candidate knowledge does not have any unresolved duplications or conflicts with existing formal knowledge; golden example checks compare the candidate knowledge with preset standard examples to confirm that the format and content conform to the specifications; and publication scope checks confirm that the publication scope of the candidate knowledge is consistent with the acceptable uses in the data requirements. In some embodiments, the pass rate threshold for deterministic verification is 0.9.

[0035] In step S111, the hash binding module performs hash binding on the candidate knowledge that has passed the deterministic verification, binding the candidate content, review report, verification result, and release package together to form a release consistency certificate. The release consistency certificate includes the hash values ​​of the candidate content, review report, verification result, and release package, as well as a combined hash value. Through hash binding, any tampering with the content of any party will result in a mismatch in the combined hash value, thus detecting whether the released content has been replaced. The hash binding result H takes the value 0 or 1, where H=1 indicates successful hash binding, and H=0 indicates no binding or binding failure.

[0036] In step S112, in some embodiments, after hash binding and before writing to the publishing area, the publishing credibility score of the candidate knowledge is calculated, and it is determined whether to allow entry into the publishing area based on the publishing credibility score.

[0037] The credibility score is calculated by weighting the evidence chain health score, usefulness score, difference score, and conflict score. The calculation formula is as follows: Equation (3) in, Assess the credibility of the publication of candidate knowledge object e. To score the health of the chain of evidence, For usefulness rating, Score the difference. Rate the level of conflict. , , , In some embodiments, to preset weighting coefficients, =0.4, =0.3, =0.2, =0.1.

[0038] The evidence chain health score is calculated by testing the source accessibility, hash consistency, review completeness, authorization timeliness, and dissemination completeness of candidate knowledge, respectively, to obtain the missing rate for each of the five dimensions. , , , , Subtracting the average missing rate of the five dimensions from 1 yields the evidence chain health score, calculated using the following formula: Equation (2) The detection methods include: Source Accessibility (checking if the original data source is still accessible); Hash Consistency (checking if the hash value of the candidate content matches that of the original snapshot); Review Integrity (checking if the review report is complete); Authorization Timeliness (checking if the authorization of the data is within its validity period (the preset number of days for authorization timeliness is 180 days in some embodiments); and Propagation Integrity (checking if the knowledge is complete and has not been tampered with during the propagation process). The missing rate for each dimension ranges from 0 to 1, where 0 indicates that the dimension is completely normal and 1 indicates that the dimension is completely missing.

[0039] When the publication credibility score is greater than or equal to a preset publication threshold, candidate knowledge is allowed to enter the publication area; when the publication credibility score is less than the preset publication threshold, candidate knowledge is rejected from entering the publication area and a revision task is generated. In some embodiments, the preset publication threshold is 0.5.

[0040] Step S113: Write the candidate knowledge that has completed hash binding (and in the embodiment where the published trust score is set to meet the standard) as formal knowledge into the publication area.

[0041] Step S114: The feedback reproduction module collects the reasons for failures, user corrections, coverage gaps, and expiration events of formal knowledge during the application process, and transforms them into the next round of data requirements, revision tasks, or retirement tasks.

[0042] The transformation process of the feedback reproduction module includes: when the application fails or the user's error indicates that a certain sub-target lacks knowledge coverage, the next round of data requirements for that sub-target is generated, and the process returns to step S104 to start a new round of knowledge production; when the application fails or the user's error indicates that existing formal knowledge contains errors, a revision task for that formal knowledge is generated, triggering a re-examination and update of that formal knowledge; when it is detected that formal knowledge has expired or its source has become invalid, a retirement task for that formal knowledge is generated, removing that formal knowledge from the publication area or marking it as invalid.

[0043] like Figure 1 and Figure 2 As shown, in some embodiments, after writing the candidate knowledge with completed hash binding as formal knowledge into the publication area, the method further includes performing evidence chain health checks and repairs on the formal knowledge in the publication area. Evidence chain health checks and repairs include three layers: a break detection layer, a compensation repair layer, and a repair entry layer.

[0044] The evidence chain includes the release package, verified memories, context projections, and the released content.

[0045] The fracture detection layer performs the following checks in sequence: source reachability check, checking whether the original source of the data is still accessible; hash consistency check, checking whether the hash value of the formal knowledge content is consistent with that at the time of publication; review integrity check, checking whether the review report is complete and has not been tampered with; authorization validity check, checking whether the authorization of the data is still valid; and dissemination integrity check, checking whether the knowledge has been completely intact and tampered with during the dissemination process.

[0046] When all five tests above are passed, it indicates that the chain of evidence is intact, and the health score is updated. When a break in the chain of evidence is detected, a remediation plan is generated, and it is determined whether the break can be supplemented with additional evidence.

[0047] The compensation and repair layer performs different repair operations based on the judgment result: when the break can be re-certified, it performs re-certification collection and re-examination, re-collects the missing evidence and reviews it, and then rebuilds the hash binding; when the break cannot be re-certified but can be replaced, it generates a replacement evidence block and replaces the invalid evidence with new evidence; when the break is irrecoverable, it performs active retirement, removing the formal knowledge from the publication area or marking it as invalid.

[0048] The repair accounting layer records the repair results as repair records, propagates the impact of the repair to relevant knowledge objects, and updates the health score. The health score is used to assess the overall credibility of formal knowledge in the publishing area and can serve as a reference factor for the coverage access control module to calculate coverage satisfaction.

[0049] The present invention will be further illustrated by a specific embodiment below.

[0050] The following is an illustrative example, where the numerical values ​​are exemplary and do not constitute a limitation on actual parameters.

[0051] A certain intelligent knowledge system receives a target to generate a formal knowledge package around a certain technical topic. The target knowledge scope includes 5 sub-targets. The coverage access control module finds that the coverage satisfaction of the publishing area for the target knowledge scope is only 0.28, with a coverage gap of 0.72, which exceeds the preset gap threshold of 0.3. The system then creates a data request and blocks the formal output.

[0052] The data requirements module records the target scope, missing knowledge points, minimum source requirements (at least 3 independent sources), acceptable uses, and prohibitions. The query combination module generates core queries, synonym queries, counterexample queries, document queries, and site queries based on the data requirements. The snapshot parsing module collects public web pages, user-uploaded materials, and internal documents, saves the original snapshots, and parses them into candidate knowledge. The deduplication and conflict removal module merges duplicate sources and records conflicts based on normalized URLs, content hashes, similarity, and version identifiers.

[0053] Candidate knowledge is written into the candidate area. An independent review module, operated by a different entity than the candidate generator, reviews the content to confirm accuracy, source reliability, and the completeness of boundary declarations before outputting a pass conclusion. A deterministic validation module checks required fields, boundary declarations, duplicate conflicts, golden examples, and release scope, achieving a pass rate of over 0.9%. A hash binding module binds the candidate content, review report, validation results, and release package for consistency.

[0054] In the calculation of the credibility score, the health score of the chain of evidence is published. The usefulness score after five-dimensional analysis is 0.92. The difference score was 0.85. The conflict score is 0.78. The weighted score is 0.1, and a credibility score is published after weighted calculation. If the value exceeds the preset publication threshold of 0.5, the candidate knowledge is allowed to enter the publication area.

[0055] In subsequent applications, if a user points out that a certain sub-goal still lacks counterexamples, the feedback reproduction module records the effect evidence and generates the next round of data requirements for that sub-goal, thus starting a new round of knowledge production.

[0056] After running for a period of time, the evidence chain health detection and repair module detects the formal knowledge in the publishing area and finds that the source of a certain formal knowledge is no longer accessible, indicating a break. The system determines that the break can be re-proven, performs supplementary evidence collection and re-examination, finds an alternative source and rebuilds the hash binding, records the repair results, and updates the health score.

[0057] This invention enables knowledge production to transform from a one-time retrieval into a complete process that is auditable, publishable, and feedback-enabled. It reduces the risk of low-reliability data directly entering the formal output and reduces duplicate collection and false coverage by quantifying coverage gaps.

[0058] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. 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 flywheel method for trusted knowledge production driven by coverage gaps, characterized in that, include: Acquire the target knowledge scope, which includes at least one sub-target; Based on the existing knowledge packages in the publishing area, calculate the coverage satisfaction degree of the publishing area for the target knowledge range, and determine the coverage gap based on the coverage satisfaction degree; When the coverage gap is greater than or equal to a preset gap threshold, a data requirement corresponding to the target knowledge scope is created, and the formal output of the target knowledge scope based on the publishing area is blocked; Based on the data requirements, generate various types of query statements, collect data from multiple sources based on the query statements, save the original snapshots of the collected data, and parse the original snapshots to obtain parsed documents; The parsed document is deduplicated and conflict recorded based on the normalized URL, content hash, similarity, and version identifier. The deduplicated extraction results are written into the candidate area as candidate knowledge. The candidate area can only be read by the review process and cannot be directly read by the formal application module. The review body, which is independent of the candidate generation body, reviews the candidate knowledge in the candidate area and outputs the review conclusion of pass, rejection, or need to be supplemented. Deterministic validation is performed on candidate knowledge whose review conclusion is passed. The deterministic validation includes mandatory field checks, boundary declaration checks, duplicate conflict checks, golden example checks, and publication scope checks. Perform hash binding on candidate knowledge that passes deterministic verification, and bind the candidate content, review report, verification result and release package in a consistent manner to form a release consistency proof; The candidate knowledge that has completed hash binding is written as formal knowledge into the publication area; and Collect the reasons for failures, user corrections, coverage gaps, and expiration events of the formal knowledge during application, and transform them into the next round of data requirements, revision tasks, or retirement tasks.

2. The method according to claim 1, characterized in that, The step of calculating the coverage satisfaction of the publishing area for the target knowledge scope based on the existing knowledge packages in the publishing area, and determining the coverage gap based on the coverage satisfaction, includes: setting an importance weight for each sub-target in the target knowledge scope, the sum of the importance weights of all sub-targets being 1 and each weight being non-negative; calculating the coverage satisfaction of the publishing area for each sub-target, the coverage satisfaction ranging from 0 to 1; multiplying the importance weight of each sub-target by the corresponding coverage satisfaction and summing the results to obtain the coverage satisfaction of the publishing area for the target knowledge scope; and subtracting the coverage satisfaction from 1 to obtain the coverage gap.

3. The method according to claim 1, characterized in that, The data requirements include the target scope, missing knowledge points, minimum source requirements, acceptable uses, and prohibitions; wherein, the minimum source requirements specify the number and type of sources required for the candidate knowledge, the acceptable uses specify the application scenarios in which the knowledge can be used, and the prohibitions specify the content that the knowledge should not contain or the scenarios in which it should not be used.

4. The method according to claim 1, characterized in that, The various query types include core queries, synonym queries, counterexample queries, file queries, and site queries. The parsed documents are deduplicated and conflict recorded based on normalized URLs, content hashes, similarity, and version identifiers. This includes: unifying different URL formats of the same webpage using normalized URLs; identifying completely identical documents using content hashes; identifying highly similar but not completely identical documents using similarity; and distinguishing different versions of the same material using version identifiers. For conflicting materials, the conflict content and source are recorded. The duplication rate and conflict rate are calculated, where the duplication rate is the proportion of duplicate sources to total collected sources, and the conflict rate is the proportion of conflicting candidate knowledge to total candidate knowledge.

5. The method according to claim 1, characterized in that, The review of candidate knowledge in the candidate area by a reviewing entity independent of the candidate generation entity includes: confirming that the reviewing entity and the candidate generation entity are different entities; the reviewing entity examining the accuracy of the content, reliability of the source, and completeness of the boundary statements of the candidate knowledge; and outputting a review conclusion of pass, rejection, or need for supplementation. When the review conclusion is that supplementation is needed, a supplementary data requirement is generated and the process returns to the data collection step. When the review conclusion is rejection, the candidate knowledge does not enter the subsequent process.

6. The method according to claim 1, characterized in that, After performing hash binding on the candidate knowledge that has passed the deterministic verification, the method further includes: calculating the publication credibility score of the candidate knowledge, wherein the publication credibility score is calculated by weighting the evidence chain health score, usefulness score, difference score, and conflict score, and the calculation formula is as follows: in, Assess the credibility of the publication of candidate knowledge object e. To score the health of the chain of evidence, For usefulness rating, Score the difference. Rate the level of conflict. The system uses a preset weighting coefficient; when the publication credibility score is greater than or equal to a preset publication threshold, the candidate knowledge is allowed to enter the publication area; and when the publication credibility score is less than the preset publication threshold, the candidate knowledge is rejected from entering the publication area and a revision task is generated.

7. The method according to claim 6, characterized in that, The health score of the chain of evidence is calculated in the following way: The missing rates for the five dimensions of candidate knowledge were obtained by testing the source accessibility, hash consistency, review completeness, authorization timeliness, and propagation completeness. ; The evidence chain health score is obtained by subtracting the average missing rate of the five dimensions from 1, and the calculation formula is as follows: in, Source accessibility checks whether the original data source is still accessible; hash consistency checks whether the hash value of the candidate content is consistent with that of the original snapshot; review integrity checks whether the review report is complete; authorization timeliness checks whether the authorization of the data is within the validity period; and dissemination integrity checks whether the knowledge is complete and has not been tampered with during the dissemination process. The missing rate for each dimension ranges from 0 to 1.

8. A flywheel system for trusted knowledge production driven by coverage gaps, characterized in that, Methods for implementing rights 1 to 7 include: The coverage access control module is used to obtain the target knowledge range, calculate the coverage satisfaction of the publishing area with the target knowledge range based on the existing knowledge packages in the publishing area, and determine the coverage gap. When the coverage gap is greater than or equal to a preset gap threshold, a data requirement is created and the formal output is blocked. The data requirements module is used to record the data requirements, which include the target scope, missing knowledge points, minimum source requirements, acceptable uses, and prohibited statements. The query combination module is used to generate core queries, synonym queries, counterexample queries, file queries, and site queries based on the data requirements; the snapshot parsing module is used to collect data from multiple sources based on the queries, save the original snapshots, and parse the original snapshots to obtain parsed documents; The deduplication and conflict resolution module is used to perform deduplication and conflict recording on the parsed document based on the normalized URL, content hash, similarity and version identifier; The candidate knowledge module is used to write the deduplicated extraction results as candidate knowledge into the candidate area. The candidate area can only be read by the review process and cannot be directly read by the formal application module. An independent review module is used by a review subject independent of the candidate generation subject to review the candidate knowledge in the candidate area and output a review conclusion of pass, rejection or need to be supplemented. The deterministic validation module is used to perform deterministic validation on candidate knowledge whose review conclusion is passed. The deterministic validation includes mandatory field checks, boundary declaration checks, duplicate conflict checks, golden example checks, and publication scope checks. The hash binding module is used to perform hash binding on candidate knowledge that has passed deterministic verification, binding the candidate content, review report, verification result and release package to form a release consistency proof; The release area is used to store the formal knowledge of completing hash bindings; and The feedback reproduction module is used to collect the reasons for failure, user corrections, coverage gaps and expiration events of the formal knowledge during the application process, and transform them into the next round of data requirements, revision tasks or retirement tasks.

9. The system according to claim 8, characterized in that, The coverage access control module is further configured to: set an importance weight for each sub-target in the target knowledge scope, the sum of the importance weights of all sub-targets being 1 and each weight being non-negative; calculate the coverage satisfaction of the publishing area for each sub-target; multiply the importance weight of each sub-target by the corresponding coverage satisfaction and sum them to obtain the coverage satisfaction of the publishing area for the target knowledge scope; subtract the coverage satisfaction from 1 to obtain the coverage gap; the independent review module is further configured to: confirm that the review subject and the candidate generation subject are different subjects; have the review subject review the accuracy of the candidate knowledge content, the reliability of the source, and the completeness of the boundary declaration; output a review conclusion of pass, rejection, or need for supplementation, wherein, when the review conclusion is that supplementation is needed, a supplementary data requirement is generated and the query combination module is notified to re-collect data.

10. The system according to claim 8, characterized in that, It also includes a release license judgment module and an evidence chain health detection and repair module; the release license judgment module is used to: calculate the release credibility score of candidate knowledge, the release credibility score being calculated by weighting the evidence chain health score, usefulness score, difference score, and conflict score; when the release credibility score is greater than or equal to a preset release threshold, the candidate knowledge is allowed to enter the release area; when the release credibility score is less than the preset release threshold, the candidate knowledge is rejected from entering the release area and a revision task is generated; wherein, the evidence chain health score obtains the missing rate of the corresponding five dimensions by detecting the source accessibility, hash consistency, review integrity, authorization timeliness, and dissemination integrity of the candidate knowledge, and sets 1 The average of the missing rates of the five dimensions is subtracted to obtain the evidence chain health detection and repair module. This module is used to perform evidence chain health detection and repair on formal knowledge in the publishing area, including: reading the evidence chain of formal knowledge, which includes the publishing package, verified memory, context projection, and published content; sequentially performing source reachability detection, hash consistency detection, review integrity detection, authorization timeliness detection, and propagation integrity detection; when a break in the evidence chain is detected, a repair plan is generated; it is determined whether the break can be re-verified; if it can be re-verified, re-verification collection and re-examination are performed, and hash binding is rebuilt; if it cannot be re-verified but can be substituted, a substitute evidence block is generated; if it is unrecoverable, active retirement is performed; the repair results are recorded and the health score is updated.

Images