A Budget-Constrained Closed-Loop Retrieval Method Driven by Evidence Gap

By employing a budget-constrained closed-loop retrieval method driven by evidence gaps, this approach addresses the issues of excessively coarse granularity of retrieval targets and insufficient differentiation of evidence roles in existing technologies, achieving precise evidence coverage and accurate conclusions even under resource-constrained conditions.

CN123115216APending Publication Date: 2026-11-17SOUTHWEAT UNIV OF SCI & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611310179.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-27
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In knowledge-intensive question answering, technical analysis, and decision support tasks, existing technologies suffer from overly coarse granularity in retrieval targets, making it easy for homogeneous materials to be counted repeatedly. They also fail to effectively distinguish the roles of different pieces of evidence, and the stopping results cannot accurately reflect whether the evidence is sufficient.

Method used

A budget-constrained closed-loop retrieval method driven by the evidence gap is adopted. By splitting candidate conclusions into assertions to be verified, identifying evidence role types, calculating evidence gaps, and prioritizing the use of limited resources to fill key gaps, the method reduces duplicate recall.

Benefits of technology

It enables precise evidence coverage judgment for the arguments to be verified in candidate conclusions, reduces the impact of duplicate materials, ensures the accuracy and efficiency of the output conclusions, and can prioritize filling key evidence gaps when resources are limited.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115216A_ABST
    Figure CN123115216A_ABST
Patent Text Reader

Abstract

This invention discloses a budget-constrained closed-loop retrieval method driven by evidence gaps. The method first breaks down candidate conclusions into arguments to be verified, then determines the dynamic evidence role of candidate evidence relative to the target argument and completes admission verification. Evidence from the same source is grouped and aggregated before participating in coverage calculations. Evidence gaps, argument conflicts, unresolved feedback, and redundancy jointly determine scheduling priority. The system generates integer quotas under constraints of basic guarantee quotas, unit limits, and a single-round total budget. It updates the cross-round evidence ledger and argument evidence status with newly acquired evidence, thereby distinguishing between sufficiency stopping, marginal benefit stopping, and resource stopping. This invention achieves the identification of missing evidence roles for each argument to be verified in candidate conclusions under constraints of retrieval count, processing time, and generation context capacity. It converts evidence gaps into integer retrieval quotas subject to a unified total budget, prioritizing limited retrieval resources for key gaps and reducing the repeated recall of similar materials.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of computer information retrieval technology, and specifically to a budget-constrained closed-loop retrieval method driven by a gap in assertion evidence. Background Technology

[0002] In knowledge-intensive tasks such as question answering, technical analysis, research reviews, and decision support, computers typically first retrieve relevant content from external databases and then generate answers or reports based on this information. These tasks often involve multiple candidate relationships, different applicable conditions, and contradictory data. Therefore, the relevance of search results to the question does not equate to the fact that every specific claim in the candidate conclusions has been adequately supported by evidence.

[0003] One existing type of technology uses keyword retrieval, semantic vector retrieval, or a combination of both to retrieve a fixed number of documents or fragments from a database, which are then processed by a generation module. While this approach can supplement external knowledge, retrieval control remains centered on the relevance of the original question to the documents. It neither breaks down candidate conclusions into separately verifiable relationships nor continuously records the coverage of supporting, challenging, boundary, mechanistic, and methodological evidence roles.

[0004] Another type of technology alternates between reasoning and retrieval, determining whether to continue retrieval or how to rewrite the query based on the phased results generated. Existing reasoning content can thus participate in subsequent retrievals, but the controlled object is usually still the reasoning chain or text fragment, and a unified state for multiple arguments and various evidentiary roles has not yet been formed. When total resources are limited, it is also difficult to determine the number of retrievals that should be performed on each argument-evidence role channel.

[0005] Another type of technology filters external materials through recall quality assessment, anomaly identification, duplicate detection, and error correction mechanisms, and adds searches when there are content conflicts or insufficient effective information. This type of method can reduce the number of low-quality fragments entering the generation process, but the evaluation is usually based on the document set corresponding to a specific query. It fails to calculate the evidence coverage after deduplication of the source, the conflict between support and challenge, and the information added across rounds around the specific claim, nor does it further convert these states into integer search quotas that conform to a unified overall budget.

[0006] In summary, the main shortcomings of existing technologies are not the insufficient number of retrieval rounds, but rather the overly coarse granularity of retrieval targets, the tendency for homogeneous materials to be repeatedly counted, the lack of differentiation between the roles of different pieces of evidence, and the inability of the stopping results to accurately reflect the sufficiency of the evidence. To address these issues, this invention uses the assertion-evidence role unit as a closed-loop control object, incorporating evidence admission, source deduplication, gap calculation, budget allocation, role-based retrieval, and stopping judgment into the same processing chain. Summary of the Invention

[0007] This invention provides a budget-constrained closed-loop retrieval method driven by evidence gaps in assertions. Under the constraints of limited retrieval times, processing time, and generation context capacity, it identifies the evidence roles that are still lacking for each assertion to be verified in candidate conclusions, and converts the evidence gaps into integer retrieval quotas that conform to a unified total budget. This allows limited retrieval resources to be prioritized for key gaps, reducing the repeated recall of similar materials.

[0008] This invention provides a budget-constrained closed-loop retrieval method driven by a gap in assertive evidence, the method comprising: The problem to be analyzed is identified and candidate conclusions are formed. The candidate conclusions are then broken down into a set of arguments to be verified, and a set of evidence role types is established. Acquire candidate evidence related to each claim to be verified, determine the dynamic evidence role type of each candidate evidence relative to each claim to be verified, and obtain a set of valid evidence through evidence admission verification; Based on the set of valid evidence, all valid evidence is grouped according to its source to obtain different source groups. Within each group, the valid contribution of a single piece of evidence is calculated and the maximum contribution value within the group is selected as the contribution of the current source group. Based on the contributions of different source groups, the current valid coverage after deduplication of sources is calculated. The difference between the current effective coverage and the target coverage is calculated to obtain the evidence gap. The evidence gap, conflicting arguments, unresolved feedback and redundancy are considered to determine the scheduling priority in the next round of budget allocation. Based on resource budget, basic guarantee quota, and quota upper limit constraints, the number of retrievals and quota execution are determined according to scheduling priority. Based on a defined number of searches, a quota is generated to perform a search task and the search is executed. The candidate results obtained are sorted by diversity and compared across rounds of information to obtain a new set of candidate evidence and the evidence admission is re-verified. Determine whether the search termination condition is met. If the condition is met, further determine whether the evidence is sufficient and decide whether to output normally or in a downgraded manner based on the result. If the condition is not met, continue the search.

[0009] Furthermore, the problem to be analyzed is identified and candidate conclusions are formed. These candidate conclusions are then broken down into a set of arguments to be verified, and a set of evidence role types is established, specifically including: From the problem to be analyzed, candidate conclusions and original applicable conditions, identify the subject, relationship, object or target result and applicable conditions one by one. Relationships that can be independently judged as true or false and have consistent applicable conditions are retained as one statement. Sentences containing multiple relationships, multiple results or different applicable conditions are split into multiple statements. When a field is missing, the missing status is recorded and no facts are created. The applicable conditions of each assertion are standardized, and the object name, working condition terminology, unit of measurement, time range and negation relationship are unified. When the object range, working condition range or time range is unknown, an unknown mark is retained. Records with known ranges that do not intersect are marked as incompatible conditions.

[0010] Furthermore, the problem to be analyzed is identified and candidate conclusions are formed. These candidate conclusions are then broken down into a set of arguments to be verified, and a set of evidence role types is established, specifically including: The set of evidence role types is used to characterize the types of roles that candidate evidence can play relative to the target assertion. Evidence role types include support, challenge, boundary, mechanism, and method.

[0011] Furthermore, candidate evidence related to each claim to be verified is obtained, the dynamic evidentiary role type of each candidate evidence relative to each claim to be verified is determined, and a valid evidence set is obtained through evidence admission verification, specifically including: Obtain the original candidate data, extract the main text, source identifier, location information, applicable conditions and recall rounds, generate content fingerprints and bind them to the target assertion, normalize the document identifier, database primary key, locator and document fingerprint according to deterministic rules, retain the rejection reason for the data whose main text or source cannot be located and do not enter the valid evidence set; Normalized features are extracted from relational direction, conclusion polarity, condition compatibility, action path and verification purpose. The scores of candidate evidence for each evidence role type are calculated. The evidence role type with the highest score is selected as the dynamic evidence role of the corresponding candidate evidence relative to the target assertion. The same evidence has different dynamic evidence roles relative to different assertions. The criteria for admission of candidate evidence are verified sequentially, including: whether the source is locatable, whether it is directly related to the target assertion, whether the dynamic role determination is credible, whether the applicable conditions are compatible, whether the content is completely duplicated, and whether the purpose of citation is consistent. If all conditions are met, the admission result is set to pass and included in the set of valid evidence. If any condition is not met, the evidence is rejected and the reason for rejection is recorded.

[0012] Furthermore, based on the set of valid evidence, all valid evidence is grouped according to its source to obtain different source groups. Within each group, the valid contribution of a single piece of evidence is calculated, and the maximum contribution value within the group is selected as the contribution of the current source group. The current valid coverage after deduplication is calculated based on the contributions of different source groups. A unified source identifier is generated based on document identifier, database primary key, locator, document fingerprint, or a verifiable combination of fields. Valid evidence with the same source identifier is grouped into the same source group, and when there is a conflict in the identifier, it enters a pending review state. For each piece of valid evidence, the contribution of that evidence to the target assertion-evidence role unit is calculated based on the admission result, relevance, dynamic role score, source quality, condition compatibility, and degree of newness of historical information. The contribution value is reduced accordingly if any key condition is insufficient, as shown in the following formula:

[0013] in, Representing the Wheel of Life Candidate evidence for the first The first unverified assertion A single valid contribution formed by a class of evidentiary roles; The evidence admission indicator is used to determine whether candidate evidence is included in the coverage calculation. It is set to one if all necessary checks are passed, and zero otherwise. Representing the Candidate evidence and the first The correlation between claims to be verified is used to determine whether the evidence directly relates to the relationship to be verified. Representing the The candidate evidence is relative to the first Each unverified assertion bears the responsibility of the first The confidence level of a role in dynamic evidence is used to measure the credibility of the role's judgment. Representing the The quality of the source of each candidate piece of evidence is used to evaluate the completeness, locatability, and reliability of the source; Representing the Candidate evidence and the first The compatibility of conditions between two unverified assertions is used to determine whether the applicable conditions of the two assertions are consistent or can be true at the same time. Representing the Wheel of Life The degree of newness of historical information in each candidate piece of evidence is used to measure its addition of sources, facts, or applicable conditions relative to the cross-cycle evidence ledger; Within each source group, the maximum effective contribution of a single piece of evidence for the same assertion-evidence role unit is selected as the contribution value for the corresponding source group, as shown in the following formula:

[0014] in, Representing the Wheel of Life The source group is the first The first unverified assertion The contribution of the source group formed by the role of evidence is used to indicate the strength of valid evidence that can be provided by the same source; The source group index is used to group abstracts, text fragments, adjacent paragraphs, or duplicate copies from the same source into the same group. The maximum value operation is used to select the highest effective contribution from multiple candidate evidences from the same source group, so that the same material will not be amplified and covered by the increase in the number of fragments after repeated segmentation. The contributions from different source groups are then summed using the following formula to obtain the effective coverage after deduplication:

[0015]

[0016] in, Represents the valid contribution of a single piece of evidence, used to indicate the contribution of a candidate piece of evidence after verification of admission, relevance, role, source, conditions, and new information; The representative source group contribution is used to indicate the contribution formed after deduplication and aggregation within the same source group. Representing the Wheel of Life The first unverified assertion The source deduplication and effective coverage achieved by the evidence-like role are used to accumulate the contributions of multiple independent source groups; Representatives participated in the One unverified assertion and the first The evidence role covers the set of source groups for computation, which is used to limit the source range of the product operation; Representing the The complementary coverage portions not yet provided by each source group are used to ensure that the marginal contribution of new sources decreases as existing coverage increases.

[0017] Furthermore, the difference between the current effective coverage and the target coverage is calculated to obtain the evidence gap. Taking into account the evidence gap, conflicting arguments, unresolved feedback, and redundancy, the scheduling priority in the next round of budget allocation is determined, specifically including: The unfilled portion between the target coverage and the current effective coverage is calculated using the following formula, and the evidence gap is obtained through non-negative truncation:

[0018] in, This represents an evidence gap, indicating the portion of the target coverage that has not yet been filled in between the current effective coverage. The target coverage is used to specify the coverage requirements that the corresponding assertion-evidence role units need to achieve; This represents the current effective coverage, indicating the extent to which the corresponding assertion-evidence role unit has been covered by the independent source group; This represents a non-negative truncation operation, used to ensure that the evidence gap is not less than zero; a positive gap is only formed when the target coverage is higher than the current effective coverage; cell gaps that have reached or exceeded the target coverage are treated as zero; For each assertion-evidence role unit, the scheduling priority is calculated using the following formula;

[0019] in, The scheduling priority of the target assertion-evidence role unit in the next round of budget allocation is used to measure the relative urgency of the corresponding unit; Representing evidence gaps, used to prioritize state units that have not yet achieved target coverage; Represents the weight of the evidence gap, used to adjust for the impact of the evidence gap; Representative arguments conflict, used to increase the priority of search channels that require challenge for clarification or boundary confirmation; The value represents the evidence role indicator. It is set to one when the evidence role is a challenge or boundary type, or when the corresponding unit undertakes the conflict clarification task; otherwise, it is set to zero. The weighting of conflicting arguments is used to moderate the impact of conflict states. This represents unresolved feedback and is used to prioritize units with issues awaiting review, unclosed feedback, or manual verification requirements. This represents the weight of unresolved feedback, used to adjust the impact of unresolved feedback. Represents redundancy, used to reduce the priority of units repeatedly occupied by duplicate material, homologous fragments, or known facts; This represents the redundancy reduction weight, used to adjust the reduction intensity of redundancy states; This represents the non-negative truncation operation, used to set the scheduling priority to zero when the weighted result is less than zero; , , and These represent four types of non-negative weights, which respectively adjust for the effects of evidence gaps, conflicting arguments, unresolved feedback, and redundancy.

[0020] Furthermore, based on resource budget, basic guarantee quota, and quota upper limit constraints, the number of retrievals is determined according to scheduling priority, specifically including:

[0021] in, Represents the allocation of the next round to the first One unverified assertion and the first The final execution quota for evidence-like roles is used to directly control the number of executions for the corresponding retrieval channels; The basic guarantee quota is used to ensure that essential evidence for key arguments has the minimum chance of being searched. Units that have already achieved target coverage, are not essential evidence, or are prohibited from being searched further have their basic guarantee quota set to zero. This represents the incremental quota, used to allocate the remaining budget based on scheduling priority. This represents a non-negativity constraint, used to ensure that no assertion-evidence role unit receives a negative quota; The upper limit constraint for the representative unit is used to ensure that the final execution quota does not exceed the available upper limit of the corresponding state unit. ; Represents the total budget constraint, used to ensure that the sum of all final execution quotas does not exceed the threshold. Total budget for a single round The total budget for a single round is measured by the number of evidence items allowed to be accepted in this round, or calculated based on the number of search requests, the number of read fragments, the number of context tokens, or the processing time.

[0022] Furthermore, based on resource budget, basic guarantee quota, and quota upper limit constraints, the number of retrievals is determined according to scheduling priority, specifically including: After deducting the basic guarantee quota, the incremental quota is determined using the following formula:

[0023] in, The optimization results, representing all incremental quotas in the next round, are used to determine the additional number of searches for each assertion-evidence role unit beyond the basic guaranteed quota. Representative assigned to the first One unverified assertion and the first Incremental quotas for evidence-based roles; Represents the scheduling priority of the current round, used to measure the necessity of increasing the number of retrieval slots to the corresponding unit; This represents logarithmic marginal returns, used to gradually decrease the returns when the quota is continuously increased for the same state unit. Represents the overall scheduling benefit of all state units, used to select the combination of integer quotas with higher benefits under the current budget constraints; This represents the total budget constraint, used to ensure that the sum of all incremental quotas does not exceed the remaining budget after deducting the basic guarantee quota; The upper limit constraint represents the unit limit, which is used to ensure that the incremental quota of each unit does not exceed the difference between the quota limit and the basic guaranteed quota. Represents an integer field constraint, allowing the calculation result to be directly converted into the number of times the retrieval task is executed.

[0024] Furthermore, based on a predetermined number of searches, a quota is used to generate a search task and execute the search. The retrieved candidate results are then sorted by diversity and compared across rounds of information to obtain a new set of candidate evidence. This is followed by a re-verification of the evidence's eligibility. Specifically, this includes: For each assertion-evidence role unit, generate the same number of retrieval tasks as the execution quota. Support tasks emphasize direct support relationships, challenge tasks emphasize counterexamples or negation relationships, boundary tasks emphasize applicable objects, working conditions and time ranges, mechanism tasks emphasize action paths, and method tasks emphasize measurement or verification methods. The assertion identifier, role identifier, focus conditions and quota sources in the tasks are all copied from the input objects. Perform a search and calculate candidate ranking scores based on query relevance, assertion matching degree, role matching degree, degree of historical newness, source independence, and redundancy within the same round. Among these, the degree of historical newness and source independence have a positive impact on the score, while redundancy within the same round has a negative impact on the score. Perform deduplication within the same round and filter candidate evidence according to the ranking scores. The candidate evidence is compared with the records in the historical cross-cycle evidence ledger by content fingerprint, source group, fact unit and applicable conditions. The normalized degree of historical new information is calculated. When the candidate evidence and the historical record are from the same source and the fact unit and applicable conditions are the same, the degree of new information is zero. When there are verifiable differences, the corresponding degree of new information is retained. After the comparison is completed, the evidence status, source group, target assertion, dynamic role and contribution relationship are written into the updated cross-cycle evidence ledger. The new candidate evidence set obtained after sorting and cross-round comparison will be returned to re-execute dynamic role determination and evidence admission verification. Content that has been accepted, rejected or completely duplicated will be retained in its historical state and will not be directly included in the current valid evidence set.

[0025] Further, it is determined whether the search termination condition is met. If the condition is met, the sufficiency of the evidence is further assessed, and the output is determined as either normal or downgraded based on the assessment result. If the condition is not met, the search continues, specifically including: Calculate the average coverage change of all valid assertion-evidence role units, and simultaneously count the number of newly added independent sources, the number of valid new evidence, and the effective gain per unit budget. The average coverage change is normalized according to the cardinality of the set of all valid assertions and the set of roles involved in the calculation. The sufficiency indicator, the continuous low gain indicator, and the budget exhaustion indicator are calculated separately. The sufficiency indicator is determined by whether all necessary roles have achieved the target coverage and whether major conflicts have been resolved. The continuous low gain indicator is determined by the average coverage change, the number of new independent sources, and the number of consecutive low gain rounds, and the threshold is calibrated by historical rounds or manually verified samples. The budget exhaustion indicator is determined by comparing the cumulative resource consumption with the cumulative budget. If the sufficiency indicator is true, a sufficiency stop is triggered and normal output is executed; if the sufficiency indicator is false but the continuous low gain indicator or the budget exhaustion indicator is true, a degraded output is triggered, and the output object clearly lists the existing evidence, missing roles, unresolved conflicts and subsequent verification directions; if all three indicators are false, a continue search instruction is generated and the next round of search is started.

[0026] This invention provides a budget-constrained closed-loop retrieval method driven by gaps in assertive evidence, which has the following beneficial effects: 1. Compared to methods that determine retrieval completion based on the frequency of keyword occurrences or the number of recalled documents, this invention applies coverage judgment to specific claims to be verified and further distinguishes between supporting, challenging, boundary, mechanistic, and methodological evidence roles. Only materials that can clearly contribute to the target claim enter the corresponding state, thus reducing the interference of background information involving only relevant variables on coverage judgment and output conclusions.

[0027] 2. Evidence contributions are first aggregated within the source group, and then accumulated across independent source groups; the degree of newness of historical information is used to identify materials that have been accepted, rejected, and completely duplicated. When the same source is segmented into multiple fragments, the coverage is not linearly amplified, and a new round of retrieval does not require repeated use of budget and context processing of known content.

[0028] 3. Evidence gaps, conflicting arguments, and unresolved feedback constitute the basis for input, while redundant states constitute the basis for deduction; integer quotas are simultaneously constrained by basic guarantees, unit limits, and total budget. This approach not only reserves a minimum retrieval opportunity for essential evidence for key arguments but also limits the continuous resource consumption of a single high-similarity channel and allows state calculation results to be directly converted into retrieval tasks.

[0029] 4. The dynamic role of evidence and the conflict-driven scheduling method ensure that challenge, boundary, mechanism, and methodological evidence can still be retrieved even when there is abundant supporting material. When support and challenges coexist, the system further investigates differences in conditions, scope of application, and action paths, without replacing conflict resolution with simple majority voting. Therefore, it can clarify the boundaries of the argument and retain the verification value of reverse evidence.

[0030] 5. This method separately judges "evidence meets requirements" and "failure to obtain new information by continuing the search". When the search ends due to continuous low gain, service abnormality or budget exhaustion, and the evidence is still insufficient, the system generates a downgraded result, retaining existing evidence, unresolved conflicts, missing roles and subsequent verification paths, to avoid misclassifying resource exhaustion as sufficient evidence.

[0031] 6. The cross-round evidence ledger continuously records assertions, dynamic evidence roles, source groups, admission results, recall rounds, status contributions, and citation purposes, enabling output citations to be traced back and their evidentiary role verified. The above effects do not depend on a specific retrieval model or fixed data size and can be verified through key assertion role completeness rate, independent source growth rate, cross-round duplication rate, effective gain per unit budget, citation role consistency rate, and the downgraded accuracy rate of assertions with insufficient evidence. Attached Figure Description

[0032] Figure 1 The overall flowchart of a budget-constrained closed-loop retrieval method driven by an assertion evidence gap is provided as an embodiment of the present invention; Figure 2 A diagram showing the connection relationship between functional modules and shared states in a budget-constrained closed-loop retrieval method driven by an assertion evidence gap, provided as an embodiment of the present invention; Figure 3 A schematic diagram illustrating the assertion construction and dynamic evidence role determination process in an assertion evidence gap-driven budget-constrained closed-loop retrieval method provided in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the evidence admission, source grouping, and coverage update process in a budget-constrained closed-loop retrieval method driven by an evidence gap provided in an embodiment of the present invention. Figure 5 A schematic diagram of the gap priority calculation and budget allocation process in a budget-constrained closed-loop retrieval method driven by gaps in assertion evidence provided in an embodiment of the present invention; Figure 6 A schematic diagram of role-conditional retrieval and cross-round deduplication process in a budget-constrained closed-loop retrieval method driven by a gap in assertion evidence provided in an embodiment of the present invention; Figure 7 A schematic diagram of the stopping, degraded output, and tracing process in a budget-constrained closed-loop retrieval method driven by an assertion evidence gap provided in an embodiment of the present invention; Figure 8 The flowchart illustrates a specific implementation of a budget-constrained closed-loop retrieval method driven by an evidence gap, as provided in one embodiment of the present invention. Detailed Implementation

[0033] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings. Similar elements in different embodiments are referred to by associated similar element reference numerals. In the following embodiments, many details are described to facilitate a better understanding of the invention. However, those skilled in the art will readily recognize that some features may be omitted in different situations, or may be replaced by other elements, materials, or methods. In some cases, certain operations related to the present invention are not shown or described in the specification. This is to avoid obscuring the core parts of the invention with excessive description. For those skilled in the art, detailed description of these related operations is not necessary; they can fully understand the related operations based on the description in the specification and general technical knowledge in the art.

[0034] Furthermore, the features, operations, or characteristics described in the specification can be combined in any suitable manner to form various embodiments. At the same time, the steps or actions in the method description can be rearranged or adjusted in a manner obvious to those skilled in the art. Therefore, the various orders in the specification and drawings are only for the clear description of a particular embodiment and do not imply a necessary order, unless otherwise stated that a particular order must be followed.

[0035] Terminology Explanation: The object of verification is the smallest relational object extracted from candidate conclusions, possessing independent identification and requiring separate evidence. The set of necessary evidence roles is a set of supporting, challenging, boundary, mechanism, or methodological roles determined based on the importance of the assertion, output risk, and verification needs. Dynamic evidence roles are the actual evidentiary function determined by combining the evidence text, the target assertion, and applicable conditions.

[0036] The Admission Result object records whether candidate evidence has entered the valid evidence set, the conditions for rejection, the reason for rejection, and the pending review status. The Source Group object is a set of evidence from the same source grouped by document identifier, locator, database primary key, document fingerprint, or verifiable combination of information. The Assumption-Evidence Role Status object records target coverage, current coverage, source group distribution, conflicts, redundancy, unresolved feedback, and update rounds. The Evidence Gap object records the non-negative difference between the target coverage and the current coverage.

[0037] Integer quota objects are non-negative integer retrieval quantities and their rationales determined for a specific assertion-evidence role unit under the constraints of the single-round total budget, basic guarantee rules, and unit upper limit. Cross-round evidence ledger objects record evidence sources, target assertions, dynamic roles, admission results, repetition states, and state contributions across rounds. Sufficiency stopping is the stopping method when necessary roles have achieved target coverage and major conflicts have been resolved; continuous low-gain stopping is the stopping method when multiple rounds of coverage changes and newly added independent sources fall below the corresponding threshold; degraded output is the output method where, at the time of stopping, the evidence is still insufficient, but existing evidence, conflicts, gaps, and subsequent verification paths are retained.

[0038] The first embodiment of this invention provides a budget-constrained closed-loop retrieval method driven by an evidence gap, executed by one or more computing devices. The computing devices are connected to a local document repository, a structured database, an external retrieval service, or a combination thereof, and are constrained by one or more resource conditions, including single-round budget, cumulative budget, context capacity, and processing time limit. This method does not judge the sufficiency of evidence based on the number of similar documents recalled for the original question; instead, it represents the evidence gap, the number of documents to be retrieved in the next round, and the stopping condition as computable and up-to-date control states.

[0039] This method first identifies the problem to be analyzed and forms candidate conclusions. These candidate conclusions are then broken down into assertions to be verified. Next, candidate evidence related to each assertion is acquired, completing dynamic evidence role determination and admission verification. Subsequently, the assertion-evidence role status is calculated based on accepted evidence. Scheduling priorities are determined by evidence gaps, assertion conflicts, unresolved feedback, and redundancy, and integer retrieval quotas are allocated under a unified budget constraint. The system generates query tasks according to the target assertion and target evidence role, writes new evidence into the cross-round evidence ledger and updates its status. Finally, based on sufficiency, marginal benefit, and remaining budget, it decides whether to continue retrieval, stop normally, or downgrade the output.

[0040] The assertion-evidence role unit is the smallest scheduling object in this method. Each unit records target coverage, current coverage, evidence gaps, source distribution, conflict status, and unresolved issues; unapproved candidate evidence must not alter these states. Thus, data retrieval, evidence determination, budget scheduling, and result output can operate continuously around the same state.

[0041] like Figure 1As shown, the problem to be analyzed is constructed from candidate conclusions and decomposed into assertions to form assertion objects to be verified. Candidate data becomes valid evidence after dynamic role determination and access verification. The system updates the cross-round evidence ledger and assertion-evidence role status with valid evidence, calculates evidence gaps, assertion conflicts, and redundancies, and then generates integer quotas and retrieval tasks under budget constraints. When the stopping condition is not met, newly recalled data returns to the evidence determination stage; when the stopping condition is met, a stopping object is formed, and it enters normal output or downgraded output according to the sufficiency status. Downgraded output retains unmet gaps, existing evidence, unresolved conflicts, missing evidence roles, and subsequent verification paths.

[0042] like Figure 2 As shown, each processing module collaborates through a shared state object. The assertion construction module writes the assertions to be verified and their applicable conditions; the dynamic role determination module and the evidence admission module record the role and processing results of candidate evidence; the state calculation module updates the source group contribution, current coverage, evidence gaps, assertion conflicts, and redundancy. The budget scheduling module generates basic guarantee quotas, incremental quotas, and final execution quotas, which the role-conditional retrieval module uses to form retrieval tasks. The evidence ledger module records the relationships between sources, recalls, acceptances, and state contributions throughout each round, while the stop and output module records the reasons for stopping, sufficiency status, degraded status, and unmet gaps. Each object uses a unified identifier and includes the update round, making quota allocation, stop decisions, and output references traceable.

[0043] When constructing an argument, the original question no longer directly serves as the sole retrieval object. The processing module extracts the subject, relation, object, or target result from the candidate conclusions and supplements it with applicable conditions, object scope, time range, argument importance, and output risk level, forming a verifiable argument that can be supported, challenged, limited, or interpreted by evidence. If a complete relation cannot be extracted temporarily, a variable-level degenerate argument is generated first, and the missing fields are recorded. Further subdivision is performed after relation evidence is obtained.

[0044] Preferably, the roles of evidence are categorized into five types: supporting, challenging, boundary, mechanistic, and methodological. Supporting evidence directly observes or verifies the target relationship; challenging evidence indicates that the relationship is not valid, has the opposite direction, or yields inconsistent results; boundary evidence defines thresholds, operating conditions, objects, time periods, or other applicable scopes; mechanistic evidence describes the mediating path, chain of action, or underlying mechanism; and methodological evidence describes the measurement, experimental, or analytical verification methods.

[0045] like Figure 3As shown, the assertions to be verified are stored as structured objects, including assertion identifier, subject, relationship, object or target result, applicable conditions, object scope, time range, assertion importance, output risk level, and necessary evidence role. The subject, relationship, and object or target result together determine the minimum relationship to be verified; applicable conditions, object scope, and time range limit its validity boundaries; assertion importance and output risk level are used to determine target coverage and budget priority. Dynamic evidence role determination simultaneously reads candidate evidence text, static role priors, and target assertions, outputting the target assertion identifier, candidate evidence identifier, dynamic evidence role, role confidence, and condition compatibility status. The same evidence can assume different roles relative to different assertions; evidence with unclear roles, misaligned conditions, or only providing background information is not counted towards effective coverage.

[0046] Candidate evidence can only change the assertion-evidence role status after passing the admission verification. The verification includes whether the evidence text and source can be located, whether the relevance of the evidence to the target assertion reaches the threshold, whether the confidence level of the dynamic role reaches the role threshold, whether the evidence conditions are compatible with the assertion conditions, and whether there are cases of complete duplication or inconsistency between the use of the evidence and the dynamic role.

[0047] Candidate evidence that fails to pass the admission criteria is not counted in the effective coverage, but the system still saves the reasons for rejection. Evidence with missing source locations, uncertain roles, or conflicting applicable conditions enters the pending review area and can be re-verified after supplementing the necessary information. The system also generates source group identifiers based on document identifiers, Uniform Resource Locators, database primary keys, document fingerprints, or combinations of author and publication time information, grouping adjacent fragments, abstract and text fragments, and duplicate copies from the same source into the same source group.

[0048] like Figure 4 As shown, candidate evidence objects are registered with evidence and source identifiers, target assertions, static role priors, dynamic evidence roles, relevance, role confidence, source quality, applicable conditions, condition compatibility, purpose of citation, degree of newness of historical information, duplication status, and recall rounds. Based on this, the system sequentially performs source location, assertion relevance, role confidence, condition compatibility, duplication status, and purpose of citation verification. Once all verifications pass, the evidence is written into the valid evidence set and categorized into the corresponding source group. Then, the system calculates the effective contribution of a single piece of evidence, the contribution of the source group, the effective coverage of source deduplication, evidence gaps, and assertion conflicts. Evidence that fails verification only retains the admission result, rejection reason, or pending review status; its effective coverage remains unchanged.

[0049] To explain Figure 4 First, define the corresponding sets and indexes for the objects of the arguments, the roles of the evidence, the candidate evidence, the source groups, and the retrieval rounds involved.

[0050]

[0051] This represents a set of assertions to be verified. Each element in the set corresponds to an assertion object with an independent identifier, and evidence needs to be obtained separately. This represents a set of evidentiary roles, including categories such as supporting, challenging, boundary, mechanism, and methodological evidence. This represents the index of the statement to be verified, used to locate a statement object in the collection; The index represents the role of evidence and is used to locate a specific type of evidence within the set. This represents the candidate evidence index, used to locate a candidate piece of evidence in the current or historical rounds. The source group index is used to locate a group of homologous evidence grouped by a unified source identifier; Represents the closed-loop retrieval round, used to mark the order of state update, evidence recall, and budget allocation, and is a non-negative integer; Represents the internal traversal position or state component index, used to perform intra-group aggregation or state calculations, and is a non-negative integer.

[0052] right Figure 4 For candidate evidence that passes the admission check, its effective contribution to the target assertion-evidence role unit is calculated using the following formula; evidence that fails the admission check is not included in the subsequent coverage calculation.

[0053]

[0054]

[0055] Representing the Wheel of Life Candidate evidence for the first The first unverified assertion The single valid contribution formed by the role of evidence is the basis for calculating the source group contribution and the effective coverage of source deduplication, and its value ranges from zero to one; The value represents the admission indicator, used to determine whether candidate evidence is included in the coverage calculation; it is set to one if all necessary checks are passed, otherwise it is set to zero. Representing the Candidate evidence and the first The correlation between the claims to be verified is used to determine whether the evidence directly relates to the relationship to be verified, and its value ranges from zero to one. Representing the The candidate evidence is relative to the first Each unverified assertion bears the responsibility of the first The confidence level of the role in dynamic evidence is used to measure the credibility of the role's judgment, and its value ranges from zero to one. Representing the The quality of the source of each candidate piece of evidence is used to evaluate the completeness, locatability, and reliability of the source, and its value ranges from zero to one. Representing the Candidate evidence and the first The condition compatibility between two unverified assertions is used to determine whether the applicable conditions of the two assertions are consistent or can be true at the same time, and the value ranges from zero to one. Representing the Wheel of Life The degree of newness of historical information for each candidate piece of evidence, used to measure its addition of sources, facts, or applicable conditions relative to the cross-cycle evidence ledger, takes a value from zero to one. The relevance threshold and role confidence threshold can be determined based on manually verified samples, historical annotation results, or a combination of both.

[0056] Within the same source group, there may be abstracts, text fragments, adjacent paragraphs, or duplicate copies. For the same assertion-evidence role unit, the following formula is preferred to take the maximum effective contribution within the group; when it is necessary to retain complementary conditions, aggregation methods such as weighted average, upper confidence bound, or truncated summation can also be used after deduplication of fact units.

[0057]

[0058] Representing the Wheel of Life The source group is the first The first unverified assertion The contribution of the source group formed by the class of evidence roles is used to represent the strength of valid evidence that can be provided by the same source, and the value ranges from zero to one. The source group index is used to group abstracts, text fragments, adjacent paragraphs, or duplicate copies from the same source into the same group. The maximum value operation is used to select the highest effective contribution from multiple candidate pieces of evidence from the same source group, ensuring that repeated segmentation of the same material does not amplify the coverage due to an increase in the number of fragments. When homologous fragments contain complementary conditions, aggregation can be performed after deduplication of fact units.

[0059] After completing the intra-group aggregation, the contributions of different source groups are accumulated using the following formula to obtain the effective coverage after source deduplication. This calculation limits the result to between zero and one, and makes the marginal contribution of new sources decrease as the existing coverage increases; when the source group set is empty, the effective coverage is zero.

[0060]

[0061]

[0062] Represents the valid contribution of a single piece of evidence, used to indicate the contribution of a candidate piece of evidence after verification of admission, relevance, role, source, conditions, and new information; The representative source group contribution is used to indicate the contribution formed after deduplication and aggregation within the same source group. Representing the Wheel of Life The first unverified assertion The effective coverage of source deduplication achieved by the evidence role is used to accumulate the contributions of multiple independent source groups, with a value ranging from zero to one; Representatives participated in the One unverified assertion and the first The evidence role covers the set of source groups for computation, which is used to limit the source range of the product operation; Representing the The complementary coverage portion not yet provided by each source group is used to ensure that the marginal contribution of new sources decreases as existing coverage increases. The calculation results are returned. Figure 4 The coverage update steps shown serve as the basis for calculating evidence gaps, argument conflicts, and subsequent scheduling priorities.

[0063]

[0064] Representing the The first unverified assertion The target coverage corresponding to the evidence role is used to specify the coverage requirements that the state unit needs to achieve; its value is determined according to the importance of the argument, the output risk and the necessity of the role, and the value ranges from zero to one. Representing the Wheel of Life The first unverified assertion The evidence gap corresponding to a type of evidence role is the non-negative difference between the target coverage and the current coverage, which takes a value of zero to one. Representing the Wheel of Life The conflict intensity of an unverified assertion is used to measure the degree to which both supportive and challenging coverage are high, and its value ranges from zero to one. Representing the The redundancy of the corresponding argument-evidence role unit in the wheel is used to measure the degree to which candidate results are occupied by source material, repeated fragments or known facts, and the value ranges from zero to one. Representing the Wheel of Life Unresolved feedback for a pending assertion is used to measure issues that have not been closed in manual review, automatic verification, or previous rounds, and takes a value of zero to one. Representing the The scheduling priority of the corresponding argument-evidence role unit in the round is used to determine the relative urgency of allocating retrieval resources in the next round, and is a non-negative real number.

[0065] The unfilled portion between the target coverage and the current effective coverage is calculated using the following formula, and the evidence gap is obtained through non-negative truncation.

[0066]

[0067] This represents an evidence gap, indicating the portion of the target coverage that has not yet been filled in between the current coverage. The target coverage is used to specify the coverage requirements that the corresponding assertion-evidence role unit needs to achieve; Represents the current effective coverage, indicating the extent to which the state cell has been covered by the independent source group; This represents a non-negative truncation operation, used to ensure that the evidence gap is not less than zero. A positive gap is only formed when the target coverage is higher than the current coverage; gaps in units that have reached or exceeded the target coverage are treated as zero. The resulting gaps are directly used in the current round of scheduling priority calculation.

[0068] When both supporting and challenging coverage of the same assertion exist simultaneously, the conflict intensity is calculated using the following formula. Conflict status is not used to exclude challenging evidence, but rather to increase the priority of tasks involving clarification of conditions, confirmation of boundaries, or revision of assertions.

[0069]

[0070] Representing the Wheel of Life The conflict intensity of an unverified assertion is used to trigger condition clarification, boundary subdivision, or assertion modification, and its value ranges from zero to one. The source deduplication coverage achieved by supporting roles representing the same unverified assertion is used to reflect the cumulative strength of supporting evidence; The source deduplication coverage achieved by challenging roles representing the same unproven assertion is used to reflect the cumulative strength of challenging evidence; The product of supporting and challenging coverage is used to ensure that conflict intensity only increases significantly when both types of coverage are simultaneously high. The results are not used to remove challenging evidence, but rather to prioritize conflict clarification and boundary confirmation tasks.

[0071] like Figure 5As shown, budget scheduling generates integer quotas based on the assertion-evidence role unit. Each quota object records the target assertion, target evidence role, evidence gap, assertion conflict, unresolved feedback, redundancy, scheduling priority, basic guarantee quota, incremental quota, final execution quota, unit cap, total budget for a single round, focus conditions, and allocation rationale. Gaps, conflicts, and unresolved feedback increase scheduling priority, while redundancy is used to suppress duplicate inputs. The system first reserves basic guarantee quotas for necessary evidence roles of key assertions, then allocates the remaining budget according to diminishing marginal returns; the two parts are added together to form the final execution quota, which, along with the focus conditions and allocation rationale, is written into the retrieval task.

[0072] For each assertion-evidence role unit, the non-negative scheduling priority is calculated using the following formula. When the target role belongs to the challenge, boundary, or conflict clarification type, the role indicator is set to one; otherwise, it is set to zero. Redundancy is used as a deduction item to suppress the continued allocation of resources to channels with a large amount of repetitive material.

[0073]

[0074] The scheduling priority of the target assertion-evidence role unit in the next round of budget allocation is used to measure the relative urgency of the unit and is a non-negative real number. Representing evidence gaps, used to prioritize state units that have not yet achieved target coverage; Represents the weight of the evidence gap, used to adjust for the impact of the evidence gap; Representative arguments conflict, used to increase the priority of search channels that require challenge for clarification or boundary confirmation; The value represents the role indicator; it is set to one when the target role is a challenge or boundary, or when the corresponding unit undertakes the conflict clarification task, otherwise it is set to zero. The weighting of conflicting arguments is used to moderate the impact of conflict states. This represents unresolved feedback and is used to prioritize status units that have issues awaiting review, unclosed feedback, or require manual verification. This represents the weight of unresolved feedback, used to adjust the impact of unresolved feedback. Represents redundancy, used to reduce the priority of state units that are repeatedly occupied by duplicate materials, homologous fragments, or known facts; This represents the redundancy reduction weight, used to adjust the reduction intensity of redundancy states; This represents the non-negative truncation operation, used to set the scheduling priority to zero when the weighted result is less than zero; , , and These represent four types of non-negative weights, which respectively adjust for the impact of evidence gaps, assertion conflicts, unresolved feedback, and redundancy; each weight can be determined based on assertion risk, manually verified samples, or historical task results.

[0075]

[0076] Representing the The first unverified assertion The basic guarantee quota corresponding to the evidence role is used to ensure that the necessary evidence role for the key argument has at least a minimum search opportunity in this round, and is a non-negative integer; Representing the The incremental quota of the corresponding argument-evidence role unit in the round is used to continue to allocate the number of searches according to the scheduling priority in addition to the basic guarantee quota, and is a non-negative integer; Representing the The final execution quota of the corresponding argument-evidence role unit in the round is used to determine the actual number of times the retrieval task is executed. It is obtained by adding the basic guarantee quota and the incremental quota, and is a non-negative integer. Representing the The total budget for a single round is used to limit the upper limit of uniform resources that can be consumed in this round, and is a non-negative integer; This represents the quota limit for the corresponding assertion-evidence role unit, used to prevent a single retrieval channel from filling the entire budget, and is a non-negative integer.

[0077] The total budget for a single round can be measured by the number of evidence items allowed in that round, or it can be calculated from the number of search requests, the number of fragments read, the number of context tokens, or the processing time. When multiple resource constraints exist, they can be uniformly converted into the same budget unit, or upper limits can be set separately in the integer allocation relationship.

[0078] The basic safeguard quota is used to ensure that essential evidence for key arguments receives a minimum search opportunity. Units that have already achieved target coverage, are not essential evidence, or are prohibited from further searching will have their basic safeguard quota reduced to zero; the sum of all basic safeguard quotas must not exceed the total budget for a single round.

[0079] After deducting the basic guarantee quota, the following formula is preferred for determining the non-negative integer incremental quota. Logarithmic marginal returns mean that the returns gradually decrease when the quota is continuously increased for the same state unit, while the unit upper limit prevents a single retrieval channel from occupying the entire budget.

[0080]

[0081] The optimized result, representing the total incremental quota for the next round, is used to determine the additional retrieval quantity for each assertion-evidence role unit beyond the basic guaranteed quota. Representative assigned to the first One unverified assertion and the first The incremental quota for evidence-like roles is a decision variable in this optimization relationship, and takes the form of a non-negative integer. Represents the scheduling priority of the current round, used to measure the necessity of adding retrieval slots to the corresponding state unit; This represents logarithmic marginal returns, used to gradually decrease the returns when the quota is continuously increased for the same state unit. Represents the overall scheduling benefit of all state units, used to select the combination of integer quotas with higher benefits under the current budget constraints; This represents the total budget constraint, used to ensure that the sum of all incremental quotas does not exceed the remaining budget after deducting the basic guarantee quota; The upper limit constraint represents the unit limit, which is used to ensure that the incremental quota of each unit does not exceed the difference between the quota limit and the basic guaranteed quota. Represents an integer field constraint, allowing the calculation result to be directly converted into the number of times the retrieval task is executed.

[0082] The final execution quota is obtained by summing the basic guaranteed quota and the incremental quota according to the following formula. The incremental quota can be solved using the maximum marginal revenue allocation, dynamic programming, branch and bound, or other deterministic integer allocation algorithms that satisfy the same constraints.

[0083]

[0084] Represents the allocation of the next round to the first One unverified assertion and the first The final execution quota for evidence-like roles is used to directly control the number of executions for the corresponding retrieval channels; This represents a basic guarantee quota, used to ensure that essential evidence for key arguments receives the minimum search opportunity; This represents the incremental quota, used to allocate the remaining budget based on scheduling priority. The final execution quota is obtained by adding the basic guarantee quota and the incremental quota. This represents a non-negativity constraint, used to ensure that no assertion-evidence role unit receives a negative quota; The upper limit constraint of the representative unit is used to ensure that the final execution quota does not exceed the available upper limit of the corresponding state unit; Represents the total budget constraint, used to ensure that the sum of all final execution quotas does not exceed the threshold. Total budget for a single round.

[0085] The obtained integer quota, together with the target assertion, target evidence role, focus condition, and reason for generation, is written into the retrieval task to control the actual number of times each retrieval channel is executed in the next round.

[0086] Role-based conditional retrieval is based on the retrieval task object. This object records the target assertion identifier, target evidence role, focus conditions, allocated quotas, generation round, and generation rationale. Different roles correspond to different retrieval intentions: the support role searches for materials that are directly observed or verified; the challenge role searches for materials that do not hold true, have opposite directions, or produce inconsistent results; the boundary role searches for thresholds, scope, operating conditions, objects, and time conditions; the mechanism role searches for mediating paths or mechanisms of action; and the method role searches for measurement, experimental, or analytical methods.

[0087] Each retrieval task can be performed through keyword retrieval, semantic vector retrieval, structured field retrieval, graph relation retrieval, or external retrieval services. Candidate ranking is based on a comprehensive evaluation of query relevance, assertion matching degree, target role matching degree, degree of novelty relative to the cross-round evidence ledger, and source independence. When candidates in the same round are similar, methods such as maximizing marginal relevance, selecting representative items after clustering, or fingerprint-based duplication suppression can be used to prevent similar fragments from filling a quota channel.

[0088] The cross-round evidence ledger records evidence identifiers, source group identifiers, target assertion identifiers, static role priors, dynamic evidence roles, relevance, role confidence, condition compatibility, effective contribution, initial recall round, acceptance round, purpose of citation, duplicate status, rejection reason, and status contribution relationship. Completely duplicated items have a historical newness level of zero; if key conditions are added to homologous fragments, a value greater than zero can be obtained; newly added independent sources containing uncovered facts take higher values.

[0089] like Figure 6 As shown, integer quotas, along with target assertions, target evidence roles, and focus conditions, constitute a role-conditional retrieval task. Candidate results undergo assertion matching, dynamic role matching, historical ledger comparison, source independence judgment, and same-round diversity rearrangement before entering the retrieval channel according to their respective quotas. Regardless of whether a candidate is accepted, rejected, or deemed a complete duplicate, the processing results are written to the cross-round evidence ledger; rejected or completely duplicated content does not occupy the effective output quota for this round. When supplementing key conditions for homologous fragments, the corresponding degree of new information is retained; candidates from independent sources that cover existing gaps are given priority in the admission process.

[0090] After each round of retrieval, the system updates source group contributions, effective coverage, evidence gaps, assertion conflicts, and redundancy status only based on evidence that has passed the admission verification. When assertions are subdivided or revised, the derivation relationship between the original assertion and the new assertion is preserved. Existing evidence must be reassessed for its dynamic role and conditional compatibility with the new assertion, and the original coverage status is not directly copied.

[0091]

[0092] Representing the The average change in coverage between rounds is used to measure the average degree of change of all valid assertion-evidence role units between adjacent rounds, and the value ranges from zero to one. Representing the The sufficiency indicator of the wheel is used to determine whether the necessary evidence for key arguments has reached the target coverage and whether high-intensity conflicts have been dealt with; it is zero or one. Representing the The continuous low gain indicator for each round is used to determine whether the average coverage change and the number of newly added independent sources have been below the corresponding threshold for several consecutive rounds, and is set to zero or one. Representing the The budget exhaustion indicator is used to determine whether the cumulative searched resources have reached the preset limit, and is set to zero or one. Representing the The stop state of the wheel is used to determine whether to terminate the search after the current round ends; it is either zero or one. Representing the The degraded state of the wheel is used to determine whether the evidence is still insufficient when stopping, and is set to zero or one.

[0093] The average coverage change between adjacent rounds is calculated using the following formula; if there is no previous round state, the initial coverage is zero. In addition to coverage change, the system also records the number of newly added independent sources, the number of valid new evidence, the rejection rate, and the effective gain per unit budget in this round.

[0094]

[0095] This represents the average change in coverage between two adjacent rounds, used to determine whether continuing the search can still significantly increase effective coverage; This represents the amount of coverage change of the same assertion-evidence role unit between the current round and the previous round, used to eliminate the influence of the direction of coverage increase or decrease on the judgment of the magnitude of change; This represents the summation operation performed on all valid arguments and evidence roles, used to summarize the coverage changes of all state units; Represents the number of valid unverified assertions, used to determine the number of assertions involved in the calculation of average coverage change; The number of representative evidence roles is used to determine the number of roles required for each valid, unverified claim in the calculation. This represents the total number of state units involved in the calculation, used to normalize coverage changes to an average value.

[0096] The coverage change threshold is set in the range of zero to one, and the threshold for adding independent sources is a non-negative integer. Both can be determined based on the historical retrieval process. This calculation is not performed when the set of arguments to be verified or the set of evidence roles is empty.

[0097] like Figure 7As shown, the stopping process sequentially checks for sufficiency of evidence, continuous low gain, and budget exhaustion, and writes the results to the stopping object. The stopping object records the stopping status and reason, sufficiency status, continuous low gain status, budget exhaustion status, degraded status, as well as unmet gaps, existing evidence, unresolved conflicts, missing evidence roles, and subsequent verification paths. When the necessary evidence roles for the key assertion reach the target coverage and high-intensity conflicts have been handled, the system outputs normally; when the evidence is not yet sufficient, but the continuous low gain or budget exhaustion conditions are met, the system outputs a downgraded version. Therefore, stagnation in retrieval or exhaustion of resources will not be mistakenly judged as sufficient evidence.

[0098] Preferably, the sufficiency condition requires that the necessary evidence for key arguments achieves target coverage, and handles high-intensity conflicts as argument revisions, clarification of boundaries, or unresolved uncertainties. The marginal benefit condition requires that the average change in coverage and the number of new independent sources remain below the corresponding thresholds for several consecutive rounds; when the cumulative budget reaches its limit, resource cessation is triggered.

[0099] The stop and degrade states are determined by the following formula. The search stops immediately when any of the conditions of sufficiency, continuous low gain, or budget exhaustion is met, but normal output only begins when the sufficiency condition is met.

[0100]

[0101] The sufficiency indicator is used to determine whether the necessary evidence roles have achieved target coverage and whether the main conflicts have been resolved; it takes one when the condition is true, and zero otherwise. Represents a continuous low-gain indicator, used to determine whether it is difficult to obtain sufficient new coverage or add independent sources by continuing the search; it takes one when the condition is met. This represents the budget exhaustion indicator, used to determine whether the available search resources have reached a preset limit; if the condition is met, one value is used. This indicates a stopped state and is used to control whether to terminate the search. This represents the maximum value among the three types of stop indication quantities; when any stop condition is met, the stop state is set to one. This indicates a downgraded status, used to distinguish between stoppage due to sufficient evidence and stoppage due to insufficient evidence; It represents the product of the stopped state and the insufficient sufficiency state; when the stopped state is one and the sufficiency indicator is zero, the downgraded state is one.

[0102] Downgraded outputs retain existing evidence, unresolved conflicts, roles with missing evidence, and subsequent verification paths, and clearly indicate the reasons for stopping, avoiding the misrepresentation of resource cessation or marginal benefit cessation as sufficient evidence.

[0103] When the sufficiency cessation condition is met, the system outputs each assertion along with its supporting evidence, challenging evidence, applicable boundaries, mechanistic or methodological evidence, and source location relationships. Upon entering the degraded output phase, assertions that do not achieve target coverage have their certainty level reduced, and the system lists existing evidence, unresolved conflicts, missing evidence roles, budget or data limitations, and directions for subsequent verification. Each citation can be traced back along the cross-round evidence ledger to the evidence text, source group, dynamic role, admission result, recall round, and status contribution relationship.

[0104] In special cases, when the set of arguments to be verified is empty, the system returns a state of supplementing or reconstructing candidate conclusions, and does not perform a targetless search; when a certain argument-evidence role unit has no available source group, its effective coverage is set to zero. If all scheduling priorities are zero and necessary roles are still insufficient, limited exploration is carried out according to the importance of the argument and the basic guarantee quota; when the cumulative budget is insufficient, it switches to degraded output.

[0105] When the sum of basic support quotas exceeds the total budget for a single round, the system reduces quotas sequentially based on the risk, importance, and role necessity of the assertion output, recording the units reduced and the reasons, and does not generate execution plans that exceed the budget. When both supporting and challenging evidence are strong, the system prioritizes boundary confirmation, conflict clarification, and assertion subdivision tasks, rather than deleting one based on the quantity of materials. When the retrieval service is abnormal, the database has no usable results, or duplicate materials are continuously recalled, the abnormal status and the evidence sufficiency status are recorded separately.

[0106] In one specific embodiment, the input task is "to analyze whether lubrication conditions under high load will increase bearing vibration, and to explain the possible effects of temperature rise and the applicable boundaries under low load conditions." The system extracts objects and conditions such as high load, lubrication conditions, bearing vibration, temperature rise, and low load from the task, forming three conclusions to be verified: bearing vibration increases under high load and insufficient lubrication; temperature rise may constitute an intermediate path for lubrication conditions to affect vibration; and the above relationship weakens under low load or within a specific temperature range.

[0107] The necessary evidence for the first claim is categorized as support, challenge, and boundary; for the second, support and mechanism; and for the third, boundary and challenge. The initial search yielded several supporting materials and measurement method data. However, after approval, it was found that multiple supporting materials belonged to the same source group. The first claim still lacks independent challenging evidence and low-load boundary evidence, and the second claim lacks mechanistic evidence connecting lubrication status, temperature rise, and vibration. The system writes these deficiencies into the corresponding state units, and does not determine the completion of the search based on the total number of recalled documents.

[0108] Subsequently, the system sequentially calculates the effective contribution of a single piece of evidence, the contribution of the source group, the effective coverage of source deduplication, evidence gaps, assertion conflicts, and scheduling priority. Supporting fragments from the same source are first aggregated within the source group; their increased quantity does not linearly amplify coverage. Challenge, boundary, and mechanistic units with larger gaps receive higher scheduling priority. Budget scheduling first reserves basic guarantee quotas for necessary roles, then allocates incremental quotas based on diminishing marginal returns, resulting in a final execution quota that does not exceed the total budget for this round.

[0109] The next round of searches will generate queries focusing on "conditions where the relationship is invalid or reversed under high load," "low load, temperature range, and equipment state boundaries," and "the action path of vibration caused by temperature rise under lubrication conditions." After comparing the recalled candidates with the cross-vessel evidence ledger, completely duplicated fragments will have zero new information, while candidates with newly added independent sources and supplemented uncovered conditions will be prioritized. Candidates can only update their corresponding status by passing source location, assertion relevance, dynamic role, condition compatibility, and duplicate status verification.

[0110] If the supporting, challenging, and boundary roles of the first assertion are covered to the target, the mechanistic role of the second assertion is covered to the target, and existing conflicts can be explained by differences in load or temperature conditions, then the sufficiency condition is met, and the system outputs the assertion with source location, applicable boundaries, and verification methods. If the cumulative budget is exhausted, and the second assertion still lacks mechanistic evidence, then this assertion is not output as a definitive conclusion, and the downgraded results list the existing indirect evidence, missing roles, and directions for subsequent verification.

[0111] In another embodiment, the evidence repository stores manufacturing process parameters, material properties, and measurement method data. The assertions to be verified are used to represent the relationship between process variables and performance indicators. The system still performs assertion splitting, dynamic role determination, source group coverage calculation, budget allocation, cross-round deduplication, and stop judgment in the aforementioned manner, only adjusting domain terminology, necessary evidence roles, and parameter configurations according to the task. The object names, role settings, and budget values ​​in this embodiment are only used to illustrate the execution relationship of the methods and do not constitute a limitation on the scope of protection.

[0112] The specific implementation steps are detailed below. The overall process includes, in sequence, S1 task initialization and construction of the assertion to be verified, S2 candidate evidence acquisition, role determination and admission, S3 source grouping and state calculation, S4 evidence gap and scheduling priority calculation, S5 integer retrieval quota allocation under budget constraints, S6 role-conditional retrieval and cross-round ledger update, and S7 stop judgment and normal or degraded output. Each sub-step is connected in the order of input, intermediate processing and output; after new candidate evidence returns to S2, it must still go through role determination, admission and state calculation again, and the coverage state must not be directly changed. Figure 8The diagram illustrates a closed loop consisting of S1 to S7: after the assertion is constructed, evidence acquisition, state calculation, gap scheduling, and budget allocation are completed sequentially before the stop judgment is entered. If further retrieval is needed, the process returns to S2; resource exhaustion and sufficient evidence are recorded using different states, and budget exhaustion is not interpreted as sufficient evidence.

[0113] In step S1, the problem to be analyzed is obtained and candidate conclusions are formed. The candidate conclusions are then split into a set of arguments to be verified, and a set of evidence role types is established.

[0114] This step first organizes the task inputs into verifiable assertions, while clarifying the evidence roles and target coverage required for each assertion, thereby defining the boundaries of subsequent evidence access.

[0115] Sub-step S1.1: Extract the smallest assertion to be verified.

[0116] Input: The problem to be analyzed, candidate conclusions, and original applicable conditions. The input is provided by the user interface, upstream analysis module, or existing task object.

[0117] Intermediate processing: The system identifies the subject, relationship, object, or target result and applicable conditions item by item. Relationships that can be independently judged for truth and have consistent applicable conditions are retained as a single assertion; sentences containing multiple relationships, multiple results, or different applicable conditions are split into multiple assertions; when a field is missing, the missing status is recorded, and no facts are fabricated. Taking "whether a certain working condition causes a certain result, and explain the applicable boundaries" as an example, the system generates causal assertions, action path assertions, and boundary assertions respectively, enabling each to obtain independent evidence.

[0118] Output: Initial set of statements to be verified S1.2 Based on this, normalize the applicable conditions of each statement.

[0119] Sub-step S1.2: Conditions for the application of the normalization conclusion.

[0120] Input: Output of sub-step S1.1 The task also includes the scope of objects, operating conditions, and timeframes specified in the task description.

[0121] Intermediate processing: Standardize object names, operating condition terminology, units of measurement, time ranges, and negation relationships. The standardized application conditions of each assertion are denoted as follows: When the scope of the object, operating conditions, or time is unknown, it is marked as unknown; records with known non-overlapping scopes are marked as incompatible conditions, and unknown conditions cannot be interpreted as unrestricted conditions. The normalization dictionary is derived from the task domain terminology and verifiable unit conversion rules, and the scope boundaries are derived from input data or upstream task objects.

[0122] Output: Normalized set of unverified assertions Each of the arguments is written in S1.3 Use this set to determine the necessary evidence roles and target coverage.

[0123] Sub-step S1.3: Determine the necessary evidence roles and target coverage.

[0124] Input: Output of sub-step S1.2 The importance of the assertions given by the task objectives, the output risks, and the verification rules.

[0125] Intermediate processing: The validation rule is first applied to the first... Each assertion generates a necessary set of roles. In this context, positional subscripts define the assertion, while textual subscripts indicate the necessary roles required by the rules; symbols... The argument is consistently positioned throughout subsequent calculations. Then, it is determined based on the importance of the argument. Risk of incorrect output and the Whether a role belongs to this set determines the validity of the assertion against the role. target coverage . Obtained by normalizing the primary and secondary relationships of tasks. Obtained by normalizing the impact level of erroneous conclusions. Calibration is performed through manual verification of samples, historical tasks, or domain validation sets. When the risk is high and challenge evidence is required, the target coverage of the challenger is increased accordingly during the calibration process; no fixed value is set when calibration records are lacking.

[0126] Output: Set of assertions to be verified Necessary role set Target Coverage Set and initial resource constraint objects S2.1 Use these objects to constrain candidate data normalization and initial resource usage.

[0127] In step S2, candidate evidence related to each claim to be verified is obtained, the dynamic evidence role type of each candidate evidence relative to each claim to be verified is determined, and a set of valid evidence is obtained through evidence admission verification.

[0128] This step receives the original candidate data, organizes it into traceable candidate evidence objects, and then determines the actual role of the evidence in conjunction with the target argument; materials that fail the verification are not included in the coverage calculation.

[0129] Sub-step S2.1: Normalize candidate data.

[0130] Input: Output of sub-step S1.3 , , and The original candidate data obtained from the database or retrieval service in this round.

[0131] Intermediate processing: Extract the main text, source identifier, location information, applicable conditions, and recall rounds; generate content fingerprints and bind them to target assertions. Round The candidate pieces of evidence are denoted as Document identifiers, database primary keys, locators, and document fingerprints are normalized according to deterministic rules; materials whose text or source cannot be located retain the reason for rejection and are not included in the valid evidence set.

[0132] Output: Set of candidate evidence S2.2 Based on this, determine the dynamic role of each candidate piece of evidence.

[0133] Sub-step S2.2: Calculate the dynamic evidence role.

[0134] Input: Output of sub-step S2.1 And the role of the target assertion and necessary evidence output by S1.3.

[0135] Intermediate processing: Extracting data based on relational direction, conclusion polarity, condition compatibility, action path, and verification purpose. Each normalized feature is used to calculate the score of candidate evidence for each role using the following formula, and the dynamic role with the highest score is selected:

[0136] Character Score Corresponding to the Candidate evidence, No. The first statement and the first Role type, subscript Used to indicate the purpose of role determination. Features The first result is obtained by comparing evidence and arguments. A normalized result, The total number of features actually extracted; weights Reflecting the role Features The impact is calculated and normalized from the role verification samples or classification model training results to maintain comparability of scores across different roles. The system uses a set of selectable roles... The one with the highest score is selected, and then... Record the dynamic role; an asterisk indicates that the value represents the optimized selection result. When the same piece of evidence has both supporting and challenging features, its dynamic role is the one with the higher score between the two.

[0137] Output: Featuring dynamic characters And the set of candidate evidence for role scores S2.3 Based on this, perform access verification.

[0138] Sub-step S2.3: Calculate the evidence admission result.

[0139] Input: Output of sub-step S2.2 In addition, rules for source location, relevance threshold, role threshold, condition compatibility, and citation usage.

[0140] Intermediate processing: For the first Round of evidence The source is verified sequentially to determine if it is locatable and consistent with the assertion. Whether directly related, roles The admission result is determined only if all six criteria are met: whether the judgment is credible, whether the applicable conditions are compatible, whether the content is completely duplicated, and whether the purpose of the citation is consistent. Set to pass. Subscript , , Separately define the evidence, arguments, and roles, and use superscripts. This indicates the retrieval round. Relevance and role thresholds are determined based on the costs of false acceptance and rejection or target metrics in manually verified samples. Condition compatibility is obtained by comparing applicable objects, operating conditions, and time ranges. Even if a role score is high, if the source cannot be located or the conditions are incompatible, the evidence remains rejected and cannot change the current coverage.

[0141] Output: Set of valid evidence Rejection of evidence set Collection to be reviewed and admission result objects S3.1 Accordingly, source grouping shall be performed only for admission evidence.

[0142] In step S3, based on the set of valid evidence, all valid evidence is grouped according to its source to obtain different source groups. Within each group, the effective contribution of a single piece of evidence is calculated, and the maximum contribution value within the group is selected as the contribution of the current source group. Based on the contributions of different source groups, the current effective coverage after source deduplication is calculated.

[0143] After the evidence is admitted, S3 first merges materials from the same source to eliminate the coverage amplification caused by repeated segmentation, and then accumulates the complementary contributions from independent sources to obtain the current coverage required for gap calculation.

[0144] Sub-step S3.1: Generate source group.

[0145] Input: Output of sub-step S2.3 and and historical cross-ship evidence ledger .

[0146] Intermediate processing: A unified source identifier is generated based on the document identifier, database primary key, locator, document fingerprint, or a verifiable combination of fields. Valid evidence with the same source identifier is grouped into the same source group. Different abstracts, text excerpts, or adjacent paragraphs are grouped into the same group as long as they can be traced back to the same source; if there is a conflict in the identifiers, the process enters a pending review state. For example, although the abstract and text excerpt of the same paper are different, they still belong to the same source group. Another independent paper belongs to the source group. .

[0147] Output: Set of source groups S3.2 Continue to calculate the individual contribution of evidence within each group.

[0148] Sub-step S3.2: Calculate the effective contribution of a single piece of evidence.

[0149] Input: Output of sub-step S3.1 And valid evidence within the group, as well as the admission, relevance, role and condition verification results output by S2.3.

[0150] Intermediate processing: Calculate the individual contribution of each valid piece of evidence to the target assertion—role unit:

[0151] Single valid contribution Corresponding to the Round Article of evidence, No. The first statement and the first Role type. Admission results. The value is 1 if the pass is successful and 0 if the fail is unsuccessful; Relevance To measure the direct relevance between evidence and conclusion, subscript Indicates correlation; For dynamic role scoring, subscript Indicates the role determination. Source quality. It is obtained by normalizing source integrity, text locationability, source type, and verification status. Indicates condition compatibility. This refers to the degree of new information obtained from ledger comparisons in S6.3. Except for the admission results, all continuous factors are normalized according to the verification sample or deterministic rules; , , Position the evidence, arguments, and roles in sequence, using superscripts. Positioning rounds. The product structure reduces the contribution when any key condition is insufficient, and evidence from the same group still needs to enter S3.3 aggregation, and cannot be accumulated as multiple independent sources.

[0152] Output: Set of single valid contributions S3.3 Based on this, the intra-group aggregation is completed.

[0153] Sub-step S3.3: Calculate the contribution of the source group.

[0154] Input: Output of sub-step S3.2 and the source group set output by S3.1 .

[0155] Intermediate processing: Within each source group, select the maximum effective contribution to the same assertion—role unit, to avoid amplifying and overwriting the same data due to an increase in the number of fragments.

[0156] Source group contribution Indicates the first Wheel Source Group On the assertion and roles Contribution The single contribution method from S3.2 is retained. The maximum value is only in the source group. Inner traversal evidence ,in , and superscript The arguments, roles, and rounds are still identified separately. When the abstract and text excerpt of the same data are grouped together, only the larger valid contribution is retained; independent data are calculated within their respective source groups.

[0157] Output: Collection of contributions from the source group S3.4 Use its cumulative complementary contributions from independent sources.

[0158] Sub-step S3.4: Calculate the effective coverage of source deduplication.

[0159] Input: Output of sub-step S3.3 And the current assertion—the effective source set of role units .

[0160] Intermediate processing: Accumulate complementary contributions from different source groups and calculate effective source deduplication coverage:

[0161] Source deduplication and coverage Corresponding to the Wheel of Reason and roles Group contribution Taken from S3.3, This represents the set of valid source groups for the unit; the product only iterates through the source groups in the set. ; , , and superscript Separately identify the assertion, role, source group, and round. Coverage is zero when the set is empty. When supporting evidence comes from two independent source groups, the contributions of both groups are accumulated based on their complementary parts; multiple fragments from the same group will not be added to the coverage repeatedly.

[0162] Output: by the currently overriding The state object is composed of the distribution of source groups and the state contribution relationship. S4.1 Based on this, compare the current coverage with the target coverage.

[0163] In step S4, the difference between the current effective coverage and the target coverage is calculated to obtain the evidence gap. The evidence gap, the conflict of arguments, the unresolved feedback and the redundancy are taken into account to determine the scheduling priority in the next round of budget allocation.

[0164] S4 takes the difference between the target coverage and the current coverage as a starting point, and also considers assertion conflicts, unresolved feedback and redundancy, to generate a scheduling basis for the next round of budget allocation.

[0165] Sub-step S4.1: Calculate the evidence gap.

[0166] Input: Output of sub-step S3.4 and the output of S1.3 .

[0167] Intermediate processing: The difference between the target coverage and the current coverage is truncated to non-negative values.

[0168] Evidence gap Corresponding to the Wheel of Reason and roles The target coverage determined by S1.3 Subtract the current coverage calculated in S3.4 get; , and superscript Separately define the assertion, role, and round. When the current coverage reaches or exceeds the target coverage, the maximum value calculation truncates the gap to zero; if it does not reach the target, the difference between the two is entered into subsequent scheduling.

[0169] Output: Evidence Gap Object The object is sent to S4.2 along with the overwrite status of S3.4.

[0170] Sub-step S4.2: Calculate the conflict of arguments.

[0171] Input: Output of sub-step S4.1 And the same assertion output by S3.4 supports coverage and challenges coverage.

[0172] Intermediate processing: Overlay the support obtained from S3.4 With challenges covered Multiply to get the first Wheel of Reason Intensity of conflict ;in and superscript Separately locate the assertion and the round, with subscripts. and These represent the supporting and challenging roles, respectively. Both types of coverage are normalized results, and the conflict only increases when both exist simultaneously; this result is used to increase the priority of the clarification task without deleting evidence from either side.

[0173] Output: Objects of Conflicting Arguments S4.3 Combines this with evidence gaps, redundancy, and unresolved feedback in the calculation.

[0174] Sub-step S4.3: Calculate scheduling priority.

[0175] Input: Output of sub-step S4.2 In addition, there are gaps in the output of S4.1, redundancy in the output of S3, and unresolved feedback.

[0176] Intermediate processing: Computation is not a negative truncation assertion—role unit scheduling priority:

[0177] Scheduling priority Corresponding to the Wheel of Reason and roles .in, The evidence gap taken from S4.1 The strength of the argument conflict is taken from S4.2. Mark character Whether to undertake conflict clarification; The result is obtained by normalizing the number and severity of pending review items and unclosed issues. The weighting is calculated based on the proportions of candidate numbers, homologous numbers, and duplicate facts. (Non-negative weighting) , , and The terms correspond to gaps, conflicts, feedback, and redundancy, respectively, with the letter subscript indicating the state type. The specific weight is determined by the verification sample or historical tasks. The plus sign in the lower right corner of the square brackets indicates that negative results are truncated to zero. In this way, results with gaps and those that play the role of conflict clarification will receive higher priority, while results repeatedly derived from the same data will be deducted from the redundancy item.

[0178] Output: Set of scheduling units Each cell is written 1. Objectives, objectives, roles, and rationale for allocation; S5.1 Based on this, calculate the basic guarantee quota for necessary roles.

[0179] In step S5, based on the resource budget, basic guarantee quota, and quota upper limit constraint, the number of retrievals is determined according to the scheduling priority to execute the quota.

[0180] S5 first guarantees the basic quotas for essential roles, then converts reliably calculable multidimensional resources into search counts, and determines the execution quotas under the constraints of integer fields, unit limits, and total budget.

[0181] Sub-step S5.1: Calculate the basic guarantee quota.

[0182] Input: Output of sub-step S4.3 And also necessity, continued search permission, rule quotas and unit limits.

[0183] Intermediate processing: For assertions—role units—that are allowed to continue searching in the next round and belong to necessary roles, the smaller value between the candidate quota given by the verification rules and the unit upper limit is taken as the assertion. and roles Basic guarantee quota superscript This indicates the next round; it is set to zero if the necessity requirement is not met or if a continued retrieval permit is not granted. Role necessity is derived from S1.3, and the continued retrieval permit is derived from the current stop and resource status. The rule quota is converted into a non-negative integer based on the assertion importance, output risk, and necessary roles. The unit limit is calculated based on the constraints of task risk, total resources, and preventing a single channel from filling the budget, and no fixed value without a source is used.

[0184] Output: Basic guarantee quota set S5.2 Based on this, deduct the number of guaranteed search attempts.

[0185] Sub-step S5.2: Unify resource budget and calculate remaining budget.

[0186] Input: Output of sub-step S5.1 And the available amount and unit retrieval consumption for each resource dimension.

[0187] Intermediate processing: Convert reliably calculable resource dimensions into single-round search counts, and deduct the basic guarantee quota.

[0188] Unified General Budget Indicates the first The maximum number of searches that can be performed in a round. (Set) It only includes resource dimensions that can be reliably converted into search counts. Positioning one dimension; This refers to the available quantity for this dimension in this round. This represents the average cost per retrieval for that dimension. The total budget is calculated by dividing the available resources for each dimension by the unit cost, rounding down, and taking the minimum value. Remaining budget. subscript This indicates the remaining amount, the value of which is the total budget minus all assertions. ,Role The next round of basic guarantee quota Then perform a non-negative truncation. Average consumption. The result is obtained by dividing the resource consumption in the historical execution record by the number of completed searches; only dimensions whose results are calculable and greater than zero participate in the unified conversion, while the remaining dimensions are retained as independent upper limits.

[0189] Output: Budget object ,include , The sum of independent resource caps and basic guarantee quotas; S5.3 calculates the incremental quota within this budget boundary.

[0190] Sub-step S5.3: Calculate the non-negative integer incremental quota.

[0191] Input: Output of sub-step S5.2 and the output of S4.3 And the basic guarantee quota for S5.1 output.

[0192] Intermediate processing: Calculate the next round of incremental quotas under the constraints of remaining budget, integer field, and unit upper limit:

[0193] Optimal Incremental Quota Vector This summarizes the incremental quotas for each unit in the next round; an asterisk indicates the optimized selection result. Unit Increment It is assigned to the argument. and roles non-negative integers, Take the current round priority calculated in S4.3; the logarithmic term is used to reflect the diminishing marginal returns when the quota is continuously increased for the same unit. The sum of all increments must not exceed the remaining budget obtained in S5.2. The unit increment must also not exceed the upper limit. Deducting basic guarantee quota The remaining margin. Here and Positioning and Roles, Superscript Indicates the current state. This indicates the next round of execution. When multiple roles require increments simultaneously, the solver allocates integer execution times within a shared budget and their respective unit limits.

[0194] Output: Incremental quota set S5.4 merges it with the basic guarantee quota.

[0195] Sub-step S5.4: Calculate the final execution quota.

[0196] Input: Output of sub-step S5.3 and S5.1 output .

[0197] Intermediate processing: This involves adjusting the basic guarantee quota of S5.1. Incremental quotas with S5.3 Add them together to arrive at the next conclusion. and roles Final execution quota superscript This indicates the next round. The system verifies each item to ensure the final quota is a non-negative integer and does not exceed the unit limit. The subscript in the upper limit This indicates the maximum allowable value. The sum of all unit quotas must not exceed the total budget obtained in S5.2. If any constraint is not satisfied, the solution is recalculated or the quota is reduced according to the risk order, and the reduced units and the reasons are recorded.

[0198] Output: Integer execution quota object S6.1 Generate a specific retrieval task based on the object.

[0199] In step S6, a quota is generated based on the determined number of searches, and the search is performed. The candidate results obtained from the search are sorted by diversity and compared across rounds of information to obtain a new set of candidate evidence, and the evidence admission verification is performed again.

[0200] Once the integer quota is determined, S6 generates a role-conditional retrieval task accordingly, performs diversity ranking and cross-round information comparison on the candidate results, and then returns the newly discovered evidence to S2 for admission verification.

[0201] Sub-step S6.1: Generate role-conditional retrieval task.

[0202] Input: Output of substep S5.4 And the target proposition, target role, and key conditions.

[0203] Intermediate processing: Generating and... for each assertion-role unit The same number of retrieval tasks are required. Support tasks emphasize direct support relationships, challenge tasks emphasize counterexamples or negation relationships, boundary tasks emphasize applicable objects, operating conditions, and time ranges, mechanism tasks emphasize action paths, and method tasks emphasize measurement or verification methods. The assertion identifiers, role identifiers, focus conditions, and quota sources in the tasks are copied from the input objects and cannot be rewritten by the retrieval module.

[0204] Output: Set of character-conditional retrieval tasks S6.2 Based on this, perform retrieval and candidate sorting.

[0205] Sub-step S6.2: Perform the retrieval and sort the candidate results.

[0206] Input: Output of substep S6.1 and historical ledgers And the original candidate results returned by the search service.

[0207] Intermediate processing: Calculate query relevance, assertion matching degree, role matching degree, degree of historical newness, source independence, and redundancy within the same round to form a candidate ranking score.

[0208] Sorting score Corresponding to the next round Candidate evidence, No. The first statement and the first Role type. Query matching items. The assertion matching terms are calculated based on the term and semantic matching results between the query and the candidate text. The role matching item is calculated based on the matching results between candidate content and the assertion subject, relation, result, and condition fields. The character trait calculation method from S2.2 is used; subscript , and These represent query, assertion, and role matching, respectively. Level of new information. The source independence was obtained from a preliminary comparison with historical ledgers before sorting. Calculated based on the difference between candidate source groups and accepted source groups, subscript Indicates source dimension; redundancy The candidate is selected based on the maximum similarity to materials already selected in the same round. Weighting. to All values ​​are non-negative and are determined by sorting and verifying samples or historical tasks; the first five items add points, and redundant items deduct points. , , and superscript By separately identifying evidence, arguments, roles, and the next round, materials that align with the boundary focus and originate from new sources are prioritized, while highly repetitive materials within the same round are suppressed.

[0209] Output: Complete deduplication within the same cycle and press [button]. Ranking of candidate evidence set S6.3 Based on this, compare with the historical ledger and calculate the degree of new information.

[0210] Sub-step S6.3: Calculate new cross-ship information and update the ledger.

[0211] Input: Output of sub-step S6.2 and historical cross-ship evidence ledger .

[0212] Intermediate processing: Compare candidate evidence with records in the historical ledger to calculate the degree of newness of historical information.

[0213] Historical new information level Corresponding to the next round 10 candidate pieces of evidence; This indicates the candidate evidence. Indicates the ledger before the update The first in Each record. The similarity function calculates a normalized result based on the content fingerprint, source group, fact unit, and applicable conditions, and then takes the maximum value from all historical records; and Locate the current evidence and historical records separately, and use superscripts. and These represent the previous and next rounds, respectively. When the ledger is empty, the maximum similarity is set to zero; when candidate evidence shares the same origin as historical records and has the same factual units and applicable conditions, the degree of new information is reduced to zero; when verifiable differences exist, the corresponding degree of new information is retained. After the comparison is complete, the evidence status, source group, target assertion, dynamic role, and contribution relationship are written into the new ledger. .

[0214] Output: New set of candidate evidence and the updated ledger The former returns to S2.1, which is then processed by S2.2, S2.3, and S3 in sequence; the recalculated state is then sent to S7.1.

[0215] In step S7, it is determined whether the retrieval termination condition is met. If the retrieval termination condition is met, it is further determined whether the evidence is sufficient, and the result of the determination is used to determine whether to output normally or in a downgraded manner. If the retrieval termination condition is not met, the retrieval continues.

[0216] After S2 and S3 complete a new round of status updates, S7 combines the latest coverage, gains from adjacent rounds, accumulated resources, and unresolved conflicts to determine whether to output normally, downgrade output, or continue searching.

[0217] Sub-step S7.1: Calculate the gain of adjacent rounds.

[0218] Input: S6.3 Return to S2 and re-execute S2 and S3 to obtain the current coverage. Previous round of coverage Valid argument set and character collection .

[0219] Intermediate processing: Calculate the average coverage change of all valid arguments—role units, and simultaneously calculate the number of new independent sources, the number of valid new evidence, and the effective gain per unit budget.

[0220] Average Coverage Change Indicates the first The state increment of a round relative to the previous round. and These are the set of valid arguments and the set of roles involved in the calculation, respectively. The cardinality of these two sets is used to normalize the total change to an average value. and Taken from S3.4, these represent the coverage of the current round and the previous round, respectively, with the absolute value retaining the range of change. , and Positioning, role, and current round (superscript) and Distinguish between adjacent rounds. Even if only some roles are retrieved in this round, the average value is still calculated based on all valid units; if any set is empty, division is not performed, but input completion or downgrading processing is initiated instead.

[0221] Output: Gain state object ,Record The number of newly added independent sources, the number of valid new evidence, and the effective gain per unit budget; S7.2 Use this object to determine the stopping method.

[0222] Sub-step S7.2: Calculate the stop, downgrade, and continue retrieval status.

[0223] Input: Output of sub-step S7.1 This includes cumulative resource consumption, cumulative budget, necessary role coverage, and unresolved conflicts.

[0224] Intermediate processing: Calculate the sufficiency indicator separately. Continuous low gain indication and budget exhaustion indicators Then calculate the stopped state and the degraded state:

[0225] Stopped state Corresponding to the Wheel, downgraded state This specifically indicates that the evidence was still insufficient when the vessel stopped. (Sufficiency indicator) Determined by whether the necessary roles achieve target coverage and whether major conflicts are resolved; continuous low-gain indicator. Combined with average coverage change The number of new independent sources and the number of consecutive low-gain rounds are determined; budget exhaustion indicator. This is derived by comparing cumulative resource consumption with cumulative budget. The three indicators are set to 1 if the condition is met, and 0 otherwise. (Subscript...) The current round is represented uniformly. The low-gain threshold and consecutive round count are determined by historical rounds or manually verified samples, and the cumulative budget is taken from the task resource constraints. A stop is triggered when any indicator is 1; if sufficiency is not met but consecutive low gain or budget exhaustion is met, the product term is simultaneously degraded. If none of the three conditions are met, the process continues to the next round of retrieval.

[0226] Output: When When a normal output object is formed; when and When a degraded output object is formed; when A request to continue searching is generated at that time. Return to S2.1 to start the next round.

[0227] Specific example: Bearing lubrication condition analysis task.

[0228] This embodiment analyzes whether lubrication conditions under high load will increase bearing vibration, and examines the possible action path of temperature rise and the applicable boundaries under low load conditions. The implementation process uses the actual relevance, source quality, applicable conditions, and resource records obtained from the task data; the following only describes how each object enters the calculation and is passed on, without pre-setting fixed thresholds, weights, or experimental results.

[0229] S1 breaks down the task into three unverified assertions: "Increased vibration under high load and insufficient lubrication," "Temperature rise may affect vibration through lubrication conditions," and "The above relationship weakens under low load or within a specific temperature range." Each assertion is assigned supporting, challenging, boundary, mechanistic, or methodological roles, with objectives covering those outlined in S1.3. Calculated.

[0230] S2 obtained three candidate pieces of evidence related to the role supporting the first argument. , and The first two are different fragments of the same data, while the last one comes from independent data. The dynamic roles of each piece of evidence are first calculated using S2.2, and then calculated using S2.3. Only those who pass all the admission criteria can enter S3.

[0231] S3 and Into ,Bundle Into The contribution of each of the three pieces of evidence is calculated individually, and fragments from the same source are first aggregated. Independent source formation Then, the support coverage of the argument is obtained by deduplication and coverage of the source in S3.4; in this way, the complementary information of the two independent sources is preserved, and two fragments of the same data are not counted repeatedly.

[0232] S4 substitutes this coverage If there are still gaps in the challenge role and the low-load boundary role, and support coverage conflicts with challenge coverage, then , and They are all included in the priority calculation, so that subsequent searches focus on counterexamples and applicable boundaries, rather than continuing to repeatedly recall supporting materials.

[0233] S5 first calculates the necessary roles. Then, the resource records are used for calculation. and The incremental quotas for challenge roles and boundary roles meet the requirements. The final number of executions is determined by get.

[0234] S6 according to Generate a challenge retrieval task, according to Generate a low-load and temperature boundary retrieval task. Candidate results are sorted by... Sort and use Compared with the historical ledger; completely duplicated materials no longer change the coverage, while new sources or new boundary conditions return to S2 for re-admission.

[0235] S7 is calculated after the new evidence re-enters S2 and S3. When all necessary roles have achieved objective coverage and major conflicts have been resolved, And output normally; when there is continuous low gain or the budget is exhausted and sufficiency is still not met, The output is downgraded, and the output objects clearly list the existing evidence, missing roles, unresolved conflicts, and directions for subsequent verification.

[0236] This processing chain is also applicable to tasks such as manufacturing processes, material properties, or measurement methods. When applying it, the corresponding assertion content, evidence roles, and parameter calibration data are replaced, and the assertion construction, evidence admission, source deduplication and coverage, gap scheduling, integer quota allocation, cross-round retrieval, and output termination are still completed according to S1 to S7; the specific retrieval model and integer solution method are not limited.

[0237] The above specific examples illustrate the present invention only to aid in understanding it and are not intended to limit the scope of the invention. Those skilled in the art can make various simple deductions, modifications, or substitutions based on the ideas of this invention. It is impossible to exhaustively list all possible embodiments here; all obvious variations or modifications derived from the technical solutions of this invention are still within the protection scope of this invention.

Claims

1. A budget-constrained closed-loop retrieval method driven by an assertion evidence gap, characterized in that, The method includes: The problem to be analyzed is identified and candidate conclusions are formed. The candidate conclusions are then broken down into a set of arguments to be verified, and a set of evidence role types is established. Acquire candidate evidence related to each claim to be verified, determine the dynamic evidence role type of each candidate evidence relative to each claim to be verified, and obtain a set of valid evidence through evidence admission verification; Based on the set of valid evidence, all valid evidence is grouped according to its source to obtain different source groups. Within each group, the valid contribution of a single piece of evidence is calculated and the maximum contribution value within the group is selected as the contribution of the current source group. Based on the contributions of different source groups, the current valid coverage after deduplication of sources is calculated. The difference between the current effective coverage and the target coverage is calculated to obtain the evidence gap. The evidence gap, conflicting arguments, unresolved feedback and redundancy are considered to determine the scheduling priority in the next round of budget allocation. Based on resource budget, basic guarantee quota, and quota upper limit constraints, the number of retrievals and quota execution are determined according to scheduling priority. Based on a defined number of searches, a quota is generated to perform a search task and the search is executed. The candidate results obtained are sorted by diversity and compared across rounds of information to obtain a new set of candidate evidence and the evidence admission is re-verified. Determine whether the search termination condition is met. If the condition is met, further determine whether the evidence is sufficient and decide whether to output normally or in a downgraded manner based on the result. If the condition is not met, continue the search.

2. The assertion evidence gap-driven budget-constrained closed-loop retrieval method as described in claim 1, characterized in that, The process involves identifying the problem to be analyzed and forming candidate conclusions. These candidate conclusions are then broken down into a set of arguments to be verified, and a set of evidence role types is established, specifically including: From the problem to be analyzed, candidate conclusions and original applicable conditions, identify the subject, relationship, object or target result and applicable conditions one by one. Relationships that can be independently judged as true or false and have consistent applicable conditions are retained as one statement. Sentences containing multiple relationships, multiple results or different applicable conditions are split into multiple statements. When a field is missing, the missing status is recorded and no facts are created. The applicable conditions of each assertion are standardized, and the object name, working condition terminology, unit of measurement, time range and negation relationship are unified. When the object range, working condition range or time range is unknown, an unknown mark is retained. Records with known ranges that do not intersect are marked as incompatible conditions.

3. The assertion evidence gap-driven budget-constrained closed-loop retrieval method as described in claim 1, characterized in that, The process involves identifying the problem to be analyzed and forming candidate conclusions. These candidate conclusions are then broken down into a set of arguments to be verified, and a set of evidence role types is established, specifically including: The set of evidence role types is used to characterize the types of roles that candidate evidence can play relative to the target assertion. Evidence role types include support, challenge, boundary, mechanism, and method.

4. The assertion evidence gap-driven budget-constrained closed-loop retrieval method as described in claim 1, characterized in that, Acquire candidate evidence related to each claim to be verified, determine the dynamic evidence role type of each candidate evidence relative to each claim to be verified, and obtain a set of valid evidence through evidence admission verification, specifically including: Obtain the original candidate data, extract the main text, source identifier, location information, applicable conditions and recall rounds, generate content fingerprints and bind them to the target assertion, normalize the document identifier, database primary key, locator and document fingerprint according to deterministic rules, retain the rejection reason for the data whose main text or source cannot be located and do not enter the valid evidence set; Normalized features are extracted from relational direction, conclusion polarity, condition compatibility, action path and verification purpose. The scores of candidate evidence for each evidence role type are calculated. The evidence role type with the highest score is selected as the dynamic evidence role of the corresponding candidate evidence relative to the target assertion. The same evidence has different dynamic evidence roles relative to different assertions. The criteria for admission of candidate evidence are verified sequentially, including: whether the source is locatable, whether it is directly related to the target assertion, whether the dynamic role determination is credible, whether the applicable conditions are compatible, whether the content is completely duplicated, and whether the purpose of citation is consistent. If all conditions are met, the admission result is set to pass and included in the set of valid evidence. If any condition is not met, the evidence is rejected and the reason for rejection is recorded.

5. The assertion evidence gap-driven budget-constrained closed-loop retrieval method as described in claim 1, characterized in that, Based on the set of valid evidence, all valid evidence is grouped according to its source to obtain different source groups. Within each group, the valid contribution of a single piece of evidence is calculated, and the maximum contribution value within the group is selected as the contribution of the current source group. Based on the contributions of different source groups, the current valid coverage after deduplication is calculated. A unified source identifier is generated based on document identifier, database primary key, locator, document fingerprint, or a verifiable combination of fields. Valid evidence with the same source identifier is grouped into the same source group, and when there is a conflict in the identifier, it enters a pending review state. For each piece of valid evidence, the contribution of that evidence to the target assertion-evidence role unit is calculated based on the admission result, relevance, dynamic role score, source quality, condition compatibility, and degree of newness of historical information. The contribution value is reduced accordingly if any key condition is insufficient, as shown in the following formula: in, Representing the Wheel of Life Candidate evidence for the first The first unverified assertion A single valid contribution formed by a class of evidentiary roles; The evidence admission indicator is used to determine whether candidate evidence is included in the coverage calculation. It is set to one if all necessary checks are passed, and zero otherwise. Representing the Candidate evidence and the first The correlation between claims to be verified is used to determine whether the evidence directly relates to the relationship to be verified. Representing the The candidate evidence is relative to the first Each unverified assertion bears the responsibility of the first The confidence level of a role in dynamic evidence is used to measure the credibility of the role's judgment. Representing the The quality of the source of each candidate piece of evidence is used to evaluate the completeness, locatability, and reliability of the source; Representing the Candidate evidence and the first The compatibility of conditions between two unverified assertions is used to determine whether the applicable conditions of the two assertions are consistent or can be true at the same time. Representing the Wheel of Life The degree of newness of historical information in each candidate piece of evidence is used to measure its addition of sources, facts, or applicable conditions relative to the cross-cycle evidence ledger; Within each source group, the maximum effective contribution of a single piece of evidence for the same assertion-evidence role unit is selected as the contribution value for the corresponding source group, as shown in the following formula: in, Representing the Wheel of Life The source group is the first The first unverified assertion The contribution of the source group formed by the role of evidence is used to indicate the strength of valid evidence that can be provided by the same source; The source group index is used to group abstracts, text fragments, adjacent paragraphs, or duplicate copies from the same source into the same group. The maximum value operation is used to select the highest effective contribution from multiple candidate evidences from the same source group, so that the same material will not be amplified and covered by the increase in the number of fragments after repeated segmentation. The contributions from different source groups are then summed using the following formula to obtain the effective coverage after deduplication: in, Represents the valid contribution of a single piece of evidence, used to indicate the contribution of a candidate piece of evidence after verification of admission, relevance, role, source, conditions, and new information; The representative source group contribution is used to indicate the contribution formed after deduplication and aggregation within the same source group. Representing the Wheel of Life The first unverified assertion The source deduplication and effective coverage achieved by the evidence-like role are used to accumulate the contributions of multiple independent source groups; Representatives participated in the One unverified assertion and the first The evidence role covers the set of source groups for computation, which is used to limit the source range of the product operation; Representing the The complementary coverage portions not yet provided by each source group are used to ensure that the marginal contribution of new sources decreases as existing coverage increases.

6. The assertion evidence gap-driven budget-constrained closed-loop retrieval method as described in claim 1, characterized in that, The difference between the current effective coverage and the target coverage is calculated to obtain the evidence gap. Taking into account the evidence gap, conflicting arguments, unresolved feedback, and redundancy, the scheduling priority in the next round of budget allocation is determined, specifically including: The unfilled portion between the target coverage and the current effective coverage is calculated using the following formula, and the evidence gap is obtained through non-negative truncation: in, This represents an evidence gap, indicating the portion of the target coverage that has not yet been filled in between the current effective coverage. The target coverage is used to specify the coverage requirements that the corresponding assertion-evidence role units need to achieve; This represents the current effective coverage, indicating the extent to which the corresponding assertion-evidence role unit has been covered by the independent source group; This represents a non-negative truncation operation, used to ensure that the evidence gap is not less than zero; a positive gap is only formed when the target coverage is higher than the current effective coverage; cell gaps that have reached or exceeded the target coverage are treated as zero; For each assertion-evidence role unit, the scheduling priority is calculated using the following formula; in, The scheduling priority of the target assertion-evidence role unit in the next round of budget allocation is used to measure the relative urgency of the corresponding unit; Representing evidence gaps, used to prioritize state units that have not yet achieved target coverage; Represents the weight of the evidence gap, used to adjust for the impact of the evidence gap; Representative arguments conflict, used to increase the priority of search channels that require challenge for clarification or boundary confirmation; The value represents the evidence role indicator. It is set to one when the evidence role is a challenge or boundary type, or when the corresponding unit undertakes the conflict clarification task; otherwise, it is set to zero. The weighting of conflicting arguments is used to moderate the impact of conflict states. This represents unresolved feedback and is used to prioritize units with issues awaiting review, unclosed feedback, or manual verification requirements. This represents the weight of unresolved feedback, used to adjust the impact of unresolved feedback. Represents redundancy, used to reduce the priority of units repeatedly occupied by duplicate material, homologous fragments, or known facts; This represents the redundancy reduction weight, used to adjust the reduction intensity of redundancy states; This represents the non-negative truncation operation, used to set the scheduling priority to zero when the weighted result is less than zero; , , and These represent four types of non-negative weights, which respectively adjust for the effects of evidence gaps, conflicting arguments, unresolved feedback, and redundancy.

7. The assertion evidence gap-driven budget-constrained closed-loop retrieval method as described in claim 1, characterized in that, Based on resource budget, basic guarantee quota, and quota upper limit constraints, the number of retrievals and the execution quota are determined according to scheduling priority, specifically including: in, Represents the allocation of the next round to the first One unverified assertion and the first The final execution quota for evidence-like roles is used to directly control the number of executions for the corresponding retrieval channels; The basic guarantee quota is used to ensure that essential evidence for key arguments has the minimum chance of being searched. Units that have already achieved target coverage, are not essential evidence, or are prohibited from being searched further have their basic guarantee quota set to zero. This represents the incremental quota, used to allocate the remaining budget based on scheduling priority. This represents a non-negativity constraint, used to ensure that no assertion-evidence role unit receives a negative quota; The upper limit constraint for the representative unit is used to ensure that the final execution quota does not exceed the available upper limit of the corresponding state unit. ; Represents the total budget constraint, used to ensure that the sum of all final execution quotas does not exceed the threshold. Total budget for a single round The total budget for a single round is measured by the number of evidence items allowed to be accepted in this round, or calculated based on the number of search requests, the number of read fragments, the number of context tokens, or the processing time.

8. The assertion evidence gap-driven budget-constrained closed-loop retrieval method as described in claim 7, characterized in that, Based on resource budget, basic guarantee quota, and quota upper limit constraints, the number of retrievals and the execution quota are determined according to scheduling priority, specifically including: After deducting the basic guarantee quota, the incremental quota is determined using the following formula: in, The optimization results, representing all incremental quotas in the next round, are used to determine the additional number of searches for each assertion-evidence role unit beyond the basic guaranteed quota. Representative assigned to the first One unverified assertion and the first Incremental quotas for evidence-based roles; Represents the scheduling priority of the current round, used to measure the necessity of increasing the number of retrieval slots to the corresponding unit; This represents logarithmic marginal returns, used to gradually decrease the returns when the quota is continuously increased for the same state unit. Represents the overall scheduling benefit of all state units, used to select the combination of integer quotas with higher benefits under the current budget constraints; This represents the total budget constraint, used to ensure that the sum of all incremental quotas does not exceed the remaining budget after deducting the basic guarantee quota; The upper limit constraint represents the unit limit, which is used to ensure that the incremental quota of each unit does not exceed the difference between the quota limit and the basic guaranteed quota. Represents an integer field constraint, allowing the calculation result to be directly converted into the number of times the retrieval task is executed.

9. The assertion evidence gap-driven budget-constrained closed-loop retrieval method as described in claim 1, characterized in that, Based on a predetermined number of searches, a quota is used to generate search tasks and execute searches. The retrieved candidate results are then sorted by diversity and compared across rounds of information to obtain a new set of candidate evidence. The evidence admission is then re-verified, specifically including: For each assertion-evidence role unit, generate the same number of retrieval tasks as the execution quota. Support tasks emphasize direct support relationships, challenge tasks emphasize counterexamples or negation relationships, boundary tasks emphasize applicable objects, working conditions and time ranges, mechanism tasks emphasize action paths, and method tasks emphasize measurement or verification methods. The assertion identifier, role identifier, focus conditions and quota sources in the tasks are all copied from the input objects. Perform a search and calculate candidate ranking scores based on query relevance, assertion matching degree, role matching degree, degree of historical newness, source independence, and redundancy within the same round. Among these, the degree of historical newness and source independence have a positive impact on the score, while redundancy within the same round has a negative impact on the score. Perform deduplication within the same round and filter candidate evidence according to the ranking scores. The candidate evidence is compared with the records in the historical cross-cycle evidence ledger by content fingerprint, source group, fact unit and applicable conditions. The normalized degree of historical new information is calculated. When the candidate evidence and the historical record are from the same source and the fact unit and applicable conditions are the same, the degree of new information is zero. When there are verifiable differences, the corresponding degree of new information is retained. After the comparison is completed, the evidence status, source group, target assertion, dynamic role and contribution relationship are written into the updated cross-cycle evidence ledger. The new candidate evidence set obtained after sorting and cross-round comparison will be returned to re-execute dynamic role determination and evidence admission verification. Content that has been accepted, rejected or completely duplicated will be retained in its historical state and will not be directly included in the current valid evidence set.

10. The assertion evidence gap-driven budget-constrained closed-loop retrieval method as described in claim 1, characterized in that, Determine if the search termination condition is met. If it is, further assess the sufficiency of the evidence and determine whether to output normally or in a downgraded manner based on the assessment result. If the search termination condition is not met, continue the search, specifically including: Calculate the average coverage change of all valid assertion-evidence role units, and simultaneously count the number of newly added independent sources, the number of valid new evidence, and the effective gain per unit budget. The average coverage change is normalized according to the cardinality of the set of all valid assertions and the set of roles involved in the calculation. The sufficiency indicator, the continuous low gain indicator, and the budget exhaustion indicator are calculated separately. The sufficiency indicator is determined by whether all necessary roles have achieved the target coverage and whether major conflicts have been resolved. The continuous low gain indicator is determined by the average coverage change, the number of new independent sources, and the number of consecutive low gain rounds, and the threshold is calibrated by historical rounds or manually verified samples. The budget exhaustion indicator is determined by comparing the cumulative resource consumption with the cumulative budget. If the sufficiency indicator is true, a sufficiency stop is triggered and normal output is executed; if the sufficiency indicator is false but the continuous low gain indicator or the budget exhaustion indicator is true, a degraded output is triggered, and the output object clearly lists the existing evidence, missing roles, unresolved conflicts and subsequent verification directions; if all three indicators are false, a continue search instruction is generated and the next round of search is started.