A multi-source evidence-based error and difficult problem management and elimination gating system

Abstract

The application discloses a kind of based on multi-source evidence's wrong difficult problem governance and elimination gate system, it is related to educational information processing and intelligent teaching auxiliary technical field, system acquisition learning result, process trajectory, prompt behavior, repeat, interval keep, migration verification, independent answer and paper back data, construct minimum elimination unit, generate tri-state support result and evidence conflict intensity, extract counter-evidence gap, restore time lag and challenge task mask, through prohibition elimination, backflow and elimination permission control wrong question pool and difficult problem pool double pool governance, task trigger and state write back.The system can pass through result chain, but there is prompt dependence, time length deviation, repeat failure, migration instability, independent answer failure or cross-end inconsistency object is identified as surface mastery risk object, to reduce the risk of result correct object being eliminated prematurely.

Description

Technical Field

[0001] This invention relates to the field of educational information processing and intelligent teaching assistance technology, specifically a gated system for error problem governance and elimination based on multi-source evidence. Background Technology

[0002] In a blended learning environment encompassing digital learning platforms, assignment systems, learning terminals, and paper-based feedback, the system typically continuously collects various types of learning evidence, including learning outcomes, answering processes, prompts, repetition feedback, phased reviews, transfer training, and independent answers. Based on this evidence, it provides services such as error correction, explanation, question assignment, or status assessment. Existing technologies supporting learning primarily fall into three categories.

[0003] The first type of approach focuses on addressing incorrect answers or weak knowledge points, generating subsequent training content based on the results of incorrect answers, knowledge point mappings, and question similarity relationships. For example, the publicly available solution with publication number CN111723193B organizes subsequent training paths around incorrect answers, knowledge point mappings, expansion of similar knowledge points, and control of recommendation ratios. This type of approach mainly focuses on explicit error objects, lacking independent governance granularity for objects where the result chain has been completed but still exhibit issues such as cue dependence, process gaps, unstable restatements, unstable intervals, unstable transfer, or failure to answer independently.

[0004] The second approach focuses on knowledge tracing and state estimation. It generates mastery or knowledge state predictions based on learners' historical interaction sequences, response records, and behavioral data, and then uses these predictions for assignment organization, personalized tutoring, and learning report generation. For example, the publicly available solutions CN111695779B and CN111695779A revolve around historical interaction sequences, attention modeling, and knowledge state prediction. This approach emphasizes improving the accuracy of state estimation and typically does not directly address the issues of gating and subsequent write-back when contradictions arise between result chains, process chains, paraphrasing chains, interval-keeping chains, transfer chains, and independent response chains.

[0005] The third approach centers on prompts, explanations, self-explanations, or coaching sessions, improving the learning process through prompt delivery, structured explanations, follow-up questions, or scaffolding-based coaching. For example, the publicly available text WO2008121323A1 and its corresponding publicly available solutions have already provided methods for organizing prompts, explanations, scaffolding, and teaching sessions. This type of approach typically uses restatements, explanations, and prompts for teaching intervention, but it lacks sufficient support for continued governance, status rewriting, and backflow control after the outcome chain has been completed.

[0006] In summary, while existing technologies can address specific issues such as error management, knowledge tracking, and explanatory intervention, they still have the following shortcomings: First, they primarily focus on those who answered incorrectly, making it difficult to independently manage those who have passed the result chain but have not yet achieved stable mastery. Second, although they can estimate the mastery status, they do not adequately utilize various conflicting pieces of evidence, making it difficult to promptly identify those who "appear to have passed but have not achieved stable mastery." Third, the prompting and repetition routes mainly serve teaching intervention, lacking a unified follow-up control mechanism for situations such as passing the repetition but failing to answer independently, or passing recently but subsequently becoming unstable. Fourth, in scenarios where paper-based feedback, terminal feedback, and manual verification coexist, there is a general lack of a unified closed-loop process for status writing, anomaly escalation, and re-entry management.

[0007] Therefore, after receiving multi-source learning evidence such as learning results, process trajectories, restatements, interval maintenance, transfer verification, independent responses, and paper feedback, how to more accurately distinguish between those who answered incorrectly and those who answered correctly but still have hidden risks, avoid the system ending governance prematurely, and continue state writing, compensation verification, and re-entering governance control when subsequent evidence becomes contradictory, unstable, or has low reliability feedback, has become a technical problem that needs to be solved in this field. Summary of the Invention

[0008] To address the shortcomings of existing error management, knowledge tracking, and explanatory tutoring technologies in terms of insufficient control over the subsequent process of the learning object, this invention aims to provide an error management and elimination gating system based on multi-source evidence. This system enables the learning object to make subsequent judgments on whether it has achieved stable mastery, even after the learning object has achieved the correct result, based on evidence such as process trajectory, prompting behavior, repetition performance, interval maintenance, transfer verification, independent answering, and written feedback. Furthermore, it prevents the system from prematurely ending the management process or maintaining the eliminated error state when there are conflicting pieces of evidence, gaps that have not been filled, or when the eliminated object becomes unstable again.

[0009] Another objective of this invention is to provide a state control chain built around false elimination suppression, enabling the system not only to identify incorrectly answered objects, but also to separate objects that have passed through the result chain but still have cue dependencies, process gaps, restatement fragility, interval instability, migration instability, or independent answer failures from general passed objects, forming independent governance objects, and thereby performing prohibition elimination, challenge verification, re-entry and state rewriting.

[0010] A further objective of this invention is to provide an implementation scheme that can accommodate system structure, computer decision chain, and anomaly recovery chain, enabling multi-source data in learning terminals, backend services, and paper-based data transmission environments to be event-driven, aligned, merged, gated, and written back at a unified object granularity, thereby improving the continuity, stability, and engineering feasibility of learning state control.

[0011] To achieve the above objectives, this invention uses the learning object as the basic unit of governance, multi-source learning evidence as input, the smallest elimination unit as the object base layer, the three-state support results and evidence conflict strength as the core identification results, the disproving gap, recovery delay and challenge task mask as the constraint results, and prohibition elimination, backflow and elimination permission as the core control results, forming a closed-loop control chain from evidence collection to state writeback.

[0012] This invention first collects learning result data, process trajectory data, prompting behavior data, repetition content data, interval maintenance data, transfer verification data, independent answering data, and paper-based feedback data, and then uniformly converts all types of records into evidence events oriented towards the learning object. By performing source normalization, time alignment, task identifier alignment, and knowledge anchor mapping on the evidence events, records from different sources can enter subsequent calculations under the same object, the same time window, and the same field caliber.

[0013] At the object-based layer, this invention constructs the smallest elimination unit based on the examination syllabus, curriculum standards, textbook catalog, historical exam questions, question bank data, collection of incorrect and difficult questions, teacher-uploaded materials, and learners' historical learning records. The smallest elimination unit not only inherits knowledge anchors, verification anchors, and source chains, but also standard expressions, standard processes, main error causes, current verification gaps, and current elimination status, used to ensure that subsequent state control chains are applied to a stable object granularity.

[0014] At the identification layer, this invention does not directly equate passing the result chain with stable mastery. Instead, it performs support allocation and multi-chain fusion on the result chain, process chain, paraphrase chain, interval maintenance chain, migration chain, and independent response chain to generate support strength for unmastered chain, apparent mastery support strength, stable mastery support strength, and evidence conflict strength. By explicitly retaining the evidence conflict strength, situations such as passing the result chain but failing independent response, passing the result chain but failing paraphrase probing, and passing recently but becoming unstable later can directly enter subsequent control.

[0015] At the constraint layer, this invention extracts disproven gaps from five directions: steps, restatements, intervals, transitions, and independent responses. It also extracts recovery delays from three directions: recovery after prompting, recovery after intervals, and recovery after transitions, thereby generating a challenge task mask. The challenge task mask is used to indicate which type of key verification should be prioritized for the current object, so that subsequent task triggering is no longer generalized reasoning, but targeted verification around the premise of approval.

[0016] At the control layer, this invention sets prohibition elimination results, reflow results, and elimination permission results. The prohibition elimination result determines whether the current object should remain in the governance chain; the reflow result determines whether the eliminated object needs to have its eliminated state revoked and re-enter governance; and the elimination permission result determines whether the object meets further release conditions after the blocking conditions are lifted. The system does not directly equate "not blocked" with "elimination allowed," but rather treats prohibition elimination, reflow, and elimination permission as a group of control results acting together on state transitions. If necessary, a set of allowed actions and elimination certificates can be further introduced as an additional release layer to further distinguish between "not blocked" and "released."

[0017] At the execution layer, this invention triggers structured paraphrasing tasks, step completion tasks, interval maintenance tasks, transition verification tasks, and independent response verification tasks based on the dual-pool governance structure, challenge task mask, and current state results. It then writes the subsequent feedback results back to the state object, thereby forming a closed loop of "risk identification, constraint generation, verification execution, state write-back, and re-control".

[0018] To achieve the above overall concept, the present invention provides a gated system for error problem governance and elimination based on multi-source evidence. The system includes at least an evidence collection and preprocessing unit, a minimum elimination unit construction unit, a latent unknowable identification unit, a counter-evidence constraint generation unit, an asymmetric gating unit, a dual-pool governance unit, and a state write-back unit.

[0019] The evidence collection and preprocessing unit is used to collect learning result data, process trajectory data, prompting behavior data, repetition content data, interval maintenance data, transfer verification data, independent answering data, and paper-based feedback data, and converts the data into evidence events oriented towards the learning object. Each evidence event includes at least a source identifier, a time identifier, a task type identifier, a knowledge anchor identifier, a prompt level, process completeness, duration deviation, expression independence, and result destination identifier. The evidence collection and preprocessing unit is also used to perform source normalization, time alignment, task identifier alignment, and knowledge anchor mapping on the evidence events, so that records from different sources can participate in subsequent judgments under a unified object caliber; among them, paper-based feedback data, after being eventified, can be used as evidence input for result chains, process chains, independent answering chains, or duration deviation risk judgments.

[0020] The minimum elimination unit construction unit is used to construct minimum elimination units corresponding to learning objects based on the examination syllabus, curriculum standards, textbook catalog, historical exam questions, question bank data, collections of incorrect and difficult questions, teacher-uploaded materials, and learners' historical learning records. Each minimum elimination unit includes at least a knowledge anchor identifier, a task type identifier, a source chain, standard expression key points, standard process key points, a verification anchor point, a current verification gap, and a current elimination status. Furthermore, the minimum elimination unit construction unit can also perform retention, splitting, and merging processes on cross-source knowledge candidate items based on source support vectors, source conflict matrices, and individual exposure frequency vectors to improve the matching degree between the object's basic layer and actual governance actions.

[0021] The implicit lack-of-knowledge identification unit is used to generate three-state support results and evidence conflict strength based on the result chain, process chain, paraphrase chain, interval maintenance chain, transfer chain, and independent response chain. The implicit lack-of-knowledge identification unit maps the result chain to "result passed, to be distinguished" support, maps stable independent response, resistance to probing, stable interval maintenance, and stable transfer verification to "stable mastery" support, maps increased cue dependence, insufficient process completeness, structured paraphrase failure, and time deviation reaching a threshold to "superficial mastery" risk support, and maps independent response failure, resistance to probing failure, interval instability, and transfer instability to "lack of mastery" risk support. It also retains the quantities formed by mutually exclusive support as evidence conflict strength. Thus, the system can distinguish between "lack of knowledge," "superficial mastery," and "stable mastery," and directly input conflict results into subsequent gating chains.

[0022] The disproving constraint generation unit is used to extract the disproving gap fingerprint, recover the time-delay signature, and generate a challenge task mask. The disproving gap fingerprint includes at least the step gap amount, the restatement gap amount, the interval gap amount, the transition gap amount, and the independent response gap amount; the recovery time-delay signature includes at least the delay from passing after prompting to passing independently, the number of rounds required for recovery after the interval, and the number of rounds required for transition recovery. The challenge task mask is generated based on the three-state support result, the evidence conflict strength, the disproving gap fingerprint, the recovery time-delay signature, and the risk trigger code, and is used to instruct the current learning object to prioritize triggering at least one of the following: independent response verification, anti-follow-up questioning restatement verification, interval preservation verification, transition variation verification, and step completion verification.

[0023] The asymmetric gating unit generates prohibition elimination results, reflow results, and elimination permission results based on the three-state support results, evidence conflict strength, disproving gap fingerprint, recovery delay signature, and challenge task mask. The asymmetric gating unit outputs a prohibition elimination result when at least one of the following conditions is met: the apparent mastery support strength reaches the prohibition elimination threshold; the evidence conflict strength reaches the prohibition elimination threshold; the disproving constraint strength reaches the prohibition elimination threshold; the duration deviation risk reaches the threshold; the result chain passes but independent response fails; the cue dependency reaches the threshold and the restatement vulnerability reaches the threshold. After the learning object is already in the eliminated state, the asymmetric gating unit outputs a reflow result if at least one of the following conditions is met: the evidence conflict strength reaches the reflow threshold; the interval descent reaches the threshold; the migration descent reaches the threshold; independent response fails again; there is cross-terminal inconsistency between paper-based response data, terminal response data, or manually verified data, reaching an anomaly threshold. Furthermore, the asymmetric gating unit can also generate a set of allowed actions based on the action inconsistency degree and the threshold corresponding to the calibration sample group, and generate an elimination certificate when the set of allowed actions includes elimination actions, does not include hold actions and does not include backflow actions, and the coverage of the key verification chain corresponding to the challenge task mask reaches the target threshold, and use it as an additional restriction condition for the elimination permission result.

[0024] The dual-pool governance unit is used to write learning objects into the incorrect question pool when the result chain fails, and to write them into the challenging question pool when the result chain passes but there are risks such as apparent mastery, evidence conflict, disproving evidence gaps, recovery lag, time deviation, independent answer failure, and structured paraphrase failure. The dual-pool governance unit also organizes structured paraphrase tasks, interval maintenance tasks, transfer verification tasks, and independent answer verification tasks corresponding to the challenge task mask, based on the knowledge anchor point, verification anchor point, and main error causes corresponding to the smallest elimination unit. In this way, incorrectly answered objects and seemingly passed objects are assigned different governance entry points, different elimination thresholds, and different reflow sensitivities within the system.

[0025] The state write-back unit updates the eliminated state when the prohibition elimination result indicates that progress is allowed, the reflow result indicates that reflow is not required, and the elimination permission result indicates that elimination is allowed. Upon the reappearance of at least one of the following subsequent evidence, the learning object is written back to the stabilizing state and re-enters the resolving state: increased conflict intensity; interval instability; migration instability; failure to respond independently; cross-end inconsistency reaching an anomaly threshold. Simultaneously, the main error cause, the most recent failure reason, and the reflow identifier are written back. With this state write-back chain, the system can not only complete a one-time judgment but also continuously use subsequent evidence to correct the formed state results.

[0026] Furthermore, the minimum elimination unit construction unit can perform more fine-grained adjustment processing on cross-source knowledge candidate items. When multiple sources provide different support for the conceptual boundaries, verification methods, or question type attribution of the same knowledge item, the system uses source support vectors, source conflict matrices, and individual exposure frequency vectors to comprehensively determine whether the item should be retained as a single minimum elimination unit, split into multiple independent minimum elimination units, or merged into adjacent units, so that subsequent gating chains and task chains can revolve around stable governance objects.

[0027] Furthermore, the asymmetric gating unit can perform aggregate calculations on the disproving gaps and recovery delays, and combine this with the strength of stable control support, apparent control support, evidence conflict strength, and key verification chain coverage to generate an action inconsistency degree. The system then establishes acceptance thresholds for candidate actions such as maintaining the current governance status, continuing challenge verification, allowing elimination, and forced repatriation based on the distribution results of historical stable samples, historical samples awaiting stabilization, and historical repatriation samples in the calibration sample group, thereby forming a set of permissible actions. Through this scheme, the system can transform the determination of "which subsequent action is most suitable at present" from a purely empirical judgment into a calibrable action selection process.

[0028] Furthermore, the system can generate a clearance certificate based on the set of allowed actions. The system only outputs a clearance certificate when the set of allowed actions includes allowed clearance actions, excludes hold actions and reflow actions, and the critical verification chain coverage corresponding to the challenge task mask reaches the target threshold. This scheme can further distinguish between "the main gating is not blocked" and "the object already has sufficient conditions for release," making it suitable for scenarios requiring stricter control over false releases.

[0029] Furthermore, the system can perform unified quantitative processing on abnormal feedback, low-reliability feedback, and cross-platform inconsistent feedback. The system calculates feedback completeness based on the number of feedback questions, coverage of key steps, and timestamp completeness; it calculates feedback reliability based on feedback completeness, cross-platform consistency, timing reasonableness, and abnormal jumps; and further combines interval drop-down, migration drop-down, and independent answer failure flags to form an unstable aggregation quantity. Then, based on whether the object is in an eliminated state, whether the unstable aggregation quantity reaches the feedback threshold, and whether the feedback reliability reaches the reliability lower limit, it outputs a forced feedback result or a downgraded spot check result. Through this scheme, the system can maintain the stability of state control in an environment where paper-based feedback, terminal feedback, and manual verification coexist.

[0030] Furthermore, the dual-pool governance unit can generate a paraphrase gating result when the structured paraphrase verification passes. The removal of the prohibition-triggered elimination restriction is only permitted if the paraphrase gating result indicates that the paraphrase has passed and is resistant to follow-up questioning, the key verification chain coverage corresponding to the challenge task mask meets the conditions, and at least one of the independent response chain, the interval-maintaining chain, and the migration chain passes again. Thus, paraphrase verification is limited to a high-value auxiliary verification step, rather than a separate elimination completion credential.

[0031] Furthermore, the asymmetric gating unit can determine the gating thresholds corresponding to the strength of apparent control support, the strength of evidence conflict, the risk of duration deviation, the interval sag, and the migration sag based on the historical stable sample set, the historical re-entry sample set, the false elimination loss, the overvalidation loss, the target false elimination rate, and the target false retention rate. The thresholds are updated upon the addition of new re-entry samples. This scheme allows the thresholds used to prohibit elimination and re-entry to be updated as the sample distribution changes, rather than relying on fixed empirical values.

[0032] The beneficial effects of this invention are: By unifying learning outcome data, process trajectory data, prompting behavior data, retelling content data, interval-maintaining data, transfer verification data, independent response data, and paper-based feedback data into evidence events, and performing source unification, time alignment, task identifier alignment, and knowledge anchor mapping on these evidence events, the problem of fragmented records from different sources and difficulty in continuous use on the same governance object in existing technologies can be overcome. Since all types of evidence are written into structurally consistent event objects before entering subsequent judgment, subsequent state control can compare result chains, process chains, retelling chains, and independent response chains within the same time window, the same object granularity, and the same field caliber. This reduces the accumulation of misjudgments caused by inconsistent sources and improves the continuity and engineering feasibility of multi-source learning evidence participating in subsequent diagnosis and write-back.

[0033] By constructing the smallest elimination unit based on the examination syllabus, curriculum standards, textbook catalog, past exam questions, question bank data, collections of incorrect and difficult questions, teacher-uploaded materials, and learners' historical learning records, and performing retention, splitting, and merging processes on cross-source knowledge candidate items, this approach can improve the problem of overly coarse or overly fine object granularity in existing error governance schemes. Existing technologies often directly use questions or broad knowledge points as subsequent governance objects, which easily leads to multiple error causes being mixed in the same object, or the same governance action failing to accurately target the actual weak knowledge segments. This application, by converging the source chain, standard expression key points, standard process key points, verification anchor points, and current verification gaps into the smallest elimination unit, ensures that subsequent dual-pool governance, task generation, and state write-back revolve around the same basic object, improving the accuracy of object segmentation and the targeting of subsequent control chains.

[0034] By generating a three-state support result comprising support strength for unacknowledged knowledge, support strength for apparent knowledge, support strength for stable knowledge, and strength of evidence conflict, this approach overcomes the limitations of existing knowledge state estimation methods that tend to either "leave the process if the result is correct" or "output only a single knowledge score." While existing technologies can estimate knowledge levels based on historical interactions, they often lack a separate state structure to characterize the risk position of objects that still exhibit cue dependence, interval instability, transfer failure, or independent response failure after the outcome chain has passed. This application retains apparent knowledge support and strength of evidence conflict, enabling the system to distinguish between objects that pass on the first attempt and those with stable knowledge, and directly transforming mutually exclusive evidence into subsequent control inputs, thus reducing the risk of premature termination of governance for objects that have passed the outcome chain.

[0035] By extracting the fingerprint of the disproving gap, recovering the time-delay signature, and generating a challenge task mask, this approach addresses the problem in existing follow-up practice organization where users "only know to continue practicing, but don't know which type of verification to supplement." Existing recommended or tutoring routes typically generate follow-up tasks based on historical errors or general difficulty, but they often lack a unified identifier to determine whether the current object lacks step completion, structured paraphrasing, interval preservation, transfer verification, or independent response evidence. This application converges step gaps, paraphrasing gaps, interval gaps, transfer gaps, and independent response gaps into a disproving gap fingerprint, and then incorporates the lag in recovery after prompts, recovery after intervals, and transfer recovery into the recovery time-delay signature. This generates a challenge task mask corresponding to the key verification chain, improving the relevance of follow-up task generation and the ability to identify false elimination risks.

[0036] By setting up an asymmetric gating chain that generates prohibition elimination results, reflow results, and elimination permission results based on three-state support results, evidence conflict strength, counter-evidence gap fingerprints, recovery delay signatures, and challenge task masks, the control gap in existing technologies, where "not being blocked is considered as allowing passage," can be improved. Since prohibition elimination, reflow, and elimination permission are separated into different control results, the system can separately handle whether the current round should be blocked, whether eliminated objects should be revoked, and when entry into the eliminated state is allowed, rather than compressing all judgments into a single conclusion. This control method allows objects that pass through the result chain but fail to respond independently, are fragile in restatement, have unstable intervals, unstable migrations, or have abnormal duration deviations to remain in the stabilization-awaited state, and continue to trigger reflow when subsequent evidence re-exposes risks, improving the stability of state control and the ability to suppress false releases.

[0037] By adding learning objects to the incorrect question pool when the result chain fails, and adding them to the difficult question pool when the result chain passes but there are risks such as apparent mastery, evidence conflict, disproving evidence gaps, recovery lag, duration deviation, independent answer failure, or structured paraphrasing failure, and then performing state write-back and re-entry when subsequent evidence shows increased conflict intensity, interval instability, transfer instability, independent answer failure, or cross-end inconsistency reaching an abnormal threshold, this approach can improve the existing system's handling of mixed incorrect and "apparently passed" objects. Existing approaches typically only focus on continuing training on incorrect questions, making it difficult to track the re-exposure risk of correctly answered objects in subsequent independent environments over the long term. This application binds the object entry point, subsequent tasks, elimination threshold, and re-entry conditions into the same closed loop through dual-pool governance and state write-back. This allows paper-based feedback, system-side verification, and subsequent independent responses to continuously participate in state updates. Furthermore, it enables the observation of governance effectiveness based on metrics such as false elimination rate, unannounced independent pass rate, abnormal return rate, and erroneous write-back rate. This improves the closure of subsequent governance, the reliability of abnormal recovery, and the continuous verifiability of eliminated states. Attached Figure Description

[0038] Figure 1 This is the main flowchart of the error handling and elimination gate control of the present invention; Figure 2 This is a flowchart illustrating the enhanced release control of the present invention. Figure 3 This is a flowchart of the abnormal return and reflow recovery process of the present invention. Detailed Implementation

[0039] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Example 1 like Figure 1 As shown, this embodiment illustrates a basic implementation of a gated system for error resolution and elimination based on multi-source evidence. This embodiment is deployed in a computer environment where a learning terminal and a backend service operate collaboratively. The learning terminal collects front-end records during the learner's answering, retelling, interval reproduction, and independent answering verification processes. The backend service performs evidence eventification, minimum elimination unit construction, three-state support result generation, contradictory evidence constraint generation, asymmetric gating, dual-pool governance, and state write-back. The backend service can be deployed on a single-machine server, container cluster, or cloud computing node. The terminal and the backend service exchange evidence events, task instructions, and state results through task interfaces and state write-back interfaces.

[0041] The system in this embodiment includes an evidence collection and preprocessing unit, a minimum elimination unit construction unit, a latent lack of knowledge identification unit, a counter-evidence constraint generation unit, an asymmetric gating unit, a dual-pool governance unit, and a state write-back unit. The evidence collection and preprocessing unit extracts raw data from test paper answer records, assignment records, prompt call records, structured paraphrasing records, interval repetition records, transfer verification records, independent answer records, and paper-based feedback records, and uniformly transcribes them into evidence events; the minimum elimination unit construction unit establishes knowledge anchors and verification anchors; the latent lack of knowledge identification unit generates three-state support results and evidence conflict strength; the counter-evidence constraint generation unit generates counter-evidence gaps, recovery delays, and challenge task masks; the asymmetric gating unit generates prohibition elimination results, reflow results, and elimination permission results; the dual-pool governance unit writes learning objects into the wrong question pool or the difficult question pool based on the gating results; and the state write-back unit is responsible for writing the most recent failure reasons, reflow identifiers, main error causes, and the current stage back to the background state object.

[0042] In this embodiment, evidence events are stored using a unified object structure. Each evidence event includes at least the following fields: learning object identifier, learner identifier, source identifier, time identifier, task type identifier, knowledge anchor identifier, verification anchor identifier, result status field, result destination identifier field, prompt level field, process completeness field, duration deviation field, expression independence field, and feedback credibility field. The learning object identifier is part of the primary key. The source identifier and time identifier are used together to eliminate duplicate feedback. The knowledge anchor identifier and verification anchor identifier are used to project the event to the corresponding smallest elimination unit. The result status field records at least three values: passed, failed, and undetermined. The result destination identifier field records the result chain destination of the evidence event within the current determination window, including at least three values: result failed, result passed but to be distinguished, and result invalid. The prompt level field uses a discrete level from zero to three, representing no prompt, light prompt, medium prompt, and strong prompt, respectively. The process completeness field uses a continuous value between zero and one, with a larger value indicating more comprehensive coverage of key steps. The duration deviation field uses a normalized positive value; a larger value indicates a more significant deviation from the baseline duration of similar tasks. The expression independence field uses a continuous value between zero and one; a larger value indicates a more independent restatement or response. The feedback credibility field is used to distinguish the credibility level of feedback from direct system connection, paper-based recognition, and manual verification.

[0043] In this embodiment, the minimum elimination unit is maintained as the object foundation layer for all subsequent governance actions. Each minimum elimination unit includes at least a unit identifier, a knowledge anchor identifier, a task type identifier, a source chain field, a standard expression field, a standard process field, a verification anchor field, a field for easily confused edges, a field for main error causes, a current verification gap field, a current elimination status field, and an update time field. The current verification gap field records the current value of step gaps, restatement gaps, interval gaps, migration gaps, and independent response gaps. The current elimination status field includes at least four states: not started, being resolved, awaiting stabilization, and eliminated. To meet the object disclosure depth requirements, the minimum elimination unit object in this embodiment also records a version field and a source batch field. The version field increments when splitting, merging, or rearranging verification anchors, and the source batch field records the data construction batch corresponding to the current unit for rollback verification during subsequent backflow or reconstruction.

[0044] In this embodiment, the error cause vector, verification gap vector, and elimination state vector are all implemented using fixed-dimensional structured objects. The error cause vector uses a five-dimensional structure, corresponding to basic missing information, step missing information, memory instability, insufficient transfer, and exposure of independent responses. The verification gap vector also uses a five-dimensional structure, corresponding to step gaps, rehearsal gaps, interval gaps, transfer gaps, and independent response gaps. The elimination state vector uses a twelve-dimensional structure, corresponding to the current stage, consecutive pass count, independent pass count, interval pass count, transfer pass count, prompt level, rehearsal gating level, evidence conflict level, recent failure reason, backflow identifier, recent update time, and priority. The dimensions of these vectors are fixed and written to the configuration object during system initialization, with initial values ​​set to all zeros. Updates are event-triggered: when a new evidence event arrives, only the corresponding dimension is updated based on the evidence chain type, without changing irrelevant dimensions; when a forced backflow occurs, the backflow identifier dimension is set to one, and the current stage dimension is rewritten as pending stabilization or being resolved; when there is no new evidence input for several consecutive periods, only the time field is updated, without actively raising the mastery status.

[0045] In this embodiment, the meaning and source of each formula variable are disclosed in the same paragraph. Among them, the variables related to support allocation are mainly derived from the evidence event object, the variables related to gaps are mainly derived from the minimum elimination unit and the feedback of the review task, the variables related to time delay are mainly derived from the statistical results of the time window, and the variables related to gating are mainly derived from the three-state support results, the evidence conflict intensity, and the feedback results of independent responses.

[0046] Working Process: The system first receives raw records from learning terminals, online learning platforms, and paper-based feedback interfaces, and merges them according to the learner's identifier and time window. For the same learner, six types of evidence chains are extracted first within a decision window: result chain, process chain, restatement chain, interval-keeping chain, transfer chain, and independent response chain. The result chain describes whether the conclusion of the question is correct; the process chain describes whether the key steps are complete; the restatement chain describes whether the learner can independently explain the question without a displayed answer; the interval-keeping chain describes whether the learner can still pass the question stably after a preset interval; the transfer chain describes whether the learner can still complete the question when faced with variant tasks; and the independent response chain describes whether the learner can answer the question independently under weak or no prompts. Each evidence chain is converted into an input that supports allocation, and the results of the allocation are used for subsequent three-state fusion.

[0047] In this embodiment, the system classifies learning objects into three states: unmastered, superficially mastered, and stably mastered. For any chain of evidence, the support allocation for different subsets of work is calculated based on the pass flag, prompt level, process completeness, expression independence, and duration deviation in the evidence event. To avoid the result chain being directly interpreted as stably mastered, this embodiment allocates support according to the type of evidence chain. The support allocation for a single chain is calculated according to formula (1): in, Indicates the chain of evidence index. Represents a subset of the working set. This indicates stable control over a single-state subset. This indicates a subset of results that have been passed but have not yet been distinguished as apparent mastery or stable mastery. This indicates signs of instability, but a distinction has not yet been made between subsets that are not yet fully understood or only superficially understood. This indicates that a single-state subset is not known. Indicates the remaining uncertain support level. This represents the credibility coefficient of the evidence chain, with a value ranging from 0.6 to 1.0. This indicates the strength of the chain of evidence, with a value ranging from zero to one. This indicates a dependency reduction amount, with a value ranging from zero to 0.8. Indicates the severity of the failure, with a value ranging from zero to one. This indicates the normalized risk intensity corresponding to the following conditions: cue dependency, process gap, paraphrase failure, interval instability, migration instability, failure to answer independently, or time deviation reaching a threshold. The value ranges from zero to one. This indicates that the result chain passes through a set of evidence. This represents the set of evidence indicating a failure in the outcome chain. This indicates a set of evidence that demonstrates independent responses, resistance to follow-up questioning, and stable pass rates based on interval maintenance or transfer verification. This represents a set of evidence related to risk exposure. This represents the set of working subsets of the evidence chain that allow for the allocation of support. For the result chain, if the object answers correctly, only the support is written. Instead of writing directly For independent response chains, paraphrasing chains, interval-keeping chains, and migration chains, if they pass stably under conditions with no or weak prompts, then priority should be given to writing support. For risk events such as increased reliance on prompts, incomplete processes, failed retelling, decline after intervals, failed transitions, failed independent responses, or abnormal time deviations, priority should be given to writing support into these areas. or .

[0048] After completing the single-chain support allocation, the system performs chain-by-chain iterative fusion on multiple evidence chains for the same learning object: only support items whose intersection equals a subset of the target working set are accumulated to obtain the unmastered support strength, apparent mastery support strength, and stable mastery support strength. The fusion order prioritizes the result chain and the process chain, followed by the paraphrased chain, the interval-maintaining chain, the transfer chain, and the independent response chain.

[0049] To avoid smoothing out mutually exclusive evidence, this embodiment does not perform normalization elimination on mutually exclusive supports. Instead, it synchronously accumulates mutually exclusive support terms with empty intersections during the fusion process to obtain the evidence conflict strength. If the result chain shows success but the independent response chain shows failure, or if the result chain shows success but the restatement chain fails in the anti-follow-up questioning phase, then the corresponding mutual exclusion support will be directly increased. In this embodiment, The value ranges from one to zero, with 0.6 set as the initial high-conflict warning value. This warning value is a parameter to be calibrated, and its initial value comes from the comparison of quantiles of historical stable samples and historical backflow samples. It can be updated in batches according to different disciplines or task families.

[0050] After obtaining the three-state support results and the strength of evidence conflict, the system further calculates the step gap. restate the gap amount Interval gap amount Migration gap Independent response gap And the delay between passing the prompt and passing independently. The number of rounds required to restore the value after the interval is normalized. Normalized value of the number of rounds required for migration recovery The strength of the contradiction constraint is obtained by summing the results according to the preset weights. The above weights use calibrable parameters, which can be initially set to 0.15 for each of the step gap and independent response gap, 0.1 for the restatement gap, 0.1 for each of the interval gap and migration gap, and 0.125 for each of the three types of time delays. They can be adjusted later based on the actual sample distribution.

[0051] In this embodiment, the system further generates a rehearsal vulnerability based on the amount of rehearsal gap, the anti-probing failure flag, and the completion status of rehearsal after prompting. The rehearsal vulnerability uses a continuous value between zero and one; a larger value indicates that the object is more likely to expose a risk of superficially passing but not achieving stable mastery during structured rehearsal or anti-probing stages. When the amount of rehearsal gap exceeds the rehearsal threshold, or the anti-probing failure flag is one, the rehearsal vulnerability enters a high-risk range; when the object completes structured rehearsal without an answer being displayed and passes anti-probing, the rehearsal vulnerability decreases. The system also generates risk trigger codes based on situations such as conflicts between the result chain and the independent response chain, conflicts between the result chain and the rehearsal chain, a downward trend in the interval chain and the migration chain, and increased prompting dependence with prominent process gaps. Risk trigger codes include at least one of the following: a conflict code between the result chain and the independent response chain, a conflict code between the result chain and the rehearsal chain, a code for instability due to resonance between the interval and the migration chain, and a code for the superposition of prompting dependence and process gaps. These codes are used to mark that the current object should have increased priority for prohibition elimination, re-entry, or challenge verification.

[0052] In this embodiment, the prohibition elimination result is generated using a hard-condition primary criterion. The system first determines whether to trigger blocking based on the strength of apparent support, the strength of evidence conflict, the strength of counter-evidence constraints, the risk of duration deviation, and key failure events. The prohibition elimination criterion is calculated according to formula (2): in, Represent the learning object The result of the prohibition elimination is that a value of 1 indicates that entry into the eliminated state is prohibited, and a value of 0 indicates that the current gate control will not directly block it. Indicates the surface's support strength; Indicates the strength of the conflict of evidence; Indicates the strength of the counter-evidence constraint; This indicates a risk of duration deviation; This indicates an independent failure flag after the result chain has passed; a value of 1 indicates failure and a value of 0 indicates success. This indicates a failure flag for the anti-follow-up paraphrase after the result chain has been passed; a value of 1 indicates failure, and a value of 0 indicates success. , , and This indicates the corresponding threshold.

[0053] For objects that are already in the eliminated state, the system further calculates the reflow criterion. The reflow criterion is calculated according to formula (3): in, Represent the learning object The return flow result is one, which indicates that the return flow has been initiated from the eliminated state to the state to be stabilized or is being resolved, and a value of zero indicates that the return flow is not triggered. Indicates the interval of decline; This indicates the amount of migration decline; This indicates that the attempt to answer independently failed again. This indicates a cross-platform inconsistency between paper-based feedback, terminal feedback, or manual verification. , and This indicates the corresponding threshold. The criterion is that if an eliminated object exposes any high-risk instability event, it can trigger a re-entry.

[0054] The system does not directly interpret "not blocked" as "cancellation allowed," but instead generates a cancellation permission result in addition to the cancellation prohibition criterion and the backflow criterion. The cancellation permission criterion is calculated according to formula (4): in, Represent the learning object The result of the removal of the license; This indicates the coverage of the key verification chain corresponding to the challenge task mask; Indicates the critical coverage threshold; , and These indicate whether the independent response chain, the interval-maintaining chain, and the migration chain have met the release conditions. The system only allows an object to enter the eliminated state if it is not blocked, has not triggered a backflow, and the critical verification chain coverage meets the requirements.

[0055] After the strength of the counter-evidence constraint and the strength of the evidence conflict are calculated, the system generates a challenge task mask based on the source of the gap, the type of conflict, and the risk trigger code. If the step gap is the largest, the step completion task is triggered first; if the restatement gap is the largest, the resistance to follow-up questioning failure flag is one, or the restatement vulnerability reaches the restatement threshold, the structured restatement task is triggered first; if the interval gap is the largest, the interval maintenance task is triggered first; if the migration gap is the largest, the migration verification task is triggered first; if the independent response gap is the largest, the hard conflict flag is one, or the conflict code between the result chain and the independent response chain is triggered, the independent response verification task is triggered first. The challenge task mask is stored in a background state object, with fields including task type, candidate task sequence, priority field, trigger reason field, generation time field, and expiration time field. The priority field is jointly determined by the strength of the counter-evidence constraint, the strength of the evidence conflict, the risk trigger code, and the current stage of the object. The update method is to recalculate the entire task when a new event arrives. If the current task has not expired and the new priority does not exceed the original priority, the original task remains unchanged.

[0056] During the dual-pool governance phase, the system first checks whether the result chain passes. If the result chain fails, the learning object is directly written into the error pool; if the result chain passes, but the calculation result of formula (2) is one, or the time deviation risk reaches the threshold, or the independent answer failure flag is one, or the anti-questioning rehearsal failure flag is one, the learning object is written into the problem pool. The error pool object mainly enters the error correction and step completion path, and the problem pool object mainly enters the pending stabilization, structured rehearsal, interval maintenance, transfer verification and independent answer kernel path. When writing to the dual pool, the system uses the learning object identifier and the minimum elimination unit identifier as key fields, and the pool type, main error cause, recent failure reason, current stage, reflow identifier and update time as value fields. If the same object already has a pool record in the current time window, the system only updates the value field and retains the creation time; if the object has been marked as eliminated but the calculation result of formula (3) is one, the reflow identifier and reflow batch field are written back at the same time.

[0057] During the state write-back phase, the system maintains a state write-back object for each learning object. This object includes at least the learning object identifier, current stage field, current pool type field, unmastered support strength field, apparent mastered support strength field, stable mastered support strength field, evidence conflict strength field, counter-evidence constraint strength field, prohibition elimination result field, reflow result field, elimination permission result field, recent failure reason field, reflow identifier field, update time field, and version field. The update rules for the current stage field are as follows: when the object first enters the governance chain, it is set to "in progress"; when the result chain passes but the prohibition elimination result is one, it is set to "pending stabilization"; when the object passes at least one of the key verifications in the independent response kernel, interval maintenance, and migration verification, and the prohibition elimination result is zero and the elimination permission result is one, it is allowed to transition to "eliminated"; when the eliminated object again meets the reflow condition of formula (3), it is written back to "pending stabilization". The version field is incremented during each stage migration to prevent old task reflow from overwriting the new state.

[0058] In this embodiment, the abnormal events include at least the following: missing evidence anomaly, incomplete return transmission anomaly, cross-end record inconsistency anomaly, and low-confidence return transmission anomaly. Anomaly judgment adopts a unified priority processing method. The priorities, from highest to lowest, are: cross-end record inconsistency anomaly, incomplete return transmission anomaly, low-confidence return transmission anomaly, and missing evidence anomaly. The threshold source uses the anomaly threshold field in the configuration object, where a return transmission completeness of less than 0.7 is considered incomplete, a cross-end key field consistency rate of less than 0.6 is considered cross-end record inconsistency, and a return transmission confidence of less than 0.5 is considered low-confidence return. The recovery condition adopts a three-option approach: "successful resampling, completed re-retransmission, or passed manual verification." When any recovery condition is met, the anomaly flag is removed, and re-entry into the main chain computation is allowed; otherwise, only the pending stabilization state or triggering a low-risk spot check is allowed, and entering the eliminated state is not permitted.

[0059] A specific execution scenario of the work process is as follows. A learner gives the correct answer to a mathematical derivation problem, but the process record shows that the key intermediate derivation steps are missing, the prompt level field is medium prompt, the anti-probing answer in the structured paraphrase fails, the actual answer time is significantly higher than the benchmark time of similar tasks, and then fails again in the independent answer task without prompts. The system first writes the result chain support of the object into the result passed partition subset, then writes the process chain and paraphrase chain support into the instability sign subset, writes the independent answer chain failure support into the unmastered risk subset, and writes the time deviation event into the apparent mastery risk subset. After the multi-chain fusion, the apparent mastery support strength increases, the evidence conflict strength increases at the same time, and the counter-evidence constraint strength exceeds the preset threshold due to the superposition of step gap, paraphrase gap and independent answer gap. Formula (2) outputs the prohibition elimination result as one. Based on this, the system writes the object into the problem pool, writes it back to the current stage as to be stabilized, and generates the structured paraphrase task and the independent answer kernel task. If the system consistently passes the interval maintenance task and the independent response task, it will recalculate the three-state support result, the prohibition elimination result, and the elimination permission result. The status will only be updated to eliminated if the prohibition elimination result is zero and the elimination permission result meets the conditions.

[0060] In this embodiment, different evidence chains are mapped to support allocations, the conflict strength of mutually exclusive evidence is preserved, and the rebuttal gap and recovery delay are uniformly introduced into the blocking criterion. After the result chain passes, the system first generates a three-state support result, then calculates the prohibition elimination result according to formula (2), calculates the return result of the eliminated object according to formula (3), and determines whether the final release condition is met according to formula (4). For objects that answer correctly after prompting, are unstable in restatement, decline after interval, have abnormal time deviation, or fail to answer independently, the system maintains the state of waiting to be stabilized or continuing verification through the evidence conflict strength, rebuttal constraint strength, and key verification chain coverage requirements, so as to reduce the probability of prematurely updating the eliminated state due to a single correct answer.

[0061] The parameter gap list in this embodiment includes the prohibition elimination threshold, initial values ​​of the credibility coefficients of each evidence chain, weight coefficients of the rebuttal constraint, anomaly criterion threshold, and key verification coverage threshold. For these parameters, the system can use an offline historical sample calibration method during initial deployment: first, select the set of objects that have been stably eliminated and the set of objects that have experienced re-entry, and then calculate the quantile intervals of the corresponding indicators. Next, write the threshold that best distinguishes the two types of objects into the configuration object. If sufficient historical samples are lacking, a conservative initial value is used, and updates are performed in batches after at least one teaching cycle. The update trigger condition is that the cumulative number of new re-entry samples reaches a preset number or the false release rate of a certain subject task family exceeds the target upper limit for two consecutive update cycles. The update frequency can be executed weekly or monthly. After the update is completed, the new version field of the configuration object is written back, and it takes effect in subsequent main chain calculations.

[0062] Example 2 like Figure 1 and Figure 2 As shown, this embodiment illustrates an enhanced implementation based on Embodiment 1. The system continues to use the evidence collection and preprocessing unit, implicit lack-of-knowledge identification unit, contradictory evidence constraint generation unit, asymmetric gating unit, dual-pool governance unit, and state write-back unit from Embodiment 1, and strengthens the cross-source construction of the minimum elimination unit, the sampled updating of the gating threshold, the generation of the allowed action set, and the additional release control for certificate elimination. This embodiment is suitable for computer implementation environments with significant differences in question types, task families, and disciplines, requiring improved release stability under different groupings.

[0063] In this embodiment, the minimum elimination unit construction unit performs retention, splitting, and merging processes on cross-source knowledge candidate entries. To this end, the system constructs a source support vector, a source conflict matrix, and an individual exposure frequency vector for each candidate entry. The source support vector is a five-dimensional structure, corresponding to the support of the exam syllabus, curriculum standards, past exam questions, question bank, and individual historical exposure, respectively. The source conflict matrix is ​​a five-by-five square matrix used to record the conflict intensity between different sources in terms of concept boundaries, verification methods, or question type attribution. The individual exposure frequency vector is a three-dimensional structure, corresponding to the number of repeated error exposures, the number of repeated instabilities, and the number of independent answer failures, respectively. These vectors and matrices are created during object initialization and updated after new source records are added or after the statistical period ends.

[0064] In this embodiment, the cross-source alignment score of candidate knowledge entries is calculated according to formula (5): in, Indicates candidate entries Cross-source alignment score, , , These represent cross-source consistency support strength, historical examination frequency, and individual exposure intensity, respectively. Indicates the cost of cross-source conflict. This represents the penalty amount for an entry's granularity deviating from the target granularity. to This represents the weighting coefficients, and their sum is one. Wherein, Derived from source support vectors, The data comes from statistics and counts of past exam questions and question banks. Derived from individual exposure frequency vectors, Originating from the source conflict matrix, The deviation calculation results are derived from the entry length, number of concepts, and number of validation anchors. The system will... Candidate entries with scores above a preset threshold and cross-source conflict costs below a preset upper limit are retained as independent minimum elimination units; candidate entries with high scores but excessive granularity penalties are split; and candidate entries with stable scores and highly overlapping main error causes are merged.

[0065] In this embodiment, the minimum elimination unit needs to be rewritten back to the object base layer after splitting or merging. During the write-back, the minimum elimination unit identifier, knowledge anchor identifier, and version field are used as key fields, and the standard expression field, standard process field, verification anchor point field, main error cause field, current verification gap field, source batch field, and update time field are used as value fields. If two or more candidate entries with similar alignment scores and a higher number of high-conflict edges in the conflict matrix appear under the same knowledge anchor point, the system prioritizes retaining the candidate entries with higher constraint levels and higher frequency of support from past exam questions, and writes the remaining candidate entries into the easily confused edge field, without directly promoting them to new minimum elimination units. This processing is used to stabilize the object base layer and reduce misclassification caused by coarse labeling.

[0066] After completing the adjustment of the minimum elimination unit, the system aggregates the step gap, restatement gap, interval gap, migration gap, and independent response gap generated in Example 1 to obtain the gap aggregation amount. This is to serve the subsequent generation of the set of allowed actions. The aforementioned gap quantities are all derived from the disproving gap field in Example 1, and are aggregated using normalized weight coefficients. to The weighted summation results in a total weight of one. The initial values ​​of the aggregation weight coefficients can be set as follows: 0.25 for each step gap and independent response gap, 0.15 for the restatement gap, and 0.175 for each interval gap and migration gap. These values ​​are then corrected during the sample update phase. The system also calculates the recovery delay aggregation for the same object, that is, it weights and sums the recovery delay after prompting, the number of interval recovery rounds, and the number of migration recovery rounds after normalizing them within the same window, using this summation as the timing input for calculating the inconsistency of subsequent actions.

[0067] To achieve stricter additional release control, this embodiment introduces a set of permitted actions after the main gating chain in Embodiment 1. The system first calculates the action inconsistency degree for each candidate action, and then determines whether the action is acceptable based on the threshold range of the calibration sample group. The action inconsistency degree is calculated according to formula (6): in, Represent the learning object Relative to candidate actions Inconsistency in actions This indicates a stable grasp of the strength of support. Indicates the surface's support strength. Indicates the strength of the conflict of evidence. This represents the aforementioned amount of gap aggregation. This indicates the amount of aggregation with time delay after recovery. Indicates candidate actions The missing amount not yet covered by the required critical verification chain. Indicates candidate actions The degree of incompatibility with the current evidence structure, to This represents the normalized weight coefficients, all of which sum to one. Candidate actions must include at least four categories: maintaining the current governance state, continuing challenge verification, allowing elimination, and mandatory revert. This is derived from the comparison between the challenge task mask and the currently completed verification chain. It is derived from the matching results of the object's current stage, current pool type, and most recent failure reason with the target action template.

[0068] In this embodiment, the set of allowed actions is determined by the action inconsistency degree and the threshold of the calibration sample group. The system establishes calibration sample groups according to subject, question type, minimum elimination unit category, and task family. Each calibration sample group includes at least historical stable elimination samples, historical samples to be supplemented for stabilization, and historical backflow samples. During offline updates, the system statistically analyzes the action inconsistency degree distribution corresponding to different candidate actions and calculates the action acceptance threshold under the target false elimination rate constraint. The set of allowed actions is calculated according to formula (7): in, Represent the learning object The set of allowed actions, Indicates candidate actions, This represents the degree of inconsistency in actions obtained from formula (6). Represent the learning object Belonging to the calibration sample group In candidate actions The system accepts actions based on the threshold. If the set of allowed actions includes both "continue challenge verification" and "allow elimination," the system prioritizes "continue challenge verification." If the set of allowed actions includes "forced reflow," the system directly prohibits advancement and writes the reflow action to the state write-back object. These are parameters to be calibrated, derived from the statistical distribution of historical stable samples and historical backflow samples. After updating, they are written back to the group field, action field, version field, and effective time field in the threshold configuration object.

[0069] After obtaining the set of permitted actions, this embodiment further generates a clearance certificate to distinguish between "not blocked" and "allowed". The clearance certificate criterion is calculated according to formula (8): in, Represent the learning object The certificate of elimination has a value of one indicating that the additional release conditions are met, and a value of zero indicating that they are not met. This indicates that the action can be eliminated. This indicates an action to maintain the current governance status. This indicates a forced reflux action. This represents the set of allowed actions obtained from formula (7). Indicates the coverage of the critical verification chain. This indicates the critical coverage threshold. Critical verification chain coverage must comprehensively consider whether the critical chains in the independent response chain, interval maintenance chain, transition chain, and restatement chain have been completed according to the challenge task mask requirements.

[0070] Working process: After the system completes the basic main chain calculation and obtains the three-state support result, evidence conflict strength, counter-evidence constraint result and dual-pool split result in Example 1, it starts the enhanced control process of this example. First, the minimum elimination unit construction unit re-aligns the cross-source knowledge candidate items. During alignment, it reads the source support vector, source conflict matrix and individual exposure frequency vector, calculates the cross-source alignment score obtained by formula (5), and completes the retention, splitting and merging of candidate items accordingly. Then, the system reads the five types of gap quantities written back in Example 1 and generates gap aggregation quantity and recovery delay aggregation quantity. Next, the system calculates the action inconsistency degree obtained by formula (6) for the four types of candidate actions: maintain, continue challenge verification, allow elimination and force backflow, and then generates the allowed action set according to formula (7). If the allowed action set contains allowed elimination actions, the system continues to calculate formula (8) to generate elimination certificate and writes the elimination certificate into the elimination permission result field; if the allowed action set contains forced backflow actions, the system directly increases the backflow priority and rewrites the current stage field in the state write-back object.

[0071] In this embodiment, the gating threshold adopts a sample-based update mechanism. The system maintains at least three sample sets: a historical stable sample set, a historical sample set to be stabilized, and a historical backflow sample set. These are used to record objects to be stabilized, objects to be stabilized for a long time, and objects that subsequently become unstable again, respectively. When updating the threshold, the system uses the sample set and calibration sample group identifier as key fields, and writes the action acceptance threshold, elimination prohibition threshold, backflow threshold, critical coverage threshold, update time, sample quantity, and version fields as value fields into the threshold configuration object. If the sample quantity of a certain group is lower than the preset minimum quantity, the enhancement layer in this embodiment is not enabled for that group, and it reverts to the main gating chain of Embodiment 1.

[0072] In this embodiment, the abnormal events are mainly identified around the failure to enable or group the reinforcement layer, including at least the following: insufficient sample anomaly, group drift anomaly, cross-source conflict anomaly, and reinforcement layer result conflict anomaly. Insufficient sample anomaly corresponds to the number of samples in a calibration sample group being lower than the minimum sample threshold; group drift anomaly corresponds to the current question type or task family of the learning object being inconsistent with its most recent calibration group and the consistency being lower than 0.5; cross-source conflict anomaly corresponds to the number of high-conflict edges in the source conflict matrix exceeding a preset upper limit; reinforcement layer result conflict anomaly corresponds to the allowed action set simultaneously containing allowed elimination actions and forced reflow actions. The above anomaly thresholds are all derived from the threshold configuration object, with priority from high to low as reinforcement layer result conflict anomaly, cross-source conflict anomaly, insufficient sample anomaly, and group drift anomaly. Recovery conditions are successful regrouping, successful sample replenishment, or successful recalculation after reverting to the main gating chain of Embodiment 1; if the recovery conditions are not met, the entire reinforcement layer is disabled, and elimination certificates are not allowed to be generated.

[0073] A specific operating scenario in this embodiment is as follows. A certain foreign language easily confused phrase object is answered correctly twice in the result chain, but fails the blank recall task, fails the context transfer sentence construction task, and only passes the independent answering task under the light prompt condition. In embodiment 1, the system has identified this object as having passed the result chain but with a risk of superficial mastery, and writes it into the problem pool. After entering this embodiment, the minimum elimination unit construction unit reads the candidate knowledge entries corresponding to the object and finds that the verification anchor support of the phrase is high in the real questions and question bank, while the textbook catalog and teacher materials have conflicts with the explanation of its easily confused edge. Therefore, the cross-source alignment score is calculated according to formula (5) and the object is split into two object basic layer units: "word meaning discrimination minimum elimination unit" and "context transfer minimum elimination unit". Subsequently, the system calculates the gap aggregation amount and finds that the paraphrasing gap and the transfer gap are dominant. Then, the action inconsistency degree is calculated for the allowed elimination action and the continued challenge verification action according to formula (6). Since the inconsistency of the continued challenge verification action is lower than the corresponding threshold, while the inconsistency of the allowed elimination action is higher than the corresponding threshold, the set of allowed actions obtained by formula (7) only includes the continued challenge verification action and does not include the allowed elimination action. Based on this, the system does not generate an elimination certificate, but continues to output the migration verification task and the silent independent answering task.

[0074] Example 1 has already achieved basic blocking of false elimination objects. In this example, the source support, individual exposure, and conflict cost of cross-source knowledge candidate entries are uniformly incorporated into the alignment score calculation through formula (5), making the construction of the smallest elimination unit closer to the governance object; the set of allowed actions is generated through formulas (6) and (7); and the coverage of the key verification chain is incorporated into the additional release condition through formula (8). Thus, the system adds object basic layer correction and additional release control in addition to the main gating chain.

[0075] The parameter gap list in this embodiment includes cross-source alignment score weighting coefficient, gap aggregation weighting coefficient, action inconsistency weighting coefficient, action acceptance threshold for each calibration sample group, critical coverage threshold, and minimum sample threshold required to enable the enhancement layer. For the above parameters to be calibrated, the system should simultaneously disclose the source of initial values, the source of calibration samples, calibration triggering conditions, and update frequency. Among them, the initial value of the cross-source alignment score weighting coefficient can be derived from the statistics of historical entry retention effects, the initial values ​​of the action acceptance threshold and critical coverage threshold can be derived from the quantile comparison of historical stable samples and historical backflow samples, and the initial value of the minimum sample threshold can be derived from the minimum statistical reliability requirements during engineering deployment. If there are not enough historical samples at the beginning of deployment, the enhancement layer in this embodiment is turned off by default, and only the alignment rules of the object base layer are retained; after the number of samples reaches the threshold, the allowed action set and certificate removal functions are enabled in batches.

[0076] Example 3 like Figures 1 to 3As shown, this embodiment illustrates the implementation of boundary and anomaly recovery based on Embodiments 1 and 2. The system continues to use the evidence event object, three-state support result, evidence conflict strength, disproving gap field, and state write-back object from Embodiment 1, as well as the minimum elimination unit update result, allowed action set, and certificate elimination result from Embodiment 2. It focuses on disclosing anomaly feedback identification, cross-device inconsistency identification, forced reflow judgment, downgraded spot check judgment, and recovery priority control after reflow. This implementation is applicable to learning support systems with paper-based feedback, weak prompt independent response feedback, cross-device synchronization, and low-confidence identification results.

[0077] In this embodiment, the system adds an anomaly identification and recovery control unit, which is interface-coupled with the status write-back unit in Embodiment 1 and the reinforcement layer control unit in Embodiment 2. The anomaly identification and recovery control unit receives independent response feedback records, paper feedback records, cross-platform synchronous records, manual verification records, and the current status of the object, and outputs anomaly classification results, reflow results, degradation action results, and recovery priority results. This unit maintains at least the following fields: feedback integrity field, feedback credibility field, cross-platform consistency field, anomaly type field, anomaly priority field, recovery action field, and recovery completion identifier field, and uniformly writes them into the anomaly status object. The learning object identifier, current stage field, and anomaly batch field serve as key fields, while the feedback integrity, feedback credibility, anomaly type, recovery action, recovery priority, update time, and version fields serve as value fields.

[0078] In this embodiment, anomalous evidence first undergoes a unified anomalous quantification phase. To this end, the system constructs a cross-end consistency matrix and an anomaly priority vector. The cross-end consistency matrix adopts a 3x3 structure, with the three dimensions corresponding to terminal response feedback, paper-based recognition feedback, and manual verification feedback, respectively. When any two types of feedback differ in answer results, time series, or key step fields, the corresponding off-diagonal elements are updated according to the intensity of the difference. The anomaly priority vector adopts a four-dimensional structure, with the four dimensions corresponding to cross-end inconsistency anomalies, incomplete feedback anomalies, low-confidence feedback anomalies, and missing evidence anomalies, respectively. Each new feedback arrives and is written to the corresponding dimension according to the anomaly triggering result. If the same object repeatedly triggers the same type of anomaly within the current judgment window, the value is incremented by a fixed step size based on the original value. Both the matrix and vector are created during object initialization and updated after each round of feedback parsing.

[0079] In this embodiment, the return integrity is calculated according to formula (9): in, Represent the learning object The integrity of the return transmission and These represent the number of questions that have been successfully submitted and the total number of questions that should be submitted, respectively. and These represent the number of valid step segments and the number of critical step segments that should be covered, respectively. and These represent the number of valid timestamp entries and the expected total number of timestamp entries, respectively. to This represents the normalized weight coefficients, all summing to one. The number of questions comes from the task feedback object, the number of step segments comes from the structured step recognition results, and the number of timestamp entries comes from the feedback logs. The system... Objects with a return value below 0.7 are marked as incomplete return anomalies. This threshold is a parameter to be calibrated and is derived from the distribution comparison between historical complete return samples and abnormal return samples.

[0080] After obtaining the completeness of the returned data, the system further calculates the reliability of the returned data. The reliability of the returned data is calculated according to formula (10): in, Represent the learning object The reliability of the return transmission This represents the return integrity obtained by formula (9). The cross-end consistency score is derived from the normalized aggregation result of the cross-end consistency matrix. The time reasonableness score is derived from the matching results of the response time and feedback sequence with the baseline window. This indicates the amount of abnormal transition markers, which originate from the mutation detection results in the results, steps, or durations of two adjacent rounds of transmission. to This represents the normalized weight coefficients, and their sum is one. If If the value is below 0.5, the system marks the object as a low-confidence postback anomaly and prohibits it from directly entering the permission removal judgment. , and All values ​​are between zero and one, and The larger the value, the more severe the abnormal jump.

[0081] In this embodiment, in order to quantify the degree of subsequent instability of the eliminated object, the system jointly reads the evidence conflict strength from Embodiment 1. The difference between the most recent interval maintenance result and the stable baseline result. The difference between the most recent migration validation results and the stable baseline results. The indicator that failed to answer independently again after being removed has been removed. and cross-platform inconsistency exception flags The variables mentioned above are derived from the state object written back in Example 1, the independent response feedback object, and the anomaly identification result in this example, and are used as direct inputs for the forced reflow determination.

[0082] The forced reflux criterion in this embodiment is calculated according to formula (11): in, Represent the learning object The forced reflow result has a value of 1, which means that the state has been forcibly withdrawn from the eliminated state and written back to the state to be stabilized or being resolved. A value of 0 means that forced reflow is not triggered. , and This indicates the corresponding threshold. The criterion is as follows: if an eliminated object experiences at least one of the following: high conflict, interval instability, migration instability, failure to respond independently again, or cross-end inconsistency, a reflow will be triggered.

[0083] When the object is not sufficient to trigger a forced return, but the return is incomplete, has low reliability, or has a minor conflict, the system enters the degradation check chain. The degradation check criterion is calculated according to formula (12): in, Represent the learning object Whether to enter the downgraded spot check or re-verification chain; a value of one indicates that a downgraded action has been triggered. Indicates the reliability of the returned data; Indicates the integrity of the returned data; The amount of mild conflict indicators comes from the weak conflict detection results between the result chain and the paraphrase chain, and between the result chain and the process chain; This indicates the critical coverage gap, which originates from the coverage difference between the challenge task mask and the completed verification chain in Example 2. to This represents the normalized weight coefficients, and their sum is one. This indicates the threshold for downgraded random checks. If as one and If the value is zero, the system will prioritize assigning low-risk spot check tasks, resampling tasks, or re-verification tasks, instead of directly canceling the eliminated status.

[0084] Working process: After receiving a new paper-based feedback, terminal feedback, or manual verification feedback, the system first writes the feedback into the abnormal state object according to the learning object identifier and abnormal batch field, then calculates the feedback integrity using formula (9), and calculates the feedback credibility using formula (10) based on the cross-end consistency matrix and abnormal jump detection results. If the object is currently in the eliminated state, the system continues to read the evidence conflict strength, interval decline, migration decline, independent answer failure flag, and cross-end inconsistency flag written back in Example 1, and determines whether to trigger forced feedback according to formula (11). If the object does not meet the forced feedback conditions, but the feedback integrity or feedback credibility is low, or there is a significant lack of key verification coverage, then determines whether to trigger downgraded spot check according to formula (12). If the forced return result is one, the system writes the current stage field back to "to be stabilized" or "being resolved", and sets the return identifier field to one; if the downgraded spot check result is one, the system does not change the current stage field, but writes the recovery action field to "low-risk spot check, re-independent response verification, or restructured restatement verification"; if neither is triggered, the object continues to execute subsequent release control along the main gating chain of Embodiment 1 and Embodiment 2.

[0085] In this embodiment, the recovery priority is determined by the anomaly priority vector and the current stage. If the cross-end inconsistency anomaly dimension is the highest value, the system prioritizes manual verification or resampling; if the incomplete feedback anomaly dimension is the highest value, the system prioritizes supplementary feedback and low-risk spot checks; if the low credibility anomaly dimension is the highest value, the system prioritizes re-independent response verification; if the missing evidence anomaly dimension is the highest value, the system prioritizes supplementary sampling and delay elimination. When a recovery action is written to the recovery task object, the learning object identifier, anomaly batch field, and recovery task type field are used as key fields, and the priority field, task deadline field, retry count field, update time field, and version field are used as value fields. The retry count is set to zero when the recovery task object is initialized. If the same recovery action fails consecutively, the retry count is incremented by one for each failure. When the retry count reaches the preset upper limit, the system automatically writes the object back to the pending stabilization state and pauses the release.

[0086] In this embodiment, anomaly quantification employs a unified threshold source management approach. The return integrity threshold, return credibility threshold, return flow threshold, return flow credibility lower limit, degradation sampling threshold, and retry upper limit all originate from the anomaly configuration object. The anomaly configuration object includes at least the threshold name field, applicable scenario field, current value field, source sample field, update time field, effective time field, and version field. If a threshold is a parameter to be calibrated, its initial value, calibration sample source, calibration trigger condition, and update frequency are recorded simultaneously. Specifically, the initial value of the return flow threshold can be derived from a comparison of the unstable aggregation distribution of historical return flow samples and historical stable samples; the initial value of the return flow credibility lower limit can be derived from the lower quantile of high-credibility return flow samples; and the initial value of the degradation sampling threshold can be derived from the distribution statistics of mildly anomalous samples. If sufficient samples are lacking in the initial deployment phase, the system adopts a conservative initial value and performs the first offline update after accumulating a preset sample size.

[0087] For various abnormal events in this embodiment, the abnormal criteria, threshold sources, priorities and recovery conditions are all written into the abnormal configuration object and abnormal status object in a unified format: the abnormal criteria correspond to at least one of the quantitative results in formulas (9) to (12), the threshold sources correspond to historical stable samples, historical abnormal samples or engineering operation and maintenance constraints, the priorities are determined in the order of cross-end inconsistent anomalies, incomplete feedback anomalies, low-confidence feedback anomalies and evidence missing anomalies, and the recovery conditions correspond to at least one of the following: successful resampling, completed refeedback, passed manual verification or successful recalculation after falling back to the main gate chain.

[0088] A specific operating scenario in this embodiment is as follows. A learning object has already met the release conditions of Embodiment 1 and Embodiment 2 in the previous cycle and has been updated to the eliminated state. Subsequently, the learner completes an independent answering task without prompting in a paper environment. The paper recognition result shows that the answer failed, but the terminal cache record shows that the previous weak prompting answer passed. At the same time, the manual verification record points out that there are key omissions in the paper steps. The system first establishes a cross-end consistency matrix for the paper feedback, terminal feedback and manual verification feedback of the object, and obtains medium feedback completeness and high feedback credibility according to formula (9) and formula (10). Then, the system reads the evidence conflict strength, interval decline, independent answering failure mark and cross-end inconsistency mark of the object, and then determines the forced feedback result as one according to formula (11). Based on this, the system cancels the eliminated state of the object, writes its current stage field back to be stabilized, and regenerates the independent answering verification task without prompting and the step completion task. If subsequent manual verification feedback indicates that there was an error in the previous paper recognition, and the recalculation no longer meets the return condition of formula (11), then the system will restore the object state according to the recovery action chain.

[0089] Examples 1 and 2 have solved the problem of identification and release control of "result chain passed but not stable control". However, when the system enters the real operating environment, another type of risk will be encountered, namely, misjudgment of the state due to low return quality, inconsistent cross-end records or subsequent instability of the eliminated object. Without the abnormal quantification and recovery chain in this example, the system may erroneously cancel the eliminated state under low quality return, or continue to maintain release when obviously unstable. This example explicitly quantifies the return quality through formulas (9) and (10), and then identifies the object that "must be forced to return" by formula (11) and the object that "must not return temporarily but must be downgraded for spot check" by formula (12). Thus, abnormal return, return reentry and recovery actions are incorporated into the same computer control chain to ensure that the cancellation, retention and recovery of the eliminated state have consistent quantitative basis.

[0090] The parameter gap list in this embodiment includes the return integrity weight coefficient, the return reliability weight coefficient, the instability aggregation weight coefficient, the return threshold, the return reliability lower limit, the degradation sampling threshold, and the recovery retry upper limit. For the above parameters to be calibrated, the system should disclose the initial value or range, the sample source, the calibration trigger condition, and the update frequency. Among them, the initial value of the return integrity weight coefficient can be derived from the historical return field missing distribution, the initial value of the instability aggregation weight coefficient can be derived from the feature importance analysis of the return samples, the initial values ​​of the return threshold and the degradation sampling threshold can be derived from the quantile comparison of historical stable samples and historical abnormal samples, and the initial value of the recovery retry upper limit can be derived from engineering operation and maintenance constraints. If the current deployment scenario has not accumulated enough cross-end return samples, this embodiment will prioritize the use of the return integrity and return reliability control chains, and will not perform fine-grained group updates on the instability aggregation weight for the time being. Updates will be performed in batches after the samples have been accumulated.

[0091] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be modified within the scope of the concept described herein by means of the above teachings or the technology or knowledge in related fields.

Claims

1. A gated system for error resolution and elimination based on multi-source evidence, characterized in that, The system includes an evidence collection and preprocessing unit, a minimum elimination unit construction unit, a latent lack of knowledge identification unit, a counter-evidence constraint generation unit, an asymmetric gating unit, a dual-pool governance unit, and a state write-back unit. The evidence collection and preprocessing unit collects learning result data, process trajectory data, prompting behavior data, rehearsal content data, interval retention data, transfer verification data, independent response data, and written feedback data, converting them into evidence events oriented towards the learning object. It also performs source normalization, time alignment, task identifier alignment, and knowledge anchor mapping on the evidence events. The minimum elimination unit construction unit... The unit is used to construct the minimum elimination unit corresponding to the learning object based on the examination syllabus, curriculum standards, textbook catalog, historical exam questions, question bank data, collection of incorrect and difficult questions, teacher-uploaded materials, and learners' historical learning records. The implicit lack of mastery identification unit is used to generate a three-state support result containing lack of mastery support strength, apparent mastery support strength, stable mastery support strength, and evidence conflict strength based on the result chain, process chain, paraphrase chain, interval maintenance chain, transfer chain, and independent response chain. The contradictory evidence constraint generation unit is used to extract the contradictory evidence gap fingerprint, recover the time-delay signature, and generate a challenge task mask. The non- The symmetric gating unit is used to generate prohibition elimination results, reflow results, and elimination permission results based on the three-state support results, evidence conflict strength, disproving gap fingerprint, recovery delay signature, and challenge task mask. The dual-pool governance unit is used to write the learning object into the error pool when the result chain fails, and to write the learning object into the challenge pool when the result chain passes but there are risks of superficial mastery, evidence conflict, disproving gap, recovery delay, duration deviation, independent answer failure, and structured paraphrase failure. The learning object is then organized according to the knowledge anchor point, verification anchor point, and main error cause corresponding to the minimum elimination unit. The structured paraphrase task, interval maintenance task, transfer verification task, and independent response verification task corresponding to the challenge task mask are described. The state write-back unit is used to update the eliminated state when the prohibition elimination result indicates that progress is allowed, the reflow result indicates that reflow is not required, and the elimination permission result indicates that elimination is allowed. When at least one of the following occurs again in subsequent evidence, the learning object is written back to the state to be stabilized and re-enters the state of being resolved: increased conflict intensity; interval instability; transfer instability; independent response failure; cross-end inconsistency reaches an abnormal threshold. At the same time, the main error cause, the most recent failure reason, and the reflow identifier are written back.

2. The error-prone problem governance and elimination gating system based on multi-source evidence according to claim 1, characterized in that, The evidence events include source identifier, time identifier, task type identifier, knowledge anchor identifier, prompt level, process completeness, duration deviation, expression independence, and result destination identifier.

3. The error-prone problem governance and elimination gating system based on multi-source evidence according to claim 1, characterized in that, The implicit lack-of-knowledge identification unit maps the result chain to the result passed and to be distinguished support, maps the stable pass of independent answering, the pass of resisting probing, the stable pass of maintaining interval, and the stable pass of transfer verification to the stable mastery support, maps the increased reliance on prompts, insufficient process completeness, failure of structured paraphrasing, and the deviation of duration reaching the threshold to the superficial mastery risk support, maps the failure of independent answering, failure of resisting probing, interval instability, and transfer instability to the lack-of-knowledge risk support, and retains the quantity formed by mutually exclusive support as the evidence conflict strength.

4. The error-prone problem governance and elimination gating system based on multi-source evidence according to claim 1, characterized in that, The counter-evidence gap fingerprint includes step gap amount, repetition gap amount, interval gap amount, migration gap amount, and independent response gap amount. The recovery time delay signature includes the delay amount from passing after prompt to passing independently, the number of rounds required for recovery after interval, and the number of rounds required for migration recovery.

5. The error-prone problem governance and elimination gating system based on multi-source evidence according to claim 1, characterized in that, The counter-evidence constraint generation unit generates a challenge task mask based on the three-state support results, evidence conflict strength, counter-evidence gap fingerprint, recovery delay signature, and risk trigger code. This mask is used to instruct the current learning object to preferentially trigger at least one of the following: independent answer verification, anti-probing restatement verification, interval maintenance verification, transfer variant verification, and step completion verification.

6. The error-prone problem governance and elimination gating system based on multi-source evidence according to claim 1, characterized in that, The asymmetric gating unit outputs the prohibition elimination result when at least one of the following conditions is met: the surface grasp support strength reaches the prohibition elimination threshold; the evidence conflict strength reaches the prohibition elimination threshold; or the counter-evidence constraint strength reaches the prohibition elimination threshold. The risk of duration deviation reaches the threshold; The result chain passes but independent response fails; the hint dependency reaches the threshold and the restatement vulnerability reaches the threshold; after the learning object is in the eliminated state, the asymmetric gating unit outputs the backflow result if at least one of the following conditions is met: the evidence conflict strength reaches the backflow threshold; the interval descent reaches the interval instability threshold; the migration descent reaches the migration instability threshold. Independent response failed again; there is a cross-platform inconsistency between paper-based response data, terminal response data, or manually verified data, and the inconsistency has reached the abnormal threshold.

7. The error-prone problem governance and elimination gating system based on multi-source evidence according to claim 1, characterized in that, The dual-pool governance unit generates a paraphrase gating result when the structured paraphrase verification passes, and allows the removal of the prohibition elimination restriction triggered by the paraphrase chain only when the paraphrase gating result shows that the paraphrase has passed and the anti-probing test has passed, the key verification chain coverage corresponding to the challenge task mask meets the conditions, and at least one of the independent response chain, the interval maintenance chain, and the migration chain passes again.

8. The error-prone problem governance and elimination gating system based on multi-source evidence according to claim 1, characterized in that, The minimum elimination unit construction unit performs retention, splitting, and merging processing on cross-source knowledge candidate entries based on the source support vector, source conflict matrix, and individual exposure frequency vector to generate a minimum elimination unit containing knowledge anchor identifier, task type identifier, source chain, standard expression key points, standard process key points, verification anchor, current verification gap, and current elimination status.

9. The error problem governance and elimination gating system based on multi-source evidence according to claim 1, characterized in that, The asymmetric gating unit also generates a set of allowed actions based on the action inconsistency and the threshold corresponding to the calibration sample group. When the set of allowed actions includes elimination actions, does not include hold actions or backflow actions, and the critical verification chain coverage corresponding to the challenge task mask reaches the target threshold, an elimination certificate is generated, and the elimination certificate is used as an additional restriction condition for the elimination permission result.

10. The error problem governance and elimination gating system based on multi-source evidence according to claim 1, characterized in that, The asymmetric gating unit determines the gating thresholds corresponding to the surface control support strength, evidence conflict strength, duration deviation risk, interval descent, and migration descent based on the historical stable sample set, historical backflow sample set, false elimination loss, overvalidation loss, target false elimination rate, and target false retention rate, and performs threshold updates after new backflow samples are added.

11. The error problem governance and elimination gating system based on multi-source evidence according to claim 1, characterized in that, The evidence collection and preprocessing unit is also used to turn the paper-based response data into events and use it as evidence input for result chains, process chains, independent response chains, or time deviation risk judgments, in order to update the status of the problem pool, eliminate permission results, or return results.

Images