Enterprise online physical examination diagnosis and expert consultation integrated implementation method and system and computer program product

Abstract

The invention relates to an enterprise digital management system, in particular to an enterprise online physical examination diagnosis and expert consultation integrated implementation method and system and a computer program product. According to the method, after a user completes mobile phone number verification and establishes an enterprise subject, a physical examination classification is selected to initialize an evaluation state Z; in the questionnaire process, credibility c is calculated based on logic consistency, data consistency and behavior consistency, c is coupled to adaptive question selection and early stop criteria, and index coverage and question quantity control are achieved; then, a rule score Pr is calculated through a rule engine, a feature vector X is constructed, a calibration score Pm is output through a calibration model, and a diagnosis level and risk vector R is generated; a diagnosis workpiece A and an evidence chain E are further generated, a workpiece hash hA is calculated, an auditing hash chain abstract hL is constructed, and staticizing and accurate failure are achieved through a cache key K; and matching experts and returning a report based on R in a consultation scene, and returning samples (X, Y) meeting gating conditions for model iteration. According to the method, the diagnosis accuracy, interpretability, traceability and high concurrency performance are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to an enterprise digital management system, and in particular to an integrated implementation method, system and computer program product for online enterprise physical examination diagnosis and expert consultation. BACKGROUND

[0002] With the continuous improvement of enterprise digital management and compliance supervision requirements, online physical examination / diagnosis platforms for small and medium-sized enterprises have gradually become an important form of enterprise management services. Such platforms usually use questionnaires as an entry point, and conduct structured quantitative assessment of enterprise status around dimensions such as finance and tax, finance, human resources, operations, governance, and strategy, and further output diagnosis reports, improvement suggestions, and some platforms also provide expert consultation / consultation and other value-added services. Enterprise physical examination systems have significant value in reducing the threshold of professional services, shortening the diagnosis cycle, and achieving large-scale delivery, but still face multiple technical challenges in engineering landing and result credibility, traceability, reproducibility, and sustainable iteration.

[0003] In the prior art, there are relatively typical "enterprise evaluation-scoring-report output" solutions. For example, Chinese patent CN113807747A provides an enterprise budget management maturity evaluation system, which generally involves constructing an evaluation index system and a questionnaire bank, collecting and assigning scores to enterprise-related data, forming sub-index results and giving evaluation conclusions, diagnostic comments or improvement suggestions, and then outputting evaluation reports and other results. Such a solution can achieve structured evaluation and report generation of enterprise management capabilities, and is suitable for solidifying expert knowledge into index weights and question mapping rules, thereby enabling automatic scoring and grade determination. However, from a system engineering perspective, the above solution often emphasizes the "questionnaire scoring and report output" main link, and for the problems of data authenticity, answer consistency, and answer behavior abnormalities that exist in online scenarios, it usually only adopts basic verification or weak constraint strategies, resulting in large deviations in scoring results when enterprise data is incomplete, contradictory, or has strategic answers. In addition, the explainability and reproducibility of the scoring results (e.g., rule version and model version solidification), as well as the subsequent consultation closed loop and iterative update mechanism, do not necessarily form a strongly coupled integrated solution.

[0004] For the direction of "questionnaire authenticity and compliance risk identification", the existing technology also proposes the idea of "questionnaire authenticity verification-risk rating-dimension score". For example, Chinese patent CN116401343A provides a data compliance analysis method, which processes the answers of users through a compliance questionnaire library, and performs form authenticity verification, violation item risk rating and dimension score on the selected questionnaire, to obtain the scoring results. This kind of scheme emphasizes improving the credibility of questionnaire scoring through authenticity verification and risk rating, which can to some extent inhibit the influence of invalid filling or low-quality filling on the scoring conclusion. But from the whole process of enterprise "physical examination-diagnosis-consultation-landing", authenticity verification often still stays at the level of "post-checking / tagging": even if abnormalities are found, they may only be used as a prompt or review mark, and "credibility" is not further used as a key control variable to drive questionnaire routing strategies (such as preferentially distributing mutual verification questions, pursuing problems, and clarifying questions) or termination conditions (such as increasing termination thresholds and extending questionnaires to supplement evidence when credibility is low), so as to achieve a dynamic balance between statistical efficiency and diagnostic accuracy. Therefore, when the enterprise physical examination scene needs to meet both "controllable question volume (experience)" and "credible results (accuracy)", authenticity verification and risk rating alone are still difficult to form a robust closed loop for complex online noisy data.

[0005] At the same time, the common value-added form of enterprise physical examination platform is the "expert consultation / advice" closed loop, that is, after the user obtains the preliminary diagnosis result, the consultation is initiated and the data is submitted, and the platform matches the experts and outputs further professional reports. In the existing technology, Chinese patent CN104866711A discloses a method for expert platform and consultation request: after the user end initiates the consultation request, the platform returns the expert list or matching result to the user, and determines the consultation arrangement in combination with the expert scheduling information. This kind of scheme solves the service organization problem of "consultation request-expert matching-scheduling confirmation", and can improve the efficiency and availability of consultation service. But for the specific scene of enterprise physical examination, if the expert matching is only based on label or simple condition screening, it is easy to appear "the expert is not suitable for the enterprise problem"; more importantly, most consultation platforms regard "consultation result" as an independent service product, rather than a feedback signal for continuous improvement of diagnosis algorithm, that is, the consultation report does not feed back to the questionnaire scoring rules or calibrate the model, and does not form an engineering closed loop of "diagnosis artifacts-evidence chain-audit chain-model iteration", which makes it difficult to improve the accuracy and generalization ability of automatic diagnosis through real consultation case accumulation after long-term operation of the system.

[0006] Further, from the perspective of system credibility and operation performance, the enterprise physical examination platform usually needs to meet the requirements of multi-role access (user end, administrator end, expert end), multi-task concurrency (questionnaire submission, report generation, data upload and download, SMS notification, etc.), and result traceability (log audit, responsibility definition). In the prior art, web page staticization and caching, log audit, tamper-proof evidence, etc. have been practiced and disclosed, but in the enterprise physical examination diagnosis scene, there is still a common engineering problem of "modular stacking": for example, after the generation of the diagnosis report, it is cached to improve access efficiency, but when the scoring rules / model version changes, the evidence chain is revised or the consultation data is supplemented, the precision invalidation of the cache and the imperfect consistency check easily cause the user end to display "old version diagnosis conclusion" or "inconsistent explanation caused by cache reuse"; again, although the log audit records operation information, there is a lack of strong binding (such as hash chain association) between the log and the diagnosis result, which makes it difficult to prove "the evidence and calculation version corresponding to a certain report" when there is a dispute or audit, thereby affecting the traceability and credibility of the platform.

[0007] In summary, the prior art provides valuable technical foundations in terms of "questionnaire evaluation and report output", "authenticity verification and risk scoring", "expert consultation matching and scheduling", etc., but still has the following shortcomings: first, there is a lack of "answer credibility" as a core control quantity to decouple the driving of adaptive questionnaire routing and termination confidence, thereby balancing the robust integration mechanism of user experience and diagnosis accuracy; second, there is a lack of engineering structure to solidify the scoring rule version, model version, and evidence chain record into reproducible "diagnosis artifacts", making it difficult to recalculate, trace and stably evolve the diagnosis conclusion; third, there is a lack of strong binding of the tamper-proof audit chain and the cache key, staticization strategy, to achieve precise consistency control of "safe and credible" and "high concurrency access"; fourth, there is a lack of closed-loop feedback and gating update mechanism between consultation results and automatic diagnosis, which prevents the expert knowledge from being continuously deposited into rule or model iteration capability. Based on the above problems, there is an urgent need for an integrated, interpretable, traceable and sustainable iteration of enterprise online physical examination diagnosis and expert consultation technical solution to improve the reliability and engineering usability of online physical examination conclusions. SUMMARY

[0008] The present application aims to provide an integrated implementation method and system of enterprise online physical examination diagnosis and expert consultation, which realizes controllable question amount and more reliable diagnosis results by coupling answer reliability assessment and adaptive questionnaire routing / early termination mechanism; and adopts rule-model fusion scoring and reproducible diagnosis artifacts / evidence chain output, combined with artifact hashing and audit hashing chain tamper-proofing traceability and cache key precise invalidation strategy, to ensure that the diagnosis conclusion is interpretable, reproducible, traceable and high-concurrency stable access; meanwhile, data security verification, expert matching and consultation result backflow calibration are realized in the expert consultation closed loop, thereby continuously improving the diagnosis accuracy and platform engineering usability.

[0009] In order to achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0010] An integrated implementation method of enterprise online physical examination diagnosis and expert consultation, comprising the following steps:

[0011] S1, identity verification and account establishment based on mobile phone verification code, collection of enterprise industry and regional information;

[0012] S2, calling target physical examination classification question bank and initializing evaluation state Z, Z containing covered indicators, risk vector R, answer reliability c and scoring rule version v r and calibration model version v m ;

[0013] S3, adaptive question selection and early termination coupled with reliability: receiving answers and behavior characteristics, calculating c according to logical consistency, data consistency and behavior consistency; selecting the next question according to the comprehensive score of "information gain IG×g(c)+coverage contribution-abnormal items" for candidate questions, and reducing the confidence threshold with c, and terminating when the coverage and reduced confidence threshold are met or the upper limit of the question amount is reached;

[0014] S4, obtaining rule score P r according to rules, and inputting P r and feature vector X into calibration model to obtain calibration score P m , and determining diagnosis grade and updating R accordingly;

[0015] S5, generating diagnosis artifacts A and evidence chain E, and recording "question, indicator, weight, contribution, risk item, feature source" in evidence chain E, and writing v r , v m to support recalculation;

[0016] S6, calculating artifact hash h A , forming audit hash h L according to key operation log by hash chain, and generating cache key K containing enterprise identifier, classification identifier, (v r , v m ), hA , h L And accurate failure with A or audit changes;

[0017] S7, receive consultation data and perform format / size limit, generate file digest hash into audit chain, perform expert matching and scheduling distribution based on risk vector R, receive expert uploaded physical examination report and associate with diagnosis artifact A for display;

[0018] S8, extract conclusion label and recommendation vector Y from physical examination report, write (X, Y) into calibration sample pool to update v m And record change summary.

[0019] As preferred, S1 step receives user registration or login request, completes identity verification based on mobile phone number, establishes user account associated with enterprise subject; receive and store enterprise basic information, which at least includes enterprise name, industry and region;

[0020] And / or, S2 step provides physical examination portal to the user based on a preset set of physical examination categories, which at least includes financial and tax physical examination, financial physical examination, human resource physical examination, operation physical examination, governance physical examination and strategy physical examination;

[0021] As preferred, S3 step performs adaptive questionnaire routing coupled with credibility for each round of question selection:

[0022] S3-1, receive user's answer to the current question, collect answer behavior characteristics;

[0023] S3-2, based on the answer and the enterprise uploaded data and / or structured form data, calculate the answer credibility c, wherein c is determined by at least three types of consistency checks: logical consistency check, data consistency check and behavior consistency check;

[0024] S3-3, calculate the comprehensive question selection score S(q) for each candidate question q, and select the question with the maximum score as the next question, wherein:

[0025] S(q) = IG(q|Z)·g(c) + λ·Cover(q,Z) - μ·Abn(q,Z);

[0026] Wherein, IG(q|Z) is the information gain or entropy reduction of question q under evaluation state Z, g(c) is the discount function about credibility c, Cover(q,Z) is the coverage contribution of question q to uncovered indicators, Abn(q,Z) is the penalty or compensation term for abnormal / inter-check questions, and λ and μ are weight coefficients;

[0027] S3-4, reduce the termination confidence based on the credibility c, terminate the questionnaire if any of the following conditions are met: the pre-set minimum coverage threshold is reached and the reduced confidence is not lower than the threshold; or the maximum number of questions is reached; and record the termination state as Z * ;

[0028] Preferably, the logical consistency check at least includes a mutual check question pair table defining at least one set of mutually exclusive or mutually supporting constraints; when the constraints conflict, the credibility c is reduced and the next question routing is switched to a check question priority mode;

[0029] Preferably, the data consistency check at least includes comparing the interval values involved in the answers with the corresponding field interval values obtained by analyzing the uploaded data; when the deviation exceeds the threshold, Abn(q, Z) is set to positive compensation to preferentially distribute follow-up questions for clarifying the source of the deviation;

[0030] Preferably, the behavior consistency check at least generates an abnormal score based on the distribution of the answering time, the number of backtracking modifications, the device fingerprint, and the features of multiple accounts answering on the same device; when the abnormal score exceeds the threshold, the reduced confidence threshold in the termination condition is increased or at least one set of check questions is forcibly added.

[0031] As a preferred, in the S4 step, the answered results after termination are executed with the rule-model fusion scoring and risk vector updating:

[0032] S4-1, get a rule score P according to a configurable scoring rule r and generate a sub-index score;

[0033] S4-2, extract a feature vector X from the answers, enterprise data and behavior features, input X and P r into a calibration model to get a calibration score P m ;

[0034] S4-3, map P m to get a diagnosis level, and output a risk vector R, wherein the risk vector R at least includes risk weights and trigger reasons according to the index dimension;

[0035] Preferably, the rule score P r is obtained by executing a scoring rule description language DSL, and the DSL at least supports question weights, index mapping, segmented scoring and mutual exclusive deduction; the diagnosis artifact writes a DSL version hash to realize rule reproducibility.

[0036] As a preferred, the diagnosis artifact in the S5 step at least includes: a diagnosis level, a total score, a sub-index score, a risk vector R, and an evidence chain E; the evidence chain E at least records the association relationship of "question-index-weight-score contribution-risk item-feature source", and writes a scoring rule version number vr With calibration model version number v m To support recalculation consistency;

[0037] And / or, in step S6, the tamper-proof audit chain is coupled with the cache key for the diagnostic artifact:

[0038] S6-1, Calculate the hash of the diagnostic workpiece content summary h A ;

[0039] S6-2. Link and store the key operation logs of the diagnostic workpiece in a hash chain manner to form an audit chain hash h. L ;

[0040] S6-3. Generate a cache key K, where K consists of at least the enterprise identifier, the physical examination category identifier, and the scoring rule version number v. r With calibration model version number v m With hash h A It is configured to trigger precise cache invalidation based on K when the diagnostic artifact or chain of evidence changes;

[0041] Preferably, the cache key K further includes the audit chain hash h. L When any critical operation log link breaks or integrity verification fails, cache reuse is prohibited and diagnostic artifacts are forcibly regenerated.

[0042] Preferably, in step S7, the system receives the expert consultation request and consultation data uploaded by the user, applies format and size restrictions to the uploaded file, generates a file digest hash and writes it into the audit chain; performs expert matching and scheduling optimization based on the risk vector R, obtains a set of candidate experts and distributes the consultation data to the expert end; receives the physical examination report uploaded by the expert end, stores it in association with the diagnostic workpiece, and then returns it for display.

[0043] Preferably, the expert matching and scheduling optimization is a multi-objective ranking or multi-objective optimization, and the objective function includes at least the matching degree between the risk vector R and the expert's expertise label, the historical consultation quality score, the conflict cost of available time slots, and the industry similarity; and outputs at least one set of candidate experts that can be confirmed by the user.

[0044] And / or, in step S7, perform consultation return calibration: extract conclusion labels and suggestion vector Y from the expert examination report, write (X, Y) into the model calibration sample pool when the preset gating conditions are met, and trigger or plan to trigger the corresponding version of the calibration model update, while recording the return identifier and model version change summary in the diagnostic workpiece.

[0045] Furthermore, this application also provides an online enterprise health check-up and expert consultation system for implementing the method, including:

[0046] The module includes: User Management Module, Enterprise Information Module, Physical Examination Classification and Question Bank Module, Adaptive Routing Module, Credibility Verification Module, Rule-Model Fusion Scoring Module, Diagnostic Workpiece and Evidence Chain Generation Module, Anti-Tampering Audit Module, Cache Key Generation and Static Caching Module, Consultation Data Management Module, Expert Matching and Consultation Module, and Consultation Backflow Caching Module.

[0047] in:

[0048] The adaptive routing module is used to calculate the comprehensive topic selection score based on information gain and coupled with credibility c, and then select the next topic.

[0049] The credibility verification module is used to output the credibility level c, which is determined by logical consistency, data consistency and behavioral consistency.

[0050] The diagnostic artifact and evidence chain generation module is used to generate a diagnostic artifact containing the scoring rule version number v. r With calibration model version number v m The chain of evidence relating to the source of the features;

[0051] The anti-tampering audit module is used to generate hash chains for key operations on diagnostic workpieces and consultation data;

[0052] The cache key generation and static cache module is used to construct cache keys using enterprise identifier, physical examination category identifier, version number and workpiece hash and to achieve precise invalidation.

[0053] The consultation feedback calibration module is used to feed the tags and suggestions extracted from the expert report back to the calibration sample pool and trigger the model version update record.

[0054] Preferably, the anti-tampering audit module writes at least the operator identifier, object identifier, operation type, timestamp, client IP, current hash and previous hash to each audit record, and provides an integrity verification interface to verify the continuity of the hash chain;

[0055] And / or, the consultation feedback calibration module sets gating conditions, which include at least an expert conclusion consistency threshold, a sample deduplication threshold, and a confidence level c threshold; samples are written into the calibration sample pool and the calibration model version number v is updated only when the gating conditions are met. m .

[0056] Furthermore, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method.

[0057] Furthermore, this application also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the method.

[0058] This invention embeds the answer credibility 'c' into adaptive question selection and early termination criteria, enabling the system to automatically switch to a verification / follow-up question priority routing strategy and dynamically raise the termination threshold when incomplete, contradictory, or abnormal submissions are found. This significantly reduces the risk of misjudgment and omission while decreasing the number of invalid questions, thus improving the stability and robustness of diagnostic conclusions. Through rule-based P... r With calibration model output P m The integrated scoring mechanism and version management bind scoring rules, model parameters, and diagnostic outputs into reproducible "diagnostic artifacts." Within these artifacts, an evidence chain of "question—indicator—weight—contribution—risk item—feature source" is generated, giving the diagnostic results interpretable, recalculated, and auditable engineering attributes. This facilitates review and accountability in enterprise, expert, and regulatory scenarios. Furthermore, by hashing the content of the diagnostic artifacts... A 1. Construct a hash chain to audit the hashes of key operation logs. L Both the result and version number are incorporated into the cache key K to achieve coordinated control of "result consistency verification + precise cache invalidation," which avoids the reuse of old conclusions / errors and improves the efficiency of high-concurrency access and static display. In the expert consultation stage, the risk vector R is used for expert matching and scheduling optimization, and the format / size restrictions and summary hashing of consultation data are applied to ensure that data transmission and process are traceable, secure and reliable. At the same time, the conclusion tags and suggestion vectors Y extracted from the expert reports are fed back to the calibration sample pool when the gating conditions are met, realizing controllable updates and continuous improvement of model iteration. Ultimately, the closed-loop optimization of enterprise health check from "collection - diagnosis - consultation - review - iteration" is achieved, which improves the overall accuracy, reliability, performance and maintainability of diagnosis. Attached Figure Description

[0059] Figure 1 System overall structure diagram (relationship between user terminal / expert terminal / management terminal / server terminal and core modules).

[0060] Figure 2 : Overall flowchart of the method (S1-S8 main process).

[0061] Figure 3 : Credibility Coupling Adaptive Topic Selection and Early Termination Flowchart (c calculation, S(q) calculation, termination condition).

[0062] Figure 4 : Illustration of fusion score and risk vector output (P) r X, P m , level mapping, R).

[0063] Figure 5 Diagnostic workpiece and evidence chain structure diagram (field relationship between A and E, version writing). Detailed Implementation

[0064] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of the present invention.

[0065] I. Terminology Explanation

[0066] To avoid ambiguity and ensure the enforceability of the claims, the main terms used in this invention are explained as follows. Unless otherwise stated, the terms in this specification have the same meaning.

[0067] 1. Enterprise Entity: This refers to the enterprise being diagnosed, characterized by at least an enterprise identifier (such as an enterprise ID or a mapping ID of the unified social credit code), enterprise name (optional), industry, region, etc. The enterprise entity is used to bind users, aggregate health check sessions, and determine the ownership of diagnostic artifacts and consultation data.

[0068] 2. Health Check Categories: This refers to the set of health check entry points categorized by business dimensions, including at least financial and tax health check, accounting health check, human resources health check, operations health check, governance health check, and strategic health check. Each category corresponds to a question bank, indicator system, scoring rule version, and report template.

[0069] 3. Indicator System: This refers to the set of indicators (which can be hierarchical) used to evaluate the status of an enterprise. For example, the primary indicator is "tax compliance," and the secondary indicators are "timeliness of declaration," "risk of abnormal deductions," and "reasonableness of tax burden," etc. There is a mapping relationship between questions and indicators, and multiple indicators can be used for a single question.

[0070] 4. Question Bank: This refers to a collection of questions organized according to the physical examination category. Questions can be of various types, including single-choice, multiple-choice, scale, fill-in-the-blank, numerical range, and attachment upload confirmation. The question bank contains metadata such as question stem, options, indicator mapping, weight, difficulty / discrimination parameters, cross-validation relationship markers, and follow-up / clarification question markers.

[0071] 5. Assessment Session: Refers to a user's complete quiz process for a specific enterprise entity under a specific health check category. Each session has a session identifier (session_id) used to associate quiz records, behavioral characteristics, data parsing fields, diagnostic artifacts, and chains of evidence and audits.

[0072] 6. Assessment State Z: Refers to the state object maintained during the adaptive assessment process, which includes at least: the set of covered indicators, the risk vector R, the answer credibility c, the candidate question set and the question parameter set, and the scoring rule version number v. r Calibration model version number v mThe system includes the sequence of questions already submitted, behavioral statistics, and a set of anomaly reason codes.

[0073] 7. Response credibility c: A quantitative indicator used to measure the authenticity and quality of responses, preferably normalized to [0, 1]. c is jointly determined by at least three types of consistency checks: logical consistency, data consistency, and behavioral consistency. In this invention, c is not only used for result labeling but also serves as a control variable for adaptive question selection and early termination threshold.

[0074] 8. Information Gain IG(q|Z): The contribution of question q to the reduction of uncertainty in state Z. It can be approximated by entropy reduction, information content or discrimination weighted uncertainty, and is used to drive adaptive question selection.

[0075] 9. Cover(q,Z): The degree to which question q supplements the uncovered indicators. It is used to ensure the coverage rate of key indicators and avoid diagnostic bias caused by insufficient coverage despite a small number of questions.

[0076] 10. Abn(q, Z): Used to express the priority of follow-up questions on "contradictions and gaps in evidence", which can increase the probability of issuing clarification questions / cross-checking questions when conflicts or data biases occur.

[0077] 11. Rule division P r The score is calculated by weighting, segmenting, and mutually exclusive deduction of answers according to the scoring rule engine (configurable rules / DSL), reflecting the expert rule system.

[0078] 12. Feature vector X: A structured input consisting of response results, behavioral statistics, enterprise data parsing fields, credibility and rule scores, used to calibrate model inference and consultation feedback.

[0079] 13. Calibration Model: A model used to map regular scores and feature inputs to calibration scores. This can be logistic regression, gradient boosting trees, lightweight neural networks, etc. The calibration model outputs a calibration score P. m It can also output explanatory information (such as feature contribution / importance).

[0080] 14. Calibration point P m The score output by the calibration model is used for final diagnostic level mapping and risk vector update.

[0081] 15. Risk Vector R: It includes at least the risk weights by indicator dimension, triggering reasons and their evidence index, and can be represented as an array of indicator risk weights and a set of cause entries, used for report interpretation, expert matching and backflow annotation.

[0082] 16. Diagnostic artifact A: A diagnostic output object designed to be "reproducible, interpretable, auditable, and cacheable." A must contain at least: diagnostic level, P... r Pm Sub-indicator scores, risk vector R, chain of evidence E, version information (v) r ,v m (e.g., generating timestamps).

[0083] 17. Evidence Chain E: A structured link used to explain the basis of the diagnostic conclusion, which should at least record the association of "topic - indicator - weight - contribution - risk item - feature source" to support recalculation consistency and audit review.

[0084] 18. Workpiece hash h A The digest obtained after normalizing and serializing diagnostic workpiece A is used for integrity verification and cache key construction.

[0085] 19. Audit hash chain digest h L The head / tail digest of the chain obtained by chaining key operation logs using chain hashing is used to prove the non-repudiation and tamper-proof nature of the process.

[0086] 20. Cache Key K: A key-value pair used for static caching, containing at least the enterprise identifier, category identifier, and version number (v). r ,v m ), workpiece hash h A With audit summary h L It is used for precise cache invalidation and consistency control.

[0087] 21. Gating conditions: Admission rules that control the entry of consultation return samples into the calibration sample pool, including at least the credibility threshold, conclusion consistency threshold, sample deduplication threshold, and report integrity threshold.

[0088] II. System Structure

[0089] like Figure 1 As shown, the system of this invention can be a cloud-based SaaS or a privately deployed system, consisting of a user terminal, an expert terminal, a management terminal, and a server terminal. The server terminal can be implemented using microservices or a layered architecture. The system includes at least the following modules, each of which can be implemented by program instructions executed on a processor:

[0090] User and Enterprise Management Module: Enables mobile phone number verification code verification, login status maintenance, enterprise entity establishment and binding, industry and regional information maintenance, session creation and access control.

[0091] Physical Examination Classification and Question Bank Module: Provides an entry point for physical examination classification, maintains question bank and question metadata, including cross-checking question comparison table, clarification question set, question parameters (difficulty, discrimination) and indicator mapping.

[0092] Credibility assessment module: Input answers, enterprise data parsing fields and behavior statistics, output credibility c and anomaly reason code set, and write the results to the assessment status Z.

[0093] Credibility-coupled adaptive routing module: Calculates the comprehensive score of each candidate question based on Z and selects the next question; at the same time, adjusts the early stopping threshold based on c to balance the number of questions and credibility.

[0094] Rule-Model Fusion Scoring Module: Calculates rule score P r Construct the feature vector X; call the calibration model to output P. m ; Generate diagnostic levels and update risk vector R; Manage rule version v r With model version v m .

[0095] Diagnostic artifact and evidence chain module: Generates diagnostic artifact A and evidence chain E, and writes them to (v r v m To support recalculation consistency.

[0096] Audit and Anti-Tampering Module: Generates audit logs for key operations such as login, answer submission, artifact generation, document upload, and expert report upload, and links them using a hash chain to form an audit summary h. L .

[0097] Staticization and caching module: Constructs a cache key K to statically cache diagnostic pages or workpiece data; when h A h L Or, when the version changes, perform a precise invalidation and trigger a recalculation / return to the source.

[0098] Consultation data management module: Implements format / size restrictions for uploaded data, generates file summary hashes, controls data distribution permissions, and allows for downloading / previewing.

[0099] Expert matching and consultation module: Based on risk vector R, expert profile and available time slots, it matches and schedules experts, distributes materials and receives expert reports.

[0100] Consultation feedback calibration module: Extracts labels and suggestion vector Y from expert reports, writes (X, Y) into the sample pool when the gating conditions are met, and triggers model updates and version recording.

[0101] III. Specific Technical Route for Implementing the Method of the Invention

[0102] like Figure 2 As shown, the method of the present invention can be implemented along the following route:

[0103] S1: Mobile number verification and account establishment, collection of basic enterprise information;

[0104] S2: Select the physical examination category and initialize the assessment status Z;

[0105] S3: Adaptive topic selection and early termination based on credibility coupling;

[0106] S4: Rule-Model Fusion Scoring Output P r P m With risk vector R;

[0107] S5: Generate diagnostic artifact A and evidence chain E and write version information;

[0108] S6: Calculate the artifact hash h A Construct the audit hash chain digest h L Generate a cache key K and invalidate it precisely;

[0109] S7: Upload consultation data and match experts, and return reports;

[0110] S8: The consultation report gating backflow forms a calibration sample pool, triggering model version iteration.

[0111] (I) S1: Mobile number verification, account establishment, and collection of basic enterprise information

[0112] Users submit their mobile phone number and verification code through the user interface. The server then calls the SMS gateway to verify the validity of the verification code and create an account. After account creation, a business entity record is created, collecting at least industry and regional information. Industry and regional information can be used for industry branch selection in the question bank, risk prior initialization, and expert matching and screening.

[0113] To ensure traceability, the system writes operations such as "register / login / bind enterprise" to the audit log. The log includes at least the operator's ID, IP address, timestamp, object ID (enterprise_id), and operation type (login / register / bind).

[0114] (II) S2: Select the physical examination category and initialize the assessment status Z

[0115] When a user selects a health check category (e.g., financial and tax health check), the server creates an assessment session and initializes the assessment state Z0. Z0 is recommended to include the following fields: company identifier, category identifier, session identifier; set of covered indicators I (initially empty); risk vector R (can be set to zero or initialized according to industry priors); credibility c (initially can be set to 1); set of question parameters (difficulty, discrimination, cross-validation flag, follow-up question flag); and candidate question set. Rating rules version number v r With calibration model version number v m (Read from configuration center and locked in session); behavior statistics container (answer duration, rollback count, modification count, device fingerprint digest, etc.); exception reason code set (initially empty). Lock version number (v r ,v mThe significance of this is that it uses a consistent version for calculation within the same session, avoiding the inability to recalculate or interpret diagnoses due to version switching in the middle.

[0116] (III) S3: Adaptive Topic Selection and Early Termination with Credibility Coupling

[0117] This step is one of the core creative realizations of this invention: within the adaptive questionnaire (similar to CAT) framework, the confidence level *c* is used both for reducing the score of the selected items and for adjusting the early stopping confidence threshold, thereby achieving a dynamic balance between "controllable number of items" and "reliable results," and demonstrating robustness to contradictory answers, noisy data, and strategic responses. The following provides a directly implementable method from the perspectives of data structure, computational process, boundary conditions, and engineering implementation.

[0118] 1. Response Collection and Behavioral Feature Recording (Sub-step S3-1)

[0119] For each question presented, the system records the following information on the user's end and submits it along with the answer: answering time (time from question display to submission); number of rollbacks and modifications; whether there were rapid clicks or abnormal screen switching (optional); device fingerprint summary (optional, under compliance requirements); network status and client version number (used for anomaly detection). The server writes this behavioral characteristic into the session behavior statistics container and records audit logs when necessary (e.g., suspicious behaviors such as abnormally frequent rollbacks or multiple accounts on the same device).

[0120] 2. Construction of credibility c (sub-step S3-2)

[0121] Trustworthiness c is determined by logical consistency, data consistency, and behavioral consistency, with the following weighted fusion as the preferred approach:

[0122] c = αc logic +βc data +γc beh ;

[0123] Parameter definition: α, β, γ are weighting coefficients, satisfying α + β + γ = 1; c logic Score for logical consistency; c data For data consistency score; c beh c represents the behavioral consistency score; c represents the overall credibility.

[0124] 2.1 Logical consistency c logic (Mutual verification question checklist)

[0125] The system pre-sets a "mutual verification question table" in the question bank to characterize the mutual exclusion / mutual verification / threshold consistency relationships between questions. For example:

[0126] 1) Mutual exclusion: Question q aAnswer "A complete invoice management system has been established = Yes" to question q b Answering "Long-term absence of invoice ledgers =

[0127] "Yes" cannot be true at the same time;

[0128] 2) Mutual verification: Question q c The answer "reported monthly and had no overdue payments" should be consistent with "number of overdue payments in the past 12 months = 0";

[0129] 3) Threshold consistency: The "Annual Sales Revenue Range" entered in the questionnaire should match the "Main Business Revenue Range" parsed from the uploaded report.

[0130] Consistency or deviation is within the tolerance range.

[0131] In implementation, mutual validation can be recorded in a rule table: (q1, q2, rule_type, rule_param, penalty), where rule_type indicates mutual exclusion / mutual verification / threshold consistency, rule_param describes the threshold or mapping function, and penalty is the deduction weight. Whenever a new answer arrives, the system retrieves the mutual validation constraints related to that question and calculates the degree of conflict, thereby updating c. logic If the conflict is significant, the abnormal reason code (such as L01 mutual exclusion conflict, L02 mutual verification failure, L03 threshold deviation) is written into the abnormal set of state Z for priority delivery of subsequent clarification questions.

[0132] 2.2 Data Consistency data (Compare with data fields)

[0133] When users upload documents (such as financial statements, tax returns, etc.) or fill out structured forms, the system performs field parsing (through template mapping or semi-structured extraction) to obtain a set of key fields (e.g., revenue, tax amount, cost structure, filing frequency, etc.). Data consistency verification compares the answered fields with the document fields; if the deviation exceeds a threshold, the system lowers the threshold. data Record the "fields to be clarified" and the deviation value.

[0134] For example, relative deviation or interval consistency checks can be used for numeric fields; interval fields can be mapped to interval codes before comparison; missing fields can be marked as "cannot be verified" and checked in the c... data The system incorporates a penalty for uncertainty. For fields requiring clarification due to deviations, the system will prioritize the corresponding clarification questions in subsequent anomaly items Abn(q, Z) to eliminate contradictions.

[0135] 2.3 Behavioral consistency beh (Abnormal response pattern recognition)

[0136] Behavioral consistency is used to identify behaviors such as "mechanical responses," "random responses," and "strategic responses." The system can employ rule-based or lightweight models.

[0137] The rules are as follows: If the duration of answering multiple questions is significantly lower than a reasonable threshold and there is no modification behavior, it is judged as abnormal; if there are frequent rollbacks and modifications in a short period of time and the modification direction is related to the mutual verification constraints, it is judged as a strategic adjustment; if the fingerprint of the same device corresponds to multiple enterprise entities answering frequently in a short period of time, it is judged as a risk of batch answering.

[0138] The model is as follows: Input the duration distribution, number of rollbacks, number of modifications, and number of screen switches into the anomaly detection model, outputting the anomaly probability, which is then mapped to c. beh .

[0139] To avoid accidental killing, c beh A "progressive deduction" approach can be adopted, and a manual review interface can be provided. The reason codes (such as B01 for excessively short duration, B02 for multiple rollbacks, and B03 for multiple accounts on the same device) are recorded in Z.

[0140] 3. Overall scoring of candidate questions and selection of the next question (sub-step S3-3)

[0141] In each round, the system selects candidates from the candidate set based on the evaluation status Z. Calculate your overall score and select the next question:

[0142] S(q)=IG(q∣Z)·g(c)+λ·Cover(q, Z)-μ·Abn(q, Z);

[0143] Parameter definitions: S(q) is the overall score of question q; IG(q|Z) is the information gain; g(c) is the credibility reduction function; Cover(q,Z) is the coverage contribution; Abn(q,Z) is the outlier; λ and μ are the weighting coefficients; c is the overall credibility.

[0144] 3.1 Implementation of Information Gain IG(q|Z)

[0145] Information gain can be implemented in various engineering ways, with the preferred approach satisfying two criteria: deployability and interpretability. Implementations include:

[0146] 1) Approximate information content based on discrimination index: The discrimination index parameter d is maintained in the question metadata. q And calculate the information content by combining the current risk vector uncertainty (e.g., the confidence interval width of a certain indicator);

[0147] 2) Based on entropy reduction: Treat the current probability distribution of the risk level of the indicator as uncertain, and estimate the entropy change of the distribution after the question is answered;

[0148] 3) Based on sample statistics: Offline statistical analysis of the contribution of questions to the final grade determination (e.g., mutual information, information gain).

[0149] It can be used online via table lookup.

[0150] Regardless of the implementation used, it is recommended to incorporate the IG calculation logic and version history into the scoring rules (version v). r Management is required to ensure recalculation consistency.

[0151] 3.2 Implementation of the reduction function g(c)

[0152] The reduction function is used to decrease the weight of "questions that pursue high information content" when credibility decreases, preventing the system from being misled by contradictory / noisy data. A simple and stable implementation can be adopted as follows:

[0153] g(c) = max(ε, min(1, c));

[0154] Parameter definitions: g(c) is the reduction result; c is the overall confidence level; ε is the minimum reduction lower limit to avoid the system losing its ability to distinguish due to excessively small values.

[0155] Alternatively, a segmented reduction approach can be adopted: when c is below the threshold, the weight of the clarification question is forcibly increased (through Abn or by directly switching the strategy mode) to improve robustness.

[0156] 3.3 Implementation of Cover(q, Z)

[0157] System maintenance key indicator set The coverage rate of the covered set I is calculated. When the coverage rate is insufficient, bonus points are awarded to questions that cover uncovered metrics. For example, if question q maps to an uncovered set of metrics ΔI, then Cover(q, Z) can be related to |ΔI| or the weights. The introduction of coverage contribution prevents the system from repeatedly asking questions on only local metrics and missing key dimensions.

[0158] 3.4 Implementation of the anomaly Abn(q, Z) (Clarification question preferred)

[0159] Exceptions are used to explicitly embed "contradiction resolution" into the topic selection mechanism. When an exception reason code or a field to be clarified exists in Z, the system will specify the corresponding set of clarification questions. Prioritize questions. In engineering implementation, a smaller Abn or a higher anomaly compensation weight can be set for clarification questions, thus making their S(q) prioritized over ordinary questions for a period of time. This mechanism is a significant difference between this invention and general authenticity verification systems: it not only detects anomalies but also feeds them back into the control loop of "what to ask next" and "when to stop".

[0160] 4. Early stopping mechanism for credibility reduction (sub-steps S3-4)

[0161] Early termination is used to reduce the number of questions and improve the user experience, but it must be coupled with confidence to avoid premature termination under noisy data. The system can maintain a base confidence level Conf (e.g., cumulative information or reciprocal of standard error) and introduce a confidence gating function f(c) to obtain the reduced confidence level Conf′:

[0162] Conf′=Conf·f(c);

[0163] Parameter definitions: Conf is the base confidence level; f(c) is the confidence gating function (f(c) can be taken as c or a piecewise function); c is the overall confidence level; Conf′ is the reduced confidence level.

[0164] The termination condition can be set as follows:

[0165] 1) The coverage rate of key indicators is not lower than the threshold (such as 80% or coverage by weight);

[0166] 2) And Conf′ is not lower than the threshold;

[0167] 3) Or the number of questions has reached the maximum limit;

[0168] 4) Or a forced termination event occurs (user logout, session timeout, etc.).

[0169] When c is low, Conf′ is suppressed, and the system will extend the questionnaire and add verification / clarification questions to reduce the risk of missed reports. Conversely, when c is high and coverage is sufficient, the system can terminate early, significantly reducing the number of questions.

[0170] 5. Boundary Conditions and Engineering Backup Strategies

[0171] To avoid unavailability of the adaptive system, it is recommended to include the following fallback strategy:

[0172] 1) Candidate problem exhaustion: If the candidate problem set is exhausted... If the filtered results are insufficient, the system can relax the filtering conditions or prioritize coverage.

[0173] Strategy-based minimum guarantee question.

[0174] 2) Low credibility: If c is consistently below the threshold, the system can mark "low credibility" in the evidence chain and suggest entering a consultation or manual review, while forcibly adding cross-check questions and restricting early termination.

[0175] 3) Missing Data: When data consistency verification cannot be performed, it is not directly judged as low credibility, but rather... data Set it to a neutral value and add an uncertainty flag, so that the system can supplement the evidence by adding questions.

[0176] 4) Unstable question parameters: For new questions or questions with insufficient samples, the weight of IG can be reduced, and mature question banks can be used first. The weight can be increased after accumulation.

[0177] The above strategies ensure that those skilled in the art can stably implement the S3 mechanism in real systems.

[0178] (iv) S4: Rule-Model Fusion Scoring Output P r P m With risk vector R

[0179] S4 is one of the core creative links: it integrates rule scoring with calibration models, preserving the interpretability of expert rules while leveraging data-driven approaches to improve accuracy and consistency, and outputting a risk vector R that can be used for consultation matching and backflow learning. A practical implementation method is given below.

[0180] 1. Rule division P r Calculation (sub-step S4-1)

[0181] Scoring rules can be described in the form of a configuration table or DSL, and at least support: question-to-index mapping; weight setting; segmented scoring and threshold mapping; mutually exclusive deductions and mutual verification weighting; upper / lower limit pruning; index aggregation (weighted sum, maximum value, piecewise function, etc.).

[0182] The rule engine takes conversation response records and necessary data fields as input, and outputs: total rule score P. r ; Sub-item score vector P r A set of risk trigger items (including trigger questions, trigger conditions, deduction / addition points, and corresponding indicators).

[0183] To ensure recalculation, the rule engine reads the version number v each time it is executed. r And write v into the diagnostic workpiece r And rule configuration summary (e.g., DSL hash).

[0184] 2. Construct the feature vector X (sub-step S4-2)

[0185] The feature vector X is the input to the calibration model and should be schema-stable and version-managed. X must contain at least the following categorical fields:

[0186] 1) Answer characteristics: key question answer codes, key question scores, number of cross-checking conflicts, number of clarification question triggers, coverage, etc.;

[0187] 2) Behavioral characteristics: Statistics on answering time (mean / quantiles), number of rollbacks, number of modifications, and count of error reason codes, etc.

[0188] 3) Data characteristics: Fields parsed from uploaded data (such as income range, tax burden range, and filing frequency) and their consistency deviation values;

[0189] 4) Credibility characteristics: c, c logic c data c beh ;

[0190] 5) Rule characteristics: P r Sub-indicator scores P r The number of trigger entries, etc.

[0191] The schema field of X is recommended to be recorded as feature_schema_hash, which corresponds to the model version v. m Binding is used to ensure that the inputs of different versions of the model are consistent and recalculated.

[0192] 3. Calibration Model Inference and Calibration Segmentation m (Sub-step S4-2)

[0193] Calibration model input X (with optional P) r Output calibration score P m P m It can be used for final ranking mapping, as well as for correcting biases in rule scores. If an interpretable model is used, the system can output feature contributions (e.g., Top-N contribution features) to aid in the construction of the chain of evidence.

[0194] Diagnostic level mapping can be defined by a configuration table, for example, by P. m Segmented mappings are categorized as low / medium / high risk. Mapping rules should also be included in version control (v) or a separate version, and a mapping version summary should be written into the artifact to ensure recalculation consistency.

[0195] 4. Construction of the risk vector R (sub-step S4-3)

[0196] The risk vector R should at least include: the risk weight, triggering cause, and evidence index for each indicator. Implementation methods include:

[0197] 1) Based on the item score P r By combining the model output with the importance of the indicator, the indicator risk weight is obtained;

[0198] 2) Map the set of triggering entries to risk cause entries;

[0199] 3) Associate the cause entries with the corresponding questions, data fields, and abnormal behavior cause codes to form an evidence index.

[0200] 4) The engineering value of R is reflected in: its use in report interpretation, expert matching, and reflow calibration label alignment.

[0201] (V) S5: Generate diagnostic artifact A and evidence chain E

[0202] S5 is one of the core creative links: it "artificializes" diagnostic results and generates an interpretable and recalculated chain of evidence, so that the output is not just a score / conclusion, but an auditable, verifiable, and accountable structured object.

[0203] 1. Field structure and generation process for diagnostic workpiece A

[0204] Diagnostic workpiece A preferably includes the following fields: company identifier, category identifier, session identifier; version number (v r v m ); Rule P r Calibration P m Diagnostic level; sub-indicator scores and risk vector R; chain of evidence E; generation timestamp and generation source (automatic / recalculation / supplementary data trigger), etc.

[0205] The generation process is as follows: After the S4 outputs the results, the artifact module encapsulates each result and performs consistency verification (field integrity, version consistency, session consistency), and then writes it to the database and cache system for staticization.

[0206] 2. The structure and construction rules of the chain of evidence E

[0207] Chain of evidence E records at least the following connections:

[0208] 1) question_id → indicator_id: The indicator that maps the question to the indicator;

[0209] 2) weight: The weight of the question / metric;

[0210] 3) Contribution: Contribution to the total score or the score of a certain indicator;

[0211] 4) risk_item: Risk item (trigger rules, trigger thresholds, etc.);

[0212] 5) Source: Feature source (question answers / data fields / behavioral features / mutual validation conflicts);

[0213] 6) version_ref: The corresponding rule version / model version reference.

[0214] Recommended rules for constructing the chain of evidence include:

[0215] 1) Directly write the trigger entries output by the rule engine;

[0216] 2) Write the key feature contributions (if any) of the model output into the "Model Evidence Section";

[0217] 3) Write the abnormal reason codes such as mutual verification conflict and data deviation into the "credibility evidence segment";

[0218] 4) Set priorities for each piece of evidence so that the user and expert terminals can display it hierarchically (e.g., displaying indicator risks first).

[0219] (Further details of the evidence follow).

[0220] Through E's structured records, a backtracking chain can be achieved from "diagnostic level → indicator risk → contribution topic / field → source evidence".

[0221] (vi) S6: Calculate h A , construct h L Generate cache key K and invalidate it precisely.

[0222] S6 is another core creative implementation of this invention: coupling "workpiece hash + audit hash chain" with cache key to achieve integrated control of security, reliability and high concurrency performance, solving the consistency problem of "old conclusions being cached and reused" that occurs in traditional systems after rule / model updates and data supplementation.

[0223] 1. Workpiece hash h A Calculation (sub-step S6-1)

[0224] After normalizing and serializing diagnostic workpiece A (sorting fields, standardizing floating-point precision, removing non-deterministic fields that do not affect the conclusion, or processing them separately), calculate the hash:

[0225] h A =Hash(Serialize(A));

[0226] Parameter definitions: A is the diagnostic artifact; Serialize(·) is the normalization serialization function; Hash(·) is the hash function; h A This is a summary of the workpiece.

[0227] The calculated h A Write to the artifact fields and use them for cache key construction. If any key field in the artifact content changes (e.g., evidence chain additions, version number changes, score changes), then h A This inevitably leads to changes, which in turn cause cache invalidation.

[0228] 2. Audit hash chain and audit digest h L (Sub-step S6-2)

[0229] The system links key operation logs in a chain: login, question submission, document upload, artifact generation, report upload, version reading, etc., can all be considered key logs. The chain hash calculation is as follows:

[0230] h i=Hash(h i-1 ||Log i );

[0231] Parameter definition: Log i For the i-th normalized log entry; h i-1 The hash of the previous log entry; h i `||` represents the current log hash; `||` is the concatenation symbol. The head / tail digest can be used as the audit digest `h`. L .

[0232] To ensure consistency, Log i The standardized format recommends fixing the field order and including: operator identifier, object identifier (enterprise_id, session_id, artifact_id, etc.), operation type, timestamp, IP, and key parameter summary (such as version number, file hash, etc.).

[0233] The uses of the audit chain include:

[0234] 1) Prove that the workpiece generation process and data transmission process have not been tampered with;

[0235] 2) Provide irrefutable evidence in disputed situations;

[0236] 3) Coupled with cache key: Cache reuse is prohibited when the audit chain changes or verification fails.

[0237] 3. Construction and precise invalidation of cache key K (sub-step S6-3)

[0238] The cache key K must contain at least the enterprise identifier, category identifier, version number, artifact hash, and audit digest.

[0239] K = <enterprise id ,category id v r ,v m h A h L >;

[0240] Parameter definition: enterprise id For corporate identification; category id For physical examination classification identification; v r v is the version number of the scoring rules. m For calibration model version number; h A For artifact hash; h L This is an audit summary.

[0241] Precision failure strategies include at least the following:

[0242] 1) Workpiece change failure: If recalculation of A leads to h A If K changes, the old cache will no longer be hit;

[0243] 2) Version Change Invalidation: If the rule or model version changes (v r or v m (Change), K changes, the old cache is automatically invalidated;

[0244] 3) Audit Change Invalidation: If the critical log chain changes (e.g., supplementary information, report upload), h L Changes occur, and the cache is not reused.

[0245] 4) Forced invalidation upon audit verification failure: If the audit chain verification fails, the system will force the cache not to be used and trigger a recalculation / alarm.

[0246] In terms of engineering implementation, the cache can be Redis / local cache / edge cache, etc., and static generation can generate HTML / JSON artifact files. The system checks the consistency of K before reading the cache, and can also check h. A h L Perform a quick verification to ensure that the displayed conclusions are consistent with the workpiece.

[0247] Through this design, the present invention binds the "secure and reliable link" and the "performance acceleration link" to the same control plane, avoiding the common problem in traditional systems of "cache hit but conclusion expired".

[0248] (vii) S7: Uploading consultation data, matching and distributing experts and sending back reports

[0249] 1. Consultation data upload and format / size limits

[0250] When a user initiates a consultation and uploads data, the system applies format whitelist and size threshold restrictions to the files and generates a file digest hash which is then written to the audit chain.

[0251] h F =Hash(FileBytes);

[0252] Parameter definition: FileBytes is the file byte stream; h F For file digest hashing. F File metadata (filename, type, size, uploader, timestamp) is written to the Log as log content. i And enter the hash chain.

[0253] Optionally, the system can also perform virus scans or content security checks and write the results to the audit log to improve the reliability of data processing.

[0254] 2. Expert matching and scheduling based on risk vector R

[0255] An expert profile should include at least: a set of expertise tags, industry experience tags, historical consultation quality scores, available time slots, and service scope. The matching process uses a risk vector R as its core input.

[0256] 1) Map the indicators with higher weights in R to expert label requirements;

[0257] 2) Calculate expert label coverage;

[0258] 3) Sort or optimize by combining available time periods with historical quality scores;

[0259] 4) Output the candidate expert set and schedule it.

[0260] During the consultation and distribution process, the system grants experts access to the consultation materials and records the distribution event in the audit chain to ensure that the flow of materials is traceable.

[0261] 3. Uploading, storing, and displaying expert reports

[0262] After experts upload their reports, the system performs format validation and generates a summary hash, which is then written into the audit chain. Simultaneously, the report is stored in association with diagnostic artifact A, forming a combined view of "online diagnosis - consultation report." When displayed on the user's end, the diagnostic level and risk vector are shown first, then expanded according to the chain of evidence, and finally the expert recommendations are presented, achieving a consistent interpretation path.

[0263] (VIII) S8: Consultation Report Gated Backflow Calibration and Model Version Update

[0264] S8 forms a closed loop of "diagnosis-consultation-iteration," which is key to improving long-term accuracy. Unlike the common "manual annotation backflow," this invention emphasizes gating conditions to avoid low-quality samples contaminating the model and records version change summaries in the workpiece to ensure traceability.

[0265] 1. Report structuring and suggestion vector Y generation

[0266] The preferred method is for experts to fill in key conclusions (such as risk level confirmation, key risk items, rectification suggestions, and evidence citations) using a structured form. The system then maps these conclusions to labels and a suggestion vector Y. If experts upload a text report, Y can also be generated through template parsing or manual structured extraction. Y should include at least: conclusion labels (aligned with the diagnosis level / risk item); suggestion category labels (tax and financial compliance, invoice management, process rectification, etc.); and key evidence citations (corresponding materials or question indexes).

[0267] 2. Gating conditions and sample pool writing

[0268] To ensure reflux quality, gating conditions are set:

[0269] 1) Credibility gating: c ≥ c min ;

[0270] 2) Report integrity gating: Key fields are complete, and evidence citations meet minimum requirements;

[0271] 3) Consistency Gating: Consistency is verified by multiple experts until a threshold is reached (optional);

[0272] 4) Deduplication gating: Sample similarity does not exceed a threshold to avoid the flood of duplicate samples.

[0273] Once the gating is satisfied, the returned sample (X, Y) is written into the calibration sample pool, and the returned event (including session ID, enterprise ID, category ID, c, gating result, and sample digest hash) is recorded into the audit chain to ensure traceability.

[0274] 3. Model update triggering and version management

[0275] Once the sample pool meets the triggering conditions (sample size, industry coverage, risk level coverage, etc.), the system will trigger or plan to trigger a model update: train / update and calibrate model parameters; generate a new model version number v′. m Calculate the model file hash and schema hash; publish to the configuration center and record a version change summary (reason for update, number of samples, changes in key metrics, etc.). After the update, use the new v′ in subsequent sessions. m Historical artifacts that have already been generated retain their old version numbers to ensure that historical results are recalculated and auditable.

[0276] In summary, this invention achieves robust evaluation of contradictory / noisy responses by simultaneously embedding the credibility criterion *c* into adaptive question selection and early termination criteria; and by using rule-based P... r With calibration P m The fusion scoring outputs a stable diagnostic level and risk vector R; interpretability and recalculation are achieved through diagnostic artifact A and evidence chain E; and artifact hash h is used to achieve this. A Audit Summary h L The strong coupling with the cache key K achieves a balance between security, reliability, and high concurrency performance; finally, continuous model iteration and version traceability are achieved through consultation-gated backflow. The above implementation provides directly implementable data structures, computational logic, and process details, which can be implemented and achieve the corresponding technical effects without creative effort by those skilled in the art.

[0277] Specific application example: Financial and tax health check-up for SMEs

[0278] I. Application Scenarios and Targets

[0279] Using a regional SME service platform as an example, the target category for the health check was "financial and tax health check." The platform conducts online health checks for enterprises from industries such as manufacturing, wholesale and retail, and information services. Enterprises submit basic information (industry, region, size, etc.), complete questionnaires online, and can upload consultation materials (VAT returns for the past 12 months, financial statement summaries, invoice statistics, etc.). The platform then outputs an online diagnostic report. When the risk is high or the enterprise actively requests it, it enters an expert consultation and generates a health check report, ultimately forming a diagnostic document.

[0280] II. System Deployment and Comparison Schemes

[0281] The system deployment adopts: user terminal (H5 / PC), expert terminal (Web), management terminal (Web), and the server includes an adaptive questionnaire module, a credibility verification module, a rule-model fusion scoring module, a diagnostic artifact / evidence chain module, a hash audit chain and cache key module, and a consultation feedback calibration module.

[0282] Two control systems were set up (to demonstrate the effectiveness of the technology):

[0283] Control group C1 (conventional questionnaire scoring system): fixed question order + fixed number of questions; no credibility-driven approach; outputs only rule scores based on scoring rules. r It generates reports; it does not artifact the chain of evidence; the logs are plaintext logs; and the cache uses a coarse-grained key of "enterprise ID + category ID".

[0284] This invention group E1 (integrated algorithm closed-loop system): enables credibility c coupled adaptive topic selection and early stopping; rule-model fusion scoring (P) r With P m ); Diagnostic workpiece A and evidence chain E; Workpiece hash h A With audit hash chain h L ;Cache key K= <enterprise,category,v r v m h A h L Precise failure detection; expert consultation and backflow calibration (gating).

[0285] III. Samples and Experimental Procedure

[0286] A total of 240 companies were selected as samples, including:

[0287] 120 companies followed the control group C1 procedure;

[0288] 120 companies followed the E1 process of this invention; the two groups of companies were kept consistent in terms of industry / size / regional distribution by stratified sampling.

[0289] The process is as follows:

[0290] 1) The company completes the financial and tax health check questionnaire;

[0291] 2) Enterprises upload consultation materials (not mandatory, but recommended for data consistency verification);

[0292] 3) The system generates an online diagnostic report;

[0293] 4) For enterprises whose diagnostic level is "high risk / requires review" or that have voluntarily applied for consultation, they will enter the expert consultation (60 enterprises will be randomly selected in each group);

[0294] 5) Expert reports serve as “artificial truth references” to evaluate the accuracy and misjudgment of online diagnoses and as candidates for return samples.

[0295] IV. Key Parameter Settings

[0296] Credibility c is calculated with the following weights: α = 0.4, β = 0.4, γ = 0.2 (logical / data / behavioral consistency);

[0297] Early cessation: Coverage of key indicators ≥ 80%, and reduced confidence level Conf′ ≥ 0.85;

[0298] Maximum number of questions: 40; Minimum number of questions: 18;

[0299] Consultation return gating: c≥0.75, consistency of expert conclusions≥0.8 (when multiple experts review), sample deduplication similarity≤0.9.

[0300] V. Definition of Evaluation Indicators

[0301] To demonstrate the effectiveness of the technology, the following indicators are set:

[0302] 1) Diagnostic accuracy: The percentage of online diagnostic levels that match the expert report levels;

[0303] 2) False positive rate (False High / False Low):

[0304] 3) False alarm of high risk: The online system judges the risk as high, but experts judge it as not high risk;

[0305] 4) High-risk cases were not reported: Online assessments indicated no high risk, but experts deemed them high risk.

[0306] 5) Review Precision: The percentage of companies marked "requires review" online that are ultimately identified by experts as having critical risks;

[0307] 6) Average number of questions and completion time: Reflects user experience;

[0308] 7) Consistency and traceability of results: After rule / model version changes or supplementary information, can it be guaranteed that the old cache is not misused and that diagnostic artifacts can be recalculated consistently?

[0309] 8) Concurrent access performance: QPS and P95 response time for homepage / diagnostic page under the same hardware.

[0310] VI. Experimental Results

[0311] Based on the above sample and process, the following statistics were obtained (accuracy calculations were performed using 60 companies that participated in the consultation as the "true value subset"):

[0312] Indicator Control group C1 Invention group E1 Diagnostic accuracy (in agreement with experts) 71.7％ 86.7％ False Low 18.3％ 6.7％ False High 10.0％ 6.6％ "Review required" hit rate 54.2％ 78.9％ Average number of questions (questions) 40 (fixed) 24.6 Average completion time (minutes) 16.4 10.1 Number of misuse of old conclusions after version change 7 times / month 0 times / month Concurrent diagnosis page P95 response (ms) 820 310

[0313] * "Number of times old conclusions are misused" refers to the number of times users are still shown the old version of the diagnostic page / old conclusions after the rule / model version is updated or the data is supplemented (through log comparison and sampling statistics).

[0314] VII. Key Processes

[0315] Select Company A (manufacturing). The difference between the "annual sales range" filled in in the questionnaire and the "main business revenue" in the uploaded data is significant.

[0316] Control group C1: The questionnaire was completed in the same order, and the respondents indicated "low risk". Expert consultation revealed that the company had abnormal tax burdens and an unreasonable input / output structure, classifying it as "medium-high risk," resulting in underreporting.

[0317] In this invention group E1: Data consistency verification is triggered, and credibility c decreases; the routing strategy automatically sends clarification questions such as "input and output structure," "declaration frequency," and "abnormal deductions"; the early termination threshold increases due to the reduction of c, and the system does not terminate prematurely; finally, the risk vector R increases significantly in the "tax compliance / input and output structure" dimension, and the diagnostic level is adjusted to "medium-high risk." The evidence chain E records the path of "source of deviation field - clarification question - change in contribution - risk item," which can directly locate the cause during expert review, proving that the results are interpretable and robust.

[0318] VIII. Consultation reflux calibration effect

[0319] From 60 consultation samples, our invention group E1 obtained 32 high-quality reflux samples ((X,Y)) through gating conditions. These were added to the calibration sample pool, and the model version (v) was updated. m After one iteration, offline replay evaluations were conducted on a subsequent batch (40 newly added companies). The diagnostic accuracy further improved from 86.7% to 89.5%, and the high-risk underreporting decreased from 6.7% to 5.0%. These results indicate that the consultation replay is controllable and effective under gating and will not degrade the model due to the introduction of noisy samples.

[0320] IX. Conclusion

[0321] Through the above practical applications and control experiments, it can be demonstrated that: in the enterprise online health check scenario, the present invention can significantly reduce the number of questions and the time, while improving the consistency between the diagnostic level and the expert conclusions. Through the integrated mechanism of the present invention, the results are more reliable, interpretable, recalculated and traceable, and have better response performance under high concurrency access conditions. Furthermore, through consultation backflow gating calibration, continuous iterative optimization is achieved, thereby achieving a comprehensive improvement in technical effect.

[0322] The foregoing description of embodiments of the present invention, through which those skilled in the art are able to implement or use the present invention, will be readily apparent to those skilled in the art. Various modifications to these embodiments will be readily apparent to those skilled in the art. The general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novelty disclosed herein.

Claims

1. A method for integrating online health check-up diagnosis and expert consultation for enterprises, characterized in that, Includes the following steps: S1, based on mobile phone verification code, completes identity verification and establishes account, and collects enterprise industry and regional information; S2, invoke the target physical examination classification question bank and initialize the assessment state Z, which includes the covered indicators, risk vector R, answer credibility c, and scoring rule version v. r With calibration model version v m ; S3, Adaptive topic selection and early termination with credibility coupling: Receive responses and behavioral characteristics, calculate c based on logical consistency, data consistency, and behavioral consistency; select the next topic based on the comprehensive score of "information gain IG × g(c) + coverage contribution − anomaly", and reduce the termination confidence threshold with c. Terminate when the coverage and confidence thresholds are met or the maximum number of topics is reached. S4, get the rule points according to the rules P r and P r The calibration score P is obtained by inputting the feature vector X into the calibration model. m Based on this, the diagnostic grade is determined and R is updated; S5 generates diagnostic artifact A and evidence chain E. Evidence chain E records "topic, indicator, weight, contribution, risk item, and feature source" and writes it to v. r v m To support recalculation; S6, Calculate the workpiece hash h A The key operation logs are formed into an audit hash h by hash chaining. L Generate a cache key K containing the enterprise identifier, category identifier, and rating rule version v. r Calibration model version v m h A h L And it becomes ineffective precisely as A or the audit changes; S7 receives consultation data and applies format / size restrictions, generates a file summary hash and inputs it into the audit chain, performs expert matching and scheduling based on the risk vector R, receives medical examination reports uploaded by experts and displays them in association with diagnostic artifact A; S8, extract the conclusion label and suggestion vector Y from the physical examination report, and write (X,Y) into the calibration sample pool to update v when the gating condition is met. m And record a summary of the changes.

2. The method according to claim 1, characterized in that, In step S1, a user registration or login request is received, identity verification based on the mobile phone number is completed, and a user account associated with the enterprise entity is established; basic enterprise information is received and stored, which includes at least the enterprise name, industry, and region. And / or, in step S2, a health check entry is provided to the user based on a preset health check category set, wherein the health check category includes at least financial and tax health check, financial health check, human resources health check, operational health check, governance health check and strategic health check.

3. The method according to claim 1, characterized in that, In step S3, adaptive questionnaire routing with credibility coupling is performed for each round of topic selection: S3-1. Receive the user's answer to the current question and collect the characteristics of the answering behavior; S3-2. Calculate the credibility of the response based on the responses and the data uploaded by the enterprise and / or structured form data. ,in It is determined by at least three types of consistency checks: logical consistency check, data consistency check, and behavioral consistency check. S3-3, For each candidate question Calculate the overall topic selection score The question with the highest score is selected as the next question, where: ； In the formula, In the evaluation state Next question Information gain or entropy decrease Regarding credibility The reduction function, For the title Coverage contribution to uncovered indicators For penalties or compensation items for abnormal / mutual verification questions, and These are the weighting coefficients; S3-4, Based on credibility The termination confidence level is reduced, and the questionnaire is terminated if either of the following conditions is met: the preset minimum coverage threshold is reached and the reduced confidence level is not lower than the threshold; or the maximum number of items is reached; and the assessment status at the time of termination is recorded as follows. ; Preferably, the logical consistency check includes at least a list of mutually verifying question pairs, and the list of mutually verifying question pairs defines at least one set of mutually exclusive or mutually verifying constraints; when constraints conflict, the reliability is reduced. And switch the routing for the next question to the verification question priority mode; Preferably, the data consistency verification includes at least comparing the range values ​​involved in the answer with the corresponding field range values ​​obtained from parsing the uploaded data; when the deviation exceeds a threshold, Set to positive compensation to prioritize the deployment of follow-up questions used to clarify the source of the deviation; Preferably, the behavior consistency verification is based at least on the distribution of answer duration, the number of rollback modifications, device fingerprints, and the answer characteristics of multiple accounts on the same device to generate an anomaly score; when the anomaly score exceeds the threshold, the reduced confidence threshold in the termination condition is increased or at least one set of verification questions is forcibly added.

4. The method according to claim 1, characterized in that, In step S4, the following rules are applied to the response results after termination: Model fusion scoring and risk vector update: S4-1. Obtain rule scores according to configurable scoring rules. And generate scores for each sub-indicator; S4-2. Extracting feature vectors from responses, company information, and behavioral characteristics. ,Will and Input calibration model to obtain calibration score ; S4-3, with The mapping yields the diagnostic level, and the risk vector is output. The risk vector It should at least include risk weights and triggering reasons by indicator dimension; Preferably, the rule is divided The scoring rules are parsed and executed using a DSL (Specific Language for Scoring Rules), which at least supports question weights, indicator mapping, segmented scoring, and mutually exclusive deductions. The diagnostic artifacts are written into the DSL version hash to ensure rule reproducibility.

5. The method according to claim 1, characterized in that, The diagnostic artifacts described in step S5 include at least: diagnostic level, total score, sub-indicator scores, and risk vector. and the chain of evidence The chain of evidence Record at least the relationships between "topic - indicator - weight - score contribution - risk item - feature source" and include the scoring rule version number. With calibration model version number To support recalculation consistency; And / or, in step S6, the tamper-proof audit chain and cache key coupling are performed on the diagnostic artifact: S6-1, Calculate the diagnostic artifact content digest hash. S6-2. Link and store the key operation logs of the diagnostic workpiece in a hash chain to form an audit chain hash. S6-3, Generate cache key ,in At least the following should be included: company logo, medical examination category logo, and scoring rule version number. With calibration model version number With hash Composition, and based on changes in diagnostic artifacts or chains of evidence. Triggering precise cache invalidation; Preferably, the cache key Further includes audit chain hashes When any critical operation log link breaks or integrity verification fails, cache reuse is prohibited and diagnostic artifacts are forcibly regenerated.

6. The method according to claim 1, characterized in that, In step S7, the system receives expert consultation requests and uploads of consultation materials initiated by the user, applies format and size restrictions to the uploaded files, generates a file digest hash, and writes it into the audit chain; based on risk vectors... Perform expert matching and scheduling optimization to obtain a set of candidate experts and distribute consultation materials to the experts; receive medical examination reports uploaded by experts, associate them with diagnostic workpieces, store them, and then send them back for display. Preferably, the expert matching and scheduling optimization are multi-objective ranking or multi-objective optimization, and the objective function includes at least a risk vector. The system calculates the matching degree with the expert's expertise tags, the quality score of historical consultations, the conflict cost of available time slots, and industry similarity; and outputs at least one set of candidate experts that the user can confirm. And / or, perform consultation reflux calibration in step S7: extract conclusion labels and recommendation vectors from the expert medical examination report. When the preset gating conditions are met, Write the data into the model calibration sample pool and trigger or plan to trigger the corresponding version of the calibration model update, while recording the return identifier and model version change summary in the diagnostic workpiece.

7. An online health check-up and expert consultation system for enterprises, used to implement the method described in any one of claims 1-6, characterized in that, include: The module includes: User Management Module, Enterprise Information Module, Physical Examination Classification and Question Bank Module, Adaptive Routing Module, Credibility Verification Module, Rule-Model Fusion Scoring Module, Diagnostic Workpiece and Evidence Chain Generation Module, Anti-Tampering Audit Module, Cache Key Generation and Static Caching Module, Consultation Data Management Module, Expert Matching and Consultation Module, and Consultation Backflow Caching Module. in: The adaptive routing module is used to base routing on information gain and coupled with reliability. Calculate the overall score and select the next question; The credibility verification module is used to output the credibility determined by logical consistency, data consistency, and behavioral consistency. ; The diagnostic artifact and evidence chain generation module is used to generate a document containing a scoring rule version number. With calibration model version number The chain of evidence relating to the source of the features; The anti-tampering audit module is used to generate hash chains for key operations on diagnostic workpieces and consultation data; The cache key generation and static cache module is used to construct cache keys using enterprise identifier, physical examination category identifier, version number and workpiece hash and to achieve precise invalidation. The consultation feedback calibration module is used to feed the tags and suggestions extracted from the expert report back to the calibration sample pool and trigger the model version update record.

8. The system according to claim 7, characterized in that: The anti-tampering audit module writes at least the operator identifier, object identifier, operation type, timestamp, client IP, current hash and previous hash to each audit record, and provides an integrity verification interface to verify the continuity of the hash chain; And / or, the consultation feedback calibration module sets gating conditions, which include at least an expert conclusion consistency threshold, a sample deduplication threshold, and a confidence level. Threshold; only when the gating condition is met will samples be written to the calibration sample pool and the calibration model version number be allowed to be updated. .

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the method of any one of claims 1–6.

10. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the method of any one of claims 1–6.

Images