Intelligent recruitment analysis method and system based on evidence chain

Abstract

The application discloses an intelligent recruitment analysis method and system based on evidence chains, and particularly relates to the technical field of computer data processing.The method comprises the following steps: receiving a post requirement text, post metadata and candidate multi-source input data, establishing an evaluation transaction and generating an evaluation tracking number; associating a post portrait version number, a capability dimension model version number, a scoring rule version number, a model service version number and a prompt parameter version number, forming a version anchoring record and determining a credible historical version baseline; structurally processing the multi-source input data, extracting evidence units, generating input snapshots and initial evidence chain maps; executing evaluation calculation under the constraint of the version anchoring record, outputting candidate evaluation results and generating output snapshots; and performing same-version recalculation or differential recalculation according to a recalculation request, and generating recalculation audit results. The version anchoring, evidence tracing and result review of the recruitment evaluation process are realized, and the interpretability and review reliability of the recruitment analysis results are improved.

Description

Technical Field

[0001] This invention relates to the field of computer data processing technology, and more specifically, to an intelligent recruitment analysis method and system based on evidence chains. Background Technology

[0002] In the ongoing recruitment process, the same recruitment system often needs to serve multiple recruitment batches over a long period. The departments responsible for these positions also continuously adjust job requirements, interview dimensions, scoring weights, and screening criteria. Simultaneously, model services, prompt parameters, and related rule configurations may change with system upgrades. Existing recruitment analysis systems typically evaluate candidates based on the currently effective job profile, competency dimension model, scoring rules, model services, and prompt parameters, combined with their resumes, Q&A records, and interview records. While these systems can output evaluation reports, they often fail to uniformly bind and freeze the actual versions of job profiles, competency dimension models, scoring rules, model services, prompt parameters, and input content used in the current evaluation. They also fail to create corresponding input and output snapshots for the evaluation results. Consequently, when candidates object to the screening conclusions, companies conduct internal audits, or need to review historical recruitment conclusions, the system can usually only provide explanations under the current conditions. It is difficult to reconstruct the calculation conditions and evidentiary basis corresponding to the formation of historical conclusions, making it difficult to recalculate historical evaluation results under the original conditions. This, in turn, affects the traceability, interpretability, and reliability of the recruitment analysis results.

[0003] To address the above problems, this invention proposes a solution. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an intelligent recruitment analysis method and system based on evidence chains to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] The intelligent recruitment analysis method based on evidence chains includes the following steps:

[0007] The system receives job requirement text, job metadata, and multi-source input data of candidates, establishes an assessment transaction, and generates an assessment tracking number. It parses the job requirement text, extracts job responsibilities, qualifications, and competency requirements, and generates a job profile based on the job metadata. It reads the assessment version object identifier corresponding to the current assessment transaction, associates it with the assessment tracking number to form a version anchoring record, calculates the version anchoring reliability comprehensive quantity based on the version time synchronization deviation, version dependency closed loop completeness, and frozen registration completeness, and determines the reliable historical version baseline when a preset threshold value is met.

[0008] Under the trusted historical version baseline, structured processing is performed on multi-source input data to form structured input results. The structured input results are mapped and evidence is extracted based on the capability dimensions in the job profile. The mapped records are solidified into evidence units, and the structured input results, evidence unit indexes, and mapping relationships are serialized to generate an input snapshot.

[0009] Under the constraints of version anchoring records, the evaluation version object bound to the evaluation tracking number is invoked. The confidence level of the evidence unit in the input snapshot is determined based on the reliability of the source, the degree of matching between the evidence and the capability dimension, the freshness of time, and the conflict status. The dimension score set and the comprehensive score are aggregated by capability dimension, and then the evidence chain object and the candidate evaluation result are generated. The input snapshot summary value, the output snapshot summary value, the version anchoring record, and the evaluation tracking number are written into the audit trace object.

[0010] Based on the recalculation request, retrieve the corresponding version anchoring record, input snapshot, input snapshot summary value, and output snapshot summary value, perform same-version recalculation, and perform differential recalculation when it is necessary to analyze the impact of version changes; calculate the recalculation offset attribution composite amount based on the dimension score offset, evidence citation drift, and conflict identification offset, and generate recalculation audit results.

[0011] In a preferred embodiment, the job metadata includes at least the job name, job level, department, and city; the assessment version object identifier includes at least the job profile version number, competency dimension model version number, scoring rule version number, model service version number, and prompt parameter version number; and the version anchoring record includes at least the assessment tracking number and the assessment version object identifier corresponding to that assessment tracking number.

[0012] In a preferred embodiment, the version time synchronization deviation is used to characterize the dispersion of the object identifiers of each evaluation version at the freeze time; the version dependency closed loop completeness is used to characterize whether the dependency relationship between the job profile, capability dimension model, scoring rules and their scoring items, and evidence types is completely closed; the freeze registration completeness is used to characterize whether the key objects related to the current evaluation transaction have been registered before entering the formal evaluation; when the version anchor credibility comprehensive quantity does not reach the preset threshold, the current evaluation transaction will not enter the subsequent evaluation calculation process.

[0013] In a preferred embodiment, the multi-source input data includes at least resume files, initial screening Q&A records, and interview audio and video data; when the multi-source input data is subjected to structured processing, a resume structured result, a Q&A structured result, and an interview structured result are formed.

[0014] In a preferred embodiment, when mapping and extracting evidence from the structured input results based on the capability dimensions in the job profile, the mapped records are solidified into evidence units; the evidence unit records at least the source type, source location, time information, associated capability dimensions, and evidence fragment summary value; the input snapshot includes at least the structured input results, the evidence unit index, and the mapping relationship between the evidence unit and the original source.

[0015] In a preferred embodiment, the source reliability is mapped by the source type of the evidence unit, and the source type includes at least resumes, Q&A, and interviews; the conflict state is used to characterize the degree of contradiction between different evidence units under the same ability dimension; the key evidence missing situation is used to characterize whether the target ability dimension required by the job profile lacks evidence units that meet the scoring requirements.

[0016] In a preferred embodiment, the evidence chain object is used to associate the score nodes of the capability dimension with the evidence units involved in the score calculation and related conflict items; the candidate evaluation result includes at least the dimension score set, the comprehensive score, the risk label, and the evidence citation list; the output snapshot is used to archive the dimension score set, the comprehensive score, the risk label, and the evidence chain object association result actually output in this evaluation.

[0017] In a preferred embodiment, the same version recalculation refers to reloading the same evaluation version object according to the evaluation version object identifier in the version anchoring record, and calling the input snapshot corresponding to the evaluation tracking number to re-execute the evaluation calculation under the same input boundary and version boundary as the historical evaluation, so as to determine whether the historical evaluation results can be obtained again under the original calculation conditions.

[0018] In a preferred embodiment, the differential recalculation refers to loading the currently effective evaluation version object and re-executing the evaluation calculation while keeping the input snapshot unchanged to obtain the current output result; when the total amount of recalculated offset attribution is higher than the preset differential threshold, attribution analysis is performed on the changes in job profile version, capability dimension model version, scoring rule version, model service version, and prompt parameter version, and the dominant offset source is determined.

[0019] In a preferred embodiment, the following modules are included:

[0020] The version anchoring module is used to receive job requirement text, job metadata, and multi-source input data of candidates, establish an assessment transaction and generate an assessment tracking number; parse the job requirement text, extract job responsibilities, job conditions and ability requirements, and generate a job profile in combination with job metadata; read the assessment version object identifier corresponding to the current assessment transaction, associate it with the assessment tracking number to form a version anchoring record; calculate the version anchoring reliability comprehensive quantity based on the version time synchronization deviation, version dependency closed loop completeness, and frozen registration completeness; and determine the reliable historical version baseline when a preset threshold value is met.

[0021] The input sealing module is used to perform structured processing on multi-source input data under the trusted historical version baseline to form structured input results, map and extract evidence based on the capability dimensions in the job profile, solidify the mapped records into evidence units, and serialize the structured input results, evidence unit index and mapping relationship to generate an input snapshot;

[0022] The evaluation output module, under the constraint of version anchoring record, calls the evaluation version object bound to the evaluation tracking number, determines the confidence level of the evidence units in the input snapshot based on the reliability of the source, the degree of matching between the evidence and the capability dimension, the freshness of time, and the conflict status, and aggregates them by capability dimension to generate a set of dimension scores and a comprehensive score, thereby generating an evidence chain object and candidate evaluation results, and writes the input snapshot summary value, output snapshot summary value, version anchoring record, and evaluation tracking number into the audit trail object;

[0023] The recalculation audit module is used to retrieve the corresponding version anchoring record, input snapshot, input snapshot summary value and output snapshot summary value according to the recalculation request, perform same-version recalculation and differential recalculation, and calculate the recalculation offset attribution comprehensive quantity based on the dimension score offset, evidence citation drift and conflict identification offset, and generate recalculation audit results.

[0024] The technical effects and advantages of this invention are as follows:

[0025] This invention does not merely record historical recruitment evaluation results retrospectively. Instead, before the formal evaluation begins, it pre-gated the evaluation process by assessing whether the current evaluation transaction meets the conditions for joint freezing, joint playback, and joint recalculation through version time synchronization deviation, version dependency closure completeness, and freeze registration completeness. Only when a reliable historical version baseline is established is subsequent evaluation computation allowed. Simultaneously, this invention uniformly serializes the structured inputs, evidence unit indexes, and mapping relationships that actually enter the computation chain and generates input snapshot summary values. This ensures that the input boundaries, version boundaries, and evidence boundaries in the recruitment evaluation process are synchronously fixed. This prevents the mixing of job profiles, scoring rules, model services, or prompt parameters with the same evaluation transaction during asynchronous version updates. It also reduces historical state distortion caused by changes in field order, time format differences, or subsequent data additions, improving the boundary stability and result consistency of the historical evaluation process.

[0026] Furthermore, in the recalculation stage, this invention does not directly regenerate new results using the currently effective version. Instead, it first restores the historical calculation boundaries based on the audit trace objects and performs a same-version recalculation. Then, while keeping the input snapshot unchanged, it individually replaces different version sources and performs local differential recalculation to determine the dominant offset source by combining dimensional score offset, evidence citation drift, and conflict identification offset. Thus, this invention achieves a closed-loop technical processing from pre-assessment gating, in-assessment binding, to post-assessment attribution. It can not only determine whether historical assessment results can be obtained again under the original calculation conditions, but also pinpoint whether the result change originates from changes in the job profile version, capability dimension model version, scoring rule version, model service version, or prompt parameter version. This improves the reliability of the recruitment analysis system's review, its audit support capabilities, and the credibility of its practical application in asynchronous version change scenarios. Attached Figure Description

[0027] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings;

[0028] Figure 1 This is a flowchart illustrating the intelligent recruitment analysis method based on the chain of evidence of the present invention.

[0029] Figure 2 This is a schematic diagram of the evaluation output driven by the chain of evidence.

[0030] Figure 3 This is a diagram illustrating historical recalculation auditing and differential attribution.

[0031] Figure 4 This is a schematic diagram of the structure of the intelligent recruitment analysis system based on the chain of evidence of the present invention. Detailed Implementation

[0032] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the job requirement text is parsed to generate a job profile. The job profile version number, capability dimension model version number, scoring rule version number, model service version number, and prompt parameter version number associated with the current assessment transaction are read and uniformly associated with the assessment tracking number to form a version anchoring record. The version anchoring credibility comprehensive quantity is calculated based on the version time synchronization deviation, version dependency closed loop completeness, and frozen registration completeness. When the version anchoring credibility comprehensive quantity meets the preset threshold value, the current assessment transaction is determined to form a credibility historical version baseline.

[0033] Step 2: Under the trusted historical version baseline, the multi-source input data of the candidates is structured and evidence is extracted according to the ability dimension in combination with the job profile to form evidence units; the structured inputs and evidence units participating in the current evaluation transaction calculation link are serialized to generate input snapshots and input snapshot summary values, and an initial evidence chain graph is generated according to the relationship between evidence units;

[0034] Step 3: Under the constraints of the version anchoring record, call the job profile version, capability dimension model version, scoring rule version, model service version, and prompt parameter version bound to the assessment tracking number to evaluate and calculate the evidence units in the input snapshot, and obtain the dimension score set and the comprehensive score; generate an evidence chain object based on the dimension score set and the evidence units involved in the calculation, output the candidate assessment results, and generate an output snapshot and an output snapshot summary value. Write the input snapshot summary value, the output snapshot summary value, the version anchoring record, and the assessment tracking number into the audit trace object;

[0035] Step 4: Receive the recalculation request and retrieve the corresponding version anchoring record, input snapshot summary value, and output snapshot summary value from the audit trace object according to the assessment trace number in the recalculation request to restore the historical calculation boundary; perform same-version recalculation based on the restored historical calculation boundary, and perform differential recalculation when it is necessary to analyze the impact of version changes; calculate the recalculation offset attribution composite amount based on the dimension score offset, evidence citation drift, and conflict identification offset, and generate the recalculation audit result.

[0036] Specifically:

[0037] Step 1: The system first receives the job requirement text, job metadata, and multi-source input data of candidates corresponding to the current recruitment task. The job metadata includes at least the job name, job level, department, and city. The system establishes a one-time assessment transaction for the received job requirement text and job metadata, generating a unique assessment tracking number for this transaction. This tracking number is used throughout the input archiving, scoring calculation, output snapshot generation, and subsequent recalculation and auditing processes of this recruitment assessment. Subsequently, the job requirement text is parsed to obtain a set of job elements. A job profile is then generated by combining the job template configuration, and the version number of the job profile corresponding to this assessment transaction is extracted simultaneously. At the same time, the system also reads the version numbers of the capability dimension model, scoring rules, model service, and prompt parameters associated with this assessment transaction, and associates these version numbers with the assessment tracking number to form a version anchoring record corresponding to this assessment transaction. The purpose of this approach is not simply to save a number of version numbers, but to establish a unique, closed, and replayable historical calculation boundary for subsequent scoring calculations. This avoids job profiles, scoring rules, or model parameters from being mixed into the same evaluation transaction after asynchronous updates, which would make it impossible to explain under what calculation conditions the historical conclusions were formed.

[0038] After establishing the version anchoring record, it does not directly proceed to the formal scoring. Instead, it further determines whether the current evaluation transaction has conditions for joint freezing, joint playback, and joint recalculation. Therefore, the version anchoring trustworthy comprehensive quantity is calculated based on the version time synchronization deviation, the version dependency closed-loop completeness, and the freeze registration completeness. The version anchoring trustworthy comprehensive quantity characterizes whether the multiple version objects depended on in this evaluation transaction already have a trustworthy historical version baseline. Preferably, the version anchoring trustworthy comprehensive quantity can be expressed as:

[0039] ;

[0040] in, This indicates the version-anchored trusted synthesis quantity. , and This represents a preset weight that satisfies: ;For example , , .

[0041] in, This represents the version time synchronization deviation, which characterizes the degree of dispersion of the job profile version, capability dimension model version, scoring rule version, model service version, and prompt parameter version bound to the firm in this assessment at the time of freezing. This represents the completeness of the version dependency closed loop, used to characterize whether the dependency relationships between the job profile version, the competency dimension model version, the scoring rule version and its scoring items, and the evidence type are completely closed. This represents the completeness of the frozen registration, indicating whether key objects related to this assessment transaction have been registered before entering the formal scoring process. Since this comprehensive quantity directly determines whether the current assessment transaction can serve as the basis for subsequent input snapshot archiving and version-based recalculation, its function is not statistical display, but rather a pre-entry control quantity for the scoring process.

[0042] Among them, version time synchronization deviation This can be directly calculated from the timestamp field in each version record table. The system reads the timestamps corresponding to the job profile version, the capability dimension model version, the scoring rule version, the model service version, and the prompt parameter version, and records them as follows: , , , and Then, calculate the difference between the maximum and minimum values ​​among the above timestamps to obtain the time span of the set of versions bound to the transaction being evaluated. Furthermore, this time span is compared with the system's configured allowed freeze time window. After normalization, we get:

[0043] ;

[0044] in,

[0045] ;

[0046] In the formula, Timestamps for different versions of the job profile. For the timestamp of the capability dimension model version, For the timestamp of the rating rule version, Timestamps for model service versions To indicate the timestamp of the parameter version, This sets the preset allowable freeze time window. When... The smaller the value, the more likely the different versions of the objects belong to the same freeze batch; when A larger value indicates that the version objects referenced by the assessment firm are more likely to span multiple update periods, and the historical calculation boundaries are more unstable. The above... , , , , All can be read directly from the version record table, the aforementioned It can be read directly from the system configuration table, so this parameter can be calculated directly and its basic parameters are all available.

[0047] Version dependency closed loop complete quantity This is used to characterize whether a complete and resolvable dependency loop is formed between the various versions of objects bound to the current evaluation transaction. Considering that the competency dimensions in the job profile need to correspond to the competency dimension model, and the scoring items and evidence types in the scoring rules need to correspond to both the job profile and the competency dimension model, the system pre-establishes a version dependency template and verifies the dependencies in the current evaluation transaction item by item according to this template. Preferably, let the total number of dependency verification items required for this evaluation transaction be... The number of items that passed the verification was The complete version dependency loop can then be represented as:

[0048] ;

[0049] The dependency verification items include at least: whether the job profile version can be associated with the capability dimension model version, whether the scoring rule version can be associated with the capability dimension model version, whether all scoring items in the scoring rule can be mapped to the capability dimensions in the current capability dimension model, and whether all evidence type requirements in the scoring rule can find corresponding records in the evidence type set required by the current job profile. It can be directly determined based on the preset dependency template, the This can be obtained by searching and statistically analyzing the version number in the current evaluation transaction item by item in the dependency table, capability dimension mapping table, and evidence type mapping table. The closer it is to 1, the more complete the dependencies of the current version set are, and the more suitable it is as a unified computational baseline for subsequent recalculated transactions.

[0050] Complete freezing registration This is used to indicate whether the key objects corresponding to this assessment transaction have been registered before entering the formal scoring stage. Since subsequent recalculation requires at least the assessment tracking number, job profile version number, competency dimension model version number, scoring rule version number, model service version number, prompt parameter version number, input snapshot registration slot, and output snapshot reserved slot, the system sets these objects as mandatory registration items. Preferably, the total number of mandatory registration items is set to... The number of projects that have been registered is Then the completeness of the frozen registration can be expressed as:

[0051] ;

[0052] In a preferred embodiment, Take 8, which corresponds to the eight key objects mentioned above; This represents the number of items out of these eight that have been successfully written into the evaluation transaction registry and whose field status is valid. When The closer the value is to 1, the more complete the recalculation and registration basis is before the formal scoring begins; conversely, if some key objects have not yet been registered, even if some version numbers have been obtained, it is not advisable to directly enter the subsequent scoring and output generation process.

[0053] After obtaining the version anchored trusted comprehensive quantity Then, the system compares it with a preset threshold value. Comparison, for example =0.5. When When the system determines that the current evaluation transaction has formed a credible historical version baseline, it allows entry into step two, where structured processing, evidence unit mapping, and input snapshot sealing are performed on the multi-source input data; when When this happens, the system marks the current evaluation transaction as pending freezing, neither outputting a formal score nor proceeding to the subsequent evaluation report generation process. Instead, it waits for version synchronization, dependency completion, or registration and repair to be completed before re-executing this step. In this way, step one not only completes the documentation and version number recording of the evaluation transaction, but also transforms the determination of whether historical calculation conditions qualify for joint freezing into a quantifiable and gated technical process before the score calculation begins. This provides a stable premise for subsequent input snapshot sealing, output snapshot generation, and recalculation of the same version based on the evaluation tracking number.

[0054] Step Two: After establishing a reliable historical version baseline in Step One, the system begins to uniformly access and archive the multi-source input data of candidates corresponding to this evaluation transaction. This multi-source input data includes at least resume files, initial screening Q&A records, and interview audio / video data. The system first uniformly categorizes the raw data from these different sources into the candidate's original information record corresponding to the current evaluation tracking number. This ensures that all subsequent structured processing, evidence extraction, conflict identification, and scoring calculations are conducted within the same evaluation transaction boundary, avoiding confusion with historical evaluation records from other recruitment batches, other positions, or the same candidate. The purpose of this setup is to first fix which inputs were actually seen in this evaluation before performing subsequent structured and mapping processing, thus ensuring a one-to-one correspondence between subsequent input snapshots and version anchor records, avoiding the need to reassemble historical facts based solely on the latest candidate information in the database. Through this process, the system first uniformly categorizes resume files, initial screening Q&A records, and interview audio / video data into the evaluation transaction corresponding to the current evaluation tracking number, and then generates structured resume results, Q&A results, and interview results respectively. This ensures that subsequent evidence extraction, conflict identification, and scoring calculations are all based on the same input boundary.

[0055] During the resume data processing stage, the system parses the format, extracts fields, and organizes the content of the resume files uploaded by candidates. It divides the educational background, work experience, skills information, and project experience in the original resume into computable structured fields, forming a structured resume result. Educational background includes fields such as school, degree, major, and start and end dates; work experience includes fields such as employer, job title, start and end dates, and job description; skills information includes skill name, proficiency description, and contextual fields; and project experience includes project name, project duration, responsibilities, results description, and statements that may be relevant to the job profile's competency dimensions. For cases with errors in expression, inconsistent time formats, inconsistent field order, or duplicate information, priority is given to field merging and time standardization to ensure that the experience information of the same candidate can be compared using a unified format during subsequent mapping. The final structured resume result is no longer manually read but is a field-based result for subsequent competency dimension mapping and evidence citation. After the above processing, the resume data, divided into computable structured fields such as educational background, work experience, skills information, and project experience, is used for subsequent evidence mapping and scoring based on competency dimensions.

[0056] During the question-and-answer data processing, the system sequentially segments the questions and answers according to the question-and-answer rounds, and then structures and organizes them according to question number, answer content, answer time, and context identifiers to form a structured result. This structuring involves more than just preserving the question and answer text; it also records the order of questions in the evaluation process, the corresponding question topics, and answer fragments that may be relevant to the competency dimensions in the job profile. Multiple answers to the same question are merged into a single answer fragment. For statements that are significantly off-topic or irrelevant to the job requirements, the original records are retained while reducing the order in which they are subsequently used for competency mapping. The structured result helps to clearly express the content and time sequence of candidates' answers to different questions during the initial screening stage, facilitating subsequent competency dimension mapping and evidence indexing. After data processing, the above questions are organized into a structured result based on question number, answer content, answer time, and context identifiers, thus enabling the establishment of subsequent competency dimension mapping and evidence indexing.

[0057] In interview data processing, when real-time or offline interview audio and video are involved in the evaluation, the system transcribes the data and, if necessary, distinguishes speakers to obtain a structured interview result with time information. The structured interview result should include speech segments, start and end times of the segments, speaker identifiers, and the spoken text content. For multi-person interviews, it is necessary to distinguish between candidate statements and interviewer questions based on speaker identifiers to avoid misinterpreting questions as evidence of candidate competence. For audio and video containing background noise, pauses, interruptions, or verbal corrections, the original time position and contextual relationships must be preserved to reconstruct the original expression environment for subsequent conflict identification or evidence retrieval. The structured interview result preserves the chronological order of the interview expressions and provides temporal coordinates for subsequent evidence location.

[0058] After completing the structured resume, structured Q&A, and structured interview results, such as Figure 2 As shown, the system does not directly output candidate profiles. Instead, it first combines the job profiles frozen in step one to aggregate and extract evidence from the three types of structured results according to ability dimensions. Specifically, the system uses the ability dimensions in the job profiles as unified evaluation coordinates, mapping experience fields in resumes, answer segments in Q&A sessions, and statement segments in interviews to their corresponding ability dimensions, forming a multi-source evidence set of candidates corresponding to those ability dimensions. For each record mapped into an ability dimension, the system further solidifies it into an evidence unit. The evidence unit at least records the source type, source location, time information, associated ability dimension, and evidence segment summary value. Among them, the source type is used to distinguish whether the evidence comes from a resume, Q&A, or interview; the source location is used to identify the specific location of the evidence in the original input. For resumes, the source can be a field path or paragraph number; for Q&A, the source can be a question number and answer sequence number; and for interviews, the source can be a time interval and speech segment number; the time information is used to record the candidate's experience time or collection time corresponding to the evidence; the associated ability dimension is used to characterize which ability dimension the evidence is included in for subsequent scoring; and the evidence segment summary value is used to fix the key content of the evidence unit so that subsequent snapshot verification and recalculation comparison can be performed.

[0059] After the evidence units are solidified, all structured inputs, evidence unit indexes, source location information, and multi-source mapping relationships corresponding to this evaluation transaction are serialized in a unified manner to generate input snapshots and input snapshot summary values. The so-called input snapshot is not a separate storage of the original files, but rather a sealing of the input content that is actually incorporated into the subsequent calculation chain of this evaluation. It should at least include the structured resume results, structured question-and-answer results, structured interview results, evidence units under each competency dimension, and the mapping relationships between evidence units and their original sources. In this way, even if the candidate subsequently supplements their resume, answers questions again, or the interview transcription service is upgraded and the same audio is transcribed into different texts, the system can still reconstruct the actual input content used when the historical conclusion of this evaluation was reached based on the input snapshots and input snapshot summary values, rather than using the current state of the system to replace reconstruction. Preferably, when generating the input snapshot and input snapshot summary value, the system does not arbitrarily concatenate the above content. Instead, it first writes the resume structured result, question-and-answer structured result, interview structured result, evidence unit index, and the mapping relationship between the evidence unit and the original source in a preset order. It uses a unified time format for the time fields involved, a unified null value representation for missing fields, and a fixed arrangement order for multi-value fields before performing serialization processing. The purpose of this processing is to ensure that the same evaluation transaction under the same input boundaries can only obtain the same input snapshot expression form, so that the input snapshot summary value truly corresponds to the fixed calculation input, rather than producing different summary results due to changes in field order, null value writing, or time format differences. Therefore, in step four, when performing the same version recalculation, it can directly determine whether the recovered input boundary is consistent with the historical input boundary based on the input snapshot summary value.

[0060] Simultaneously with the generation of the input snapshot, the system also generates an initial evidence chain graph corresponding to this evaluation transaction based on the relationships between the evidence units. This initial evidence chain graph is used to characterize the supporting, conflicting, and missing relationships between different evidence units. A supporting relationship refers to two or more evidence units from different sources pointing to the same capability dimension and mutually corroborating each other. A conflicting relationship refers to evidence units from different sources exhibiting significant inconsistencies in terms of time, years, responsibilities, or project experience. A missing relationship refers to a situation where, although the key capability dimensions required for the job profile have scoring requirements, sufficient evidence units to support the calculation of that dimension have not yet been formed. The purpose of generating the initial evidence chain graph here is not to complete the final scoring in advance, but rather to fix how evidence enters the subsequent scoring chain, ensuring that subsequent consistency checks and evidence chain binding are based on the solidified evidence units.

[0061] Step 3: As Figure 3As shown, after a credible historical version baseline has been established in step one and the input snapshot has been sealed and evidence units have been solidified in step two, the system begins to perform formal assessment calculations under the constraints of the version anchoring record. At this time, the system calls the job profile version, capability dimension model version, scoring rule version, model service version, and prompt parameter version bound to the current assessment tracking number, performs dimension-by-dimensional calculations on the evidence units solidified in the input snapshot, and outputs the dimension score set, comprehensive score, and risk label. Here, "same version condition" means that the job profile, capability dimension definition, scoring rules, model calling method, and prompt parameters participating in this assessment are all consistent with the version information frozen and registered in step one. The candidate input content read for assessment is also based on the input snapshot sealed in step two, and the latest status in the candidate database is no longer read back. In this way, the assessment process is limited to the same input boundary, the same version boundary, and the same evidence boundary, so that the subsequently generated assessment report and output snapshot can correspond one-to-one with the corresponding historical conditions, and also provides a premise for subsequent same-version recalculation. Furthermore, in steps three and four, whenever a model service call is involved, the system uses the model service version number and prompt parameter version number registered in the version anchoring record as the actual call basis, instead of using the current system default configuration or subsequent updated configuration. If the corresponding model service version number or prompt parameter version number is missing in the audit trace object, the current assessment transaction will not enter the formal assessment, or the current recalculation request will not enter the same version recalculation. In this way, the model service call boundary, along with the job profile version, capability dimension model version, and scoring rule version, are included in the same historical calculation boundary, thereby avoiding the distortion of recalculation results caused by only restoring some version objects.

[0062] In the specific calculation, the system first calculates the confidence level of each piece of evidence. The confidence level is not solely determined by the source type, but rather by a comprehensive consideration of source reliability, the degree of matching between the evidence and capability dimensions, timeliness, and whether the evidence relates to a point of conflict. Preferably, for any piece of evidence e, its confidence level can be expressed as:

[0063] ;

[0064] in, The confidence level of evidence unit e is represented, and its value range is preferably limited to [0,1]. , , and For the preset weights, satisfy ;For example In the formula, This indicates the reliability of the source, and its value is directly mapped from the source type field in the evidence unit. For example, a preset value can be used when the source type is a resume, another preset value can be used when the source type is a question and answer, and another preset value can be used when the source type is an interview. This mapping relationship is stored in the source reliability configuration table. This indicates the degree of semantic matching between the evidence unit and the current capability dimension. Its value can be calculated from the similarity between the evidence fragment text vector and the capability dimension template vector. It is preferable to use the normalized similarity result so that its value is in the range of [0,1]. The time freshness is indicated by a value calculated based on the difference between the current assessment point in time and the time elapsed corresponding to the evidence; preferably, it can be expressed as:

[0065] ;

[0066] in, This indicates the time difference between the current assessment date and the end date of the experience corresponding to the evidence. This indicates the window with a preset validity period; both can be obtained directly from the date field and the configuration field. This indicates a conflict indicator; it is set to 1 if the evidence unit has been referenced by any conflict item in the conflict set, and 0 otherwise. This process automatically lowers the confidence level of evidence units from the same source that are involved in a conflict, while increasing the confidence level of evidence units from more reliable sources, better matching the capability dimension, and more recent in time.

[0067] After obtaining the confidence level of each evidence unit, the system aggregates evidence units from resumes, Q&As, and interviews according to the ability dimension, obtaining the matching degree of each source under that ability dimension. Let the set of evidence units corresponding to the d-th ability dimension under the resume source be denoted as . The set of evidence units corresponding to the question and answer source is as follows: The set of evidence units corresponding to the source of the interview is as follows: The matching degree of the three types of sources can be preferably expressed as:

[0068] ;

[0069] ;

[0070] ;

[0071] in, , and These represent the matching degree of the resume, Q&A, and interview under the ability dimension d, respectively. , and These represent the number of evidence units in the three sets of evidence units mentioned above; The system assigns the confidence level of the evidence to the corresponding evidence unit. Through this aggregation method, the system can transform the evidence units that have been solidified in step two into computable matching degrees from different sources under the same capability dimension, and the source and composition of each matching degree can be traced back to the specific evidence unit.

[0072] Building upon this, the system further calculates the capability dimension score by incorporating conflict sets and missing key evidence. Preferably, for any capability dimension d, its dimension score can be expressed as:

[0073] ;

[0074] in, This represents the dimensional score of capability dimension d; , and These represent the source weights of resumes, Q&A sessions, and interviews for this ability dimension, respectively, satisfying the following conditions: Its value can be directly obtained based on the evidence type constraints or scoring rule configuration table in the job profile. This represents a conflict penalty term, used to characterize the degree to which identified conflicts under the current capability dimension reduce the score; preferably, let the set of conflict terms associated with capability dimension d be... The severity of the j-th conflict term is Conflict score The number of conflicting items is The maximum severity level is Then it can be expressed as:

[0075] ;

[0076] in, and All can be read directly from the conflict set. This represents a penalty for missing key evidence, used to characterize whether the necessary evidence for the current competency dimension requirements of the job profile has been met; preferably, let the expected number of key evidence items for competency dimension d in the job profile be... The number of valid evidence units that have actually been formed so far is Then it can be expressed as:

[0077] ;

[0078] in, Evidence type constraints and essential condition configurations derived from job profiles This score is calculated by counting the number of evidence units with a confidence level not lower than a preset threshold under the current capability dimension. After this processing, if a capability dimension has a small amount of evidence but it is insufficient to support a complete judgment for that dimension, the system will not simply award a high score, but will correct it through a penalty for missing key evidence. For those calculated using the formula... The system can further truncate it to the [0,100] interval.

[0079] After calculating the scores for each competency dimension, the system further generates a comprehensive score based on the dimension weights in the job profile. Preferably, let the set of competency dimensions in the job profile be D, and the weight corresponding to the d-th competency dimension be... The overall score can then be expressed as:

[0080] ;

[0081] in, This indicates the candidate's overall score; Let d be the weight of the d-th capability dimension in the job profile, and the sum of the weights of all dimensions is 1; This is the risk penalty coefficient; This represents a risk label penalty item, used to reflect the impact of integrity risk, experience authenticity risk, or other preset risk labels on the overall score. Preferably, let F be the set of risk labels generated in the current assessment results, and let the level value of the k-th risk label be... The number of risk labels is The upper limit of the risk level is Then it can be expressed as:

[0082] ;

[0083] The types and levels of risk labels are both triggered by the scoring rules and conflict sets and written into the risk label results.

[0084] After the dimensional scores and overall scores are calculated, an evidence chain object is generated based on the dimensional score set, candidate profile set, and conflict set. Each capability dimension's score node is associated with the evidence units and related conflict items involved in the calculation under that dimension score, forming a ternary binding relationship between dimension score, evidence units, and conflict items. Ideally, at least two evidence units should be associated with each dimension score node. If the associated evidence is insufficient, a penalty for missing key evidence is triggered under that dimension, and it is marked as a low score due to lack of evidence in the subsequent evaluation report. If there are conflict items under that dimension and the penalty for conflict is significantly increased, it is marked as a low score due to conflict in the subsequent evaluation report. The generated evidence chain object can display the key evidence and the basis for score reduction for each dimension in the subsequent evaluation report, allowing auditors to directly locate the specific evidence units and conflict items involved in this scoring during review, without having to search through the original resume, Q&A, and full transcript of the interview.

[0085] After the evidence chain object is generated, the system outputs a candidate evaluation report based on the job profile, dimensional score set, overall score, and evidence chain object. The evaluation report includes at least the score for each capability dimension, key evidence, conflict explanations, and review recommendations. When a dimension is penalized due to a conflict item in the conflict set, the report simultaneously displays the conflict explanation corresponding to the score reduction. When a dimension is penalized due to missing key evidence, the report simultaneously displays a message indicating insufficient evidence. Simultaneously, the system serializes the dimensional score set, overall score, risk tags, evidence citation list, conflict item list, and report summary generated in this evaluation transaction, generating an output snapshot and calculating the output snapshot summary value. This output snapshot summary value, along with the input snapshot summary value generated in step two, the version anchoring record generated in step one, and the current evaluation tracking number, is written into the audit trace object.

[0086] Step Four: After completing the assessment report output in Step Three and writing the input snapshot summary value, output snapshot summary value, and version anchoring record into the audit trail object, the system has the basic conditions to replay and verify the recruitment assessment. When subsequent candidate disputes, internal spot checks, audit reviews, or updates to job profiles, scoring rules, or model services require comparison with historical conclusions, the system first receives a recalculation request. Based on the assessment tracking number carried in the recalculation request, it retrieves the corresponding job profile version number, competency dimension model version number, scoring rule version number, input snapshot summary value, output snapshot summary value, and the output result summary already generated in Step Three from the audit trail object. The key point here is not simply reading a historical report text, but restoring the calculation boundaries involved in the assessment at that time, that is, restoring which version of the job profile, competency dimension model, and scoring rule was used, and what the input snapshots involved in the calculation were. In this way, the system restores the conditions under which historical conclusions were formed before the recalculation begins, thus avoiding the direct substitution of the latest job configurations, scoring rules, or candidate information in the current database for historical states, which would render the so-called recalculation merely a re-evaluation. By restoring the historical version number set corresponding to the evaluation tracking number, the input snapshot summary value, and the output snapshot summary value, the system can reconstruct the computational boundaries at the time when historical conclusions were formed before the recalculation begins, thereby preventing the current state from being mistakenly used as a historical state in the recalculation.

[0087] After restoring the aforementioned historical calculation conditions, the system prioritizes performing a recalculation based on the same version. Specifically, without invoking the latest current configuration, the system reloads the corresponding job profile, competency dimension model, and scoring rules according to the version number recorded in the audit trace object. Simultaneously, it retrieves the input snapshot corresponding to the assessment tracking number, ensuring that the evidence confidence calculation, dimension score calculation, comprehensive score calculation, and evidence chain binding processes in step three are re-executed under the same input and version boundaries as the historical assessment. Through this recalculation, the system obtains a new set of recalculation output results and serializes these output results to form a recalculation output snapshot summary value, denoted as... Simultaneously, the system reads the historical output snapshot summary value from the audit trail object and records it as... .when When the system determines that the historical evaluation result can be obtained again under the original calculation conditions, it indicates that the credible historical version baseline formed in step one is consistent with the input and output boundaries formed in steps two and three; when If the system determines that the historical evaluation result cannot be fully reproduced under the original conditions, it further triggers a difference comparison process to check whether the difference stems from inconsistent input snapshots, incomplete version recovery, or changes in the result items during the calculation process. The purpose of this approach is to transform the reproducibility of historical results into a directly verifiable summary consistency issue, rather than relying on subjective judgment based on manual report reading.

[0088] If the system needs to further analyze the impact of rule upgrades, job profile updates, or model service replacements, then differential recalculation will be performed after completing the same-version recalculation. Differential recalculation still uses the input snapshot already saved in step two as its input basis, but instead of using historical version configurations, it loads the currently effective job profile version, competency dimension model version, scoring rule version, and related service parameters, and re-executes the evaluation calculation process in step three on the same input snapshot. This way, the current output result retains only version changes as a variable compared to the aforementioned historical output results, while fixing the input content. This allows for a clearer observation of whether the difference in results is caused by version changes, rather than by changes in candidate information. To uniformly quantify this difference, the system calculates a comprehensive recalculation offset attribution based on dimension score offset, evidence citation drift, and conflict identification offset. Preferably, the comprehensive recalculation offset attribution can be expressed as:

[0089] ;

[0090] in, This represents the total amount of attribution for recalculated offset. , and This represents a preset weight that satisfies: ;For example , , .

[0091] In the formula, This represents the dimensional score offset, used to characterize the overall difference between historical and current output results in each capability dimension score; This represents the evidence citation drift, used to characterize the degree of substitution between the two outputs on the evidence citation set; This represents the conflict identification offset, which characterizes the degree of change in the conflict item identification result and conflict severity between the two outputs.

[0092] Among them, the dimension score offset The score can be directly calculated from the scores of the capability dimensions in the historical output and the current output. Let the set of capability dimensions be K, and the score of the k-th capability dimension in the historical results be... The score in the current results is The weight of this ability dimension in the job profile is... The upper limit of a single dimension score is Therefore, the preferred option is:

[0093] ;

[0094] in, Read directly from the dimension weight vector in the job profile. The set of dimensional scores from historical output snapshots. The set of dimensional scores from the current difference recalculation results. The upper limit of 100 for the dimension scores in step three can be adopted. This parameter reflects the overall drift of the scores of each dimension behind the change in the total score. When some key dimensions change significantly due to rule updates or model updates, It will increase accordingly.

[0095] Evidence citation drift This is used to characterize the degree of change in the evidence citation list between historical and current output results. Since this invention emphasizes that the scoring result must be bound to the chain of evidence, even if the overall score changes little, a significant shift should be considered as a change in the set of evidence supporting that score. Let the set of evidence citation identifiers in the historical results be... The set of evidence citation identifiers in the current results is Therefore, the preferred option is:

[0096] ;

[0097] in, and The elements in the table are all evidence unit numbers or evidence item numbers, derived from historical evidence chain objects and evidence chain objects generated by the current differential recalculation. If the evidence units referenced in the two results are completely identical, then... Set to 0; if the evidence cited in the two results is completely different, then Approaching 1.

[0098] Conflict identification offset This is used to characterize the degree of change between historical and current output results in the conflict set. Since the multi-source consistency verification result directly participates in the calculation of the scoring penalty item, and changes in the conflict set synchronously affect the score correction and result interpretation of the affected capability dimensions, the conflict identification offset can reflect the degree of difference between historical and current output results in conflict handling logic. To simultaneously reflect whether there are changes in conflict items and whether the conflict severity has changed, it is preferable to use... Represented as:

[0099] ;

[0100] in, and For local weights, satisfying ,For example , In the formula, Let the set of conflicting items in the historical results be denoted as . The set of conflicting items in the current result is ,but:

[0101] ;

[0102] Among them, the set intersection and union operations use a combination of conflict type and association capability dimension as the matching key. Let represent the change in conflict severity, and let be the number of conflict items that can be matched between historical and current results. The severity of the i-th matching conflict in the historical results is The severity in the current results is The maximum severity level is Then it can be expressed as:

[0103] ;

[0104] in, and The severity field is directly derived from the historical conflict set and the current conflict set. This represents the upper limit for conflict severity; in this example, it can be set to 3. After this processing, if the two results not only identify different conflicting items but also make different judgments on the severity of the same conflicting item, then... This will be adjusted accordingly, thus making the biased attribution more closely aligned with the actual scoring correction logic.

[0105] After obtaining the composite value of the recalculated offset attribution Then, the system can determine whether the deviation between historical results and current results warrants further attribution analysis. If... If the result is below the preset difference threshold, it indicates that although the current system version has changed, the deviation between the recalculated result under the current input snapshot and the historical result is small. In this case, the system can directly record the current difference recalculation result as having insignificant offset. If the deviation exceeds the preset difference threshold, it indicates that the version change has had a substantial impact on the evaluation conclusion, and the system continues to perform hierarchical attribution of the offset source; specifically, the system can keep the input snapshot unchanged and adopt a local difference recalculation method that only replaces a single version source to test the version changes of job profile, capability dimension model, scoring rules, model service, and prompt parameter separately. Let the set of version sources to be compared be J. For the j-th type of version source, the system calculates the local offset value corresponding to its individual replacement. Furthermore, take:

[0106] ;

[0107] in, This indicates the version source category that maximizes the local offset value. The system uses this to determine the version number. The class version change was identified as the primary source of the result offset.

[0108] After completing both the same-version recalculation and differential recalculation, the system will also generate a recalculation audit result for this recalculation process. The recalculation audit result includes at least: the assessment tracking number corresponding to the recalculation request, the set of historical version numbers used, the same-version recalculation output snapshot summary value, the differential recalculation output snapshot summary value, the recalculation offset attribution aggregate, the dominant offset source, and necessary review prompts. If the same-version recalculation is successful and If the system marks the historical assessment report as reproducible under the original conditions in the recalculation audit results, then the system will indicate in the differential recalculation results that the historical assessment report can be reproduced under the original conditions; if the differential recalculation shows... If the deviation exceeds the preset difference threshold and there is a clear dominant source of deviation, the system will simultaneously write the dominant source and the corresponding deviation description into the recalculation audit results, so that recruitment specialists, interviewers, or auditors can directly view it during subsequent review. After this processing, the output of step four is not just a conclusion on whether it is consistent, but a set of structured recalculation results that simultaneously includes historical reproducibility judgments and explanations of the reasons for deviations.

[0109] Example 2: The design of the intelligent recruitment analysis system based on the chain of evidence of the present invention is based on the method in Example 1, specifically as follows... Figure 4 As shown, it includes a version anchoring module, an input sealing module, an evaluation output module, and a recalculation audit module.

[0110] The version anchoring module receives the job requirement text, job metadata, and multi-source input data of candidates corresponding to the current recruitment task. It establishes a one-time evaluation transaction for the received job requirement text and job metadata, and generates a unique evaluation tracking number for this evaluation transaction. It also parses the job requirement text to obtain a set of job elements, generates a job profile based on the job template configuration, and simultaneously extracts the version number of the job profile corresponding to this evaluation transaction. Furthermore, it reads the version numbers of the capability dimension model, scoring rules, model service, and prompt parameters associated with this evaluation transaction, and associates these version numbers with the evaluation tracking number to form a version anchoring record corresponding to this evaluation transaction. Finally, it calculates the version anchoring reliability comprehensive value based on the version time synchronization deviation, version dependency closed-loop completeness, and frozen registration completeness, and compares the version anchoring reliability comprehensive value with a preset threshold. If the comparison result meets the requirements, it determines that the current evaluation transaction has formed a reliable historical version baseline, allowing it to enter the subsequent input sealing and evaluation processing flow.

[0111] The input sealing module is used to uniformly access and seal the multi-source input data of the candidates after a credible historical version baseline has been formed for the current evaluation transaction. It is also used to uniformly classify resume files, initial screening Q&A records, and interview audio / video data into the evaluation transaction corresponding to the current evaluation tracking number, and generate structured resume results, Q&A results, and interview results respectively. Furthermore, it is used to combine the job profile frozen in step one to aggregate and extract evidence from the structured resume results, Q&A results, and interview results according to ability dimensions, forming a multi-source evidence set of candidates corresponding to the ability dimensions, and further solidifying the records mapped into the ability dimensions into evidence units. It is also used to uniformly serialize all structured inputs, evidence unit indexes, source location information, and multi-source mapping relationships corresponding to this evaluation transaction, generating input snapshots and input snapshot summary values. Finally, it is used to generate an initial evidence chain graph corresponding to this evaluation transaction based on the relationships between evidence units, to characterize the support relationships, conflict relationships, and missing relationships between different evidence units.

[0112] The evaluation output module, under the constraint of version anchoring records, calls the job profile version, competency dimension model version, scoring rule version, model service version, and prompt parameter version bound to the current evaluation tracking number. It performs dimension-by-dimensional calculations on the evidence units already fixed in the input snapshot and outputs a set of dimension scores, a comprehensive score, and risk labels. It also calculates the confidence level of each evidence unit and aggregates evidence units from resumes, Q&As, and interviews by competency dimension to obtain the matching degree of each source under that competency dimension. Furthermore, it calculates competency dimension scores by combining conflict sets and key evidence missing cases, and generates a comprehensive score based on the dimension weights in the job profile. The system is also used to generate evidence chain objects based on the dimensional score set, candidate profile set, and conflict set, and to establish associations between the score nodes of each capability dimension and the evidence units and related conflict items involved in the calculation under the dimensional score; it is also used to output candidate evaluation reports based on job profiles, dimensional score sets, comprehensive scores, and evidence chain objects, and to serialize the dimensional score set, comprehensive score, risk label, evidence citation list, conflict item list, and report summary formed in this evaluation transaction in a unified manner, generating output snapshots and output snapshot summary values, and writing the output snapshot summary values ​​together with the input snapshot summary values, version anchoring records, and the current evaluation tracking number into the audit trace object.

[0113] The recalculation audit module is used after step three, when the assessment report is output and the input snapshot summary value, output snapshot summary value, and version anchoring record are written to the audit trace object, to receive a recalculation request. Based on the assessment tracking number carried in the recalculation request, it retrieves the corresponding job profile version number, competency dimension model version number, scoring rule version number, input snapshot summary value, output snapshot summary value, and output result summary from the audit trace object to restore the calculation boundaries at the time the historical conclusion was formed. It is also used to prioritize performing a recalculation of the same version after restoring the historical calculation conditions to determine whether the historical assessment result can be obtained again under the original calculation conditions. Furthermore, it is used to continue performing differential recalculation when it is necessary to analyze the impact of version changes, and based on... The calculation of the dimensional score offset, evidence citation drift, and conflict identification offset yields a comprehensive attribution measure for recalculating the offset, determining whether the result offset reaches a level requiring further attribution analysis. It is also used to perform stratified attribution of changes in job profile versions, capability dimension model versions, scoring rule versions, model service versions, and prompt parameter versions when the offset reaches a preset differential threshold, and to determine the dominant source of the current result offset. Furthermore, it is used to generate a recalculation audit result, which includes at least the assessment tracking number corresponding to the recalculation request, the set of historical version numbers used, the snapshot summary value of the same version recalculation output, the snapshot summary value of the differential recalculation output, the comprehensive attribution measure for recalculation offset, the dominant offset source, and necessary review prompts.

[0114] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0115] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0116] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An intelligent recruitment analysis method based on evidence chains, characterized in that, Includes the following steps: Receive job requirement text, job metadata, and multi-source input data from candidates; establish assessment transactions and generate assessment tracking numbers. The job requirement text is parsed to extract job responsibilities, qualifications and ability requirements. A job profile is generated by combining job metadata. The evaluation version object identifier corresponding to the current evaluation task is read and associated with the evaluation tracking number to form a version anchoring record. The version anchoring credibility comprehensive quantity is calculated based on the version time synchronization deviation, the version dependency closed loop completeness, and the frozen registration completeness. The credibility historical version baseline is determined when the preset threshold value is met. Under the trusted historical version baseline, structured processing is performed on multi-source input data to form structured input results. The structured input results are mapped and evidence is extracted based on the capability dimensions in the job profile. The mapped records are solidified into evidence units, and the structured input results, evidence unit indexes, and mapping relationships are serialized to generate an input snapshot. Under the constraints of version anchoring records, the evaluation version object bound to the evaluation tracking number is invoked. The confidence level of the evidence unit in the input snapshot is determined based on the reliability of the source, the degree of matching between the evidence and the capability dimension, the freshness of time, and the conflict status. The dimension score set and the comprehensive score are aggregated by capability dimension, and then the evidence chain object and the candidate evaluation result are generated. The input snapshot summary value, the output snapshot summary value, the version anchoring record, and the evaluation tracking number are written into the audit trace object. Based on the recalculation request, retrieve the corresponding version anchoring record, input snapshot, input snapshot summary value, and output snapshot summary value, perform same-version recalculation, and perform differential recalculation when it is necessary to analyze the impact of version changes; calculate the recalculation offset attribution composite amount based on the dimension score offset, evidence citation drift, and conflict identification offset, and generate recalculation audit results.

2. The intelligent recruitment analysis method based on evidence chain according to claim 1, characterized in that: The job metadata includes at least the job name, job level, department, and city; the assessment version object identifier includes at least the job profile version number, competency dimension model version number, scoring rule version number, model service version number, and prompt parameter version number; the version anchoring record includes at least the assessment tracking number and the assessment version object identifier corresponding to that assessment tracking number.

3. The intelligent recruitment analysis method based on evidence chain according to claim 2, characterized in that: The version time synchronization deviation is used to characterize the degree of dispersion of the object identifier of each evaluation version at the freezing time; the version dependency closed loop completeness is used to characterize whether the dependency relationship between job profile, capability dimension model, scoring rules and their scoring items, and evidence type is completely closed; the freeze registration completeness is used to characterize whether the key objects related to the current evaluation task have been registered before entering the formal evaluation. When the version anchored trusted synthesis quantity does not reach the preset threshold, the current evaluation transaction will not enter the subsequent evaluation calculation process.

4. The intelligent recruitment analysis method based on evidence chain according to claim 1, characterized in that: The multi-source input data includes at least resume files, initial screening Q&A records, and interview audio and video data; when the multi-source input data is processed in a structured manner, a resume structured result, a Q&A structured result, and an interview structured result are generated.

5. The intelligent recruitment analysis method based on evidence chain according to claim 4, characterized in that: When mapping and extracting evidence from structured input results based on the competency dimensions in the job profile, the mapped records are solidified into evidence units; each evidence unit records at least the source type, source location, time information, associated competency dimensions, and evidence fragment summary value; the input snapshot includes at least the structured input results, evidence unit index, and the mapping relationship between the evidence unit and the original source.

6. The intelligent recruitment analysis method based on evidence chain according to claim 1, characterized in that: The reliability of the source is mapped by the source type of the evidence unit, which includes at least resumes, Q&A, and interviews; the conflict state is used to characterize the degree of contradiction between different evidence units under the same ability dimension; the key evidence missing situation is used to characterize whether the target ability dimension required by the job profile lacks evidence units that meet the scoring requirements.

7. The intelligent recruitment analysis method based on evidence chain according to claim 6, characterized in that: The evidence chain object is used to associate the score nodes of the capability dimension with the evidence units involved in the score calculation and related conflict items; the candidate evaluation result includes at least the dimension score set, the comprehensive score, the risk label, and the evidence citation list; the output snapshot is used to archive the dimension score set, the comprehensive score, the risk label, and the evidence chain object association result of the actual output of this evaluation.

8. The intelligent recruitment analysis method based on evidence chain according to claim 1, characterized in that: The same version recalculation refers to reloading the same evaluation version object according to the evaluation version object identifier in the version anchoring record, and calling the input snapshot corresponding to the evaluation tracking number to re-execute the evaluation calculation under the same input boundary and version boundary as the historical evaluation, so as to determine whether the historical evaluation results can be obtained again under the original calculation conditions.

9. The intelligent recruitment analysis method based on evidence chain according to claim 8, characterized in that: The differential recalculation refers to loading the currently effective evaluation version object and re-executing the evaluation calculation while keeping the input snapshot unchanged, in order to obtain the current output result; when the total amount of recalculated offset attribution is higher than the preset differential threshold, attribution analysis is performed on the changes in job profile version, capability dimension model version, scoring rule version, model service version, and prompt parameter version, and the dominant offset source is determined.

10. An intelligent recruitment analysis system based on a chain of evidence, characterized in that: The analysis system is used to implement the method according to any one of claims 1-9, and includes the following modules: The version anchoring module is used to receive job requirement text, job metadata and candidate multi-source input data, establish evaluation transactions and generate evaluation tracking numbers; The job requirement text is parsed to extract job responsibilities, qualifications and ability requirements. A job profile is generated by combining job metadata. The evaluation version object identifier corresponding to the current evaluation task is read and associated with the evaluation tracking number to form a version anchoring record. The version anchoring credibility comprehensive quantity is calculated based on the version time synchronization deviation, the version dependency closed loop completeness, and the frozen registration completeness. The credibility historical version baseline is determined when the preset threshold value is met. The input sealing module is used to perform structured processing on multi-source input data under the trusted historical version baseline to form structured input results, map and extract evidence based on the capability dimensions in the job profile, solidify the mapped records into evidence units, and serialize the structured input results, evidence unit index and mapping relationship to generate an input snapshot; The evaluation output module, under the constraint of version anchoring record, calls the evaluation version object bound to the evaluation tracking number, determines the confidence level of the evidence units in the input snapshot based on the reliability of the source, the degree of matching between the evidence and the capability dimension, the freshness of time, and the conflict status, and aggregates them by capability dimension to generate a set of dimension scores and a comprehensive score, thereby generating an evidence chain object and candidate evaluation results, and writes the input snapshot summary value, output snapshot summary value, version anchoring record, and evaluation tracking number into the audit trail object; The recalculation audit module is used to retrieve the corresponding version anchoring record, input snapshot, input snapshot summary value and output snapshot summary value according to the recalculation request, perform same-version recalculation and differential recalculation, and calculate the recalculation offset attribution comprehensive quantity based on the dimension score offset, evidence citation drift and conflict identification offset, and generate recalculation audit results.

Images