A Case Business Review and Analysis Method and System Based on Multi-Agent Collaboration
By standardizing formats and establishing review relationship chains, and collaboratively processing case materials, the inconsistency issues in the conclusion positioning and review process of multiple agents were resolved, and stable review result generation was achieved.
Patent Information
- Application Number
- CN202611511168.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-28
- Publication Date
- 2026-11-17
AI Technical Summary
When processing conclusions from multiple agents, the existing case review platform has difficulty accurately identifying the connection between the same facts, and the review conclusions are prone to discrepancies after the materials are supplemented or the rules are updated, resulting in inconsistent review processes.
By acquiring case business materials and their location, formation time, and review rule versions, standardizing the format and dividing the content into segments, forming standard review data, establishing review relationship chains, retaining adjacent content segments and collaboratively sending review tasks, cross-checking different conclusions, and forming the final review result.
It improves the consistency of multi-agent conclusions and the stability of the review process, ensuring that the review results remain consistent in the common processing context and reducing the differences in conclusions caused by reprocessing materials.
Smart Images

Figure CN123115874A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of case business review and analysis technology, and in particular to a case business review and analysis method and system based on multi-agent collaboration. Background Technology
[0002] As case materials are gradually digitized, case review and analysis platforms need to process case files, tables, scanned images, transcripts, and audio / video transcribed text simultaneously. Existing platforms first perform text recognition or field extraction according to material type, and then send different material fragments to the corresponding review models or modules based on procedures, documents, evidence, and cross-reviews. Each review model or module independently reads the assigned materials and generates problem conclusions, material location information, and credibility information according to its loaded review rules. After receiving multiple conclusions, the platform uses credibility ranking, fixed weight merging, or manual selection to form a summary result.
[0003] This processing method can complete different review tasks in parallel. However, when dividing review materials, it is sometimes difficult to fully preserve the connection between the same fact in different documents. Material fragments located at the task boundary may only be sent to one of the review models. After material supplementation, document revision, or review rule update, different review models may also form conclusions based on different material versions or different review rule versions. When multiple review models give different conclusions on the same fact, the aggregation end usually obtains the final conclusion and its credibility. It is inconvenient to reconstruct the material fragments and review rules that each conclusion went through along the material processing order. The material location or review rule where the conflict first occurred is not easy to determine directly. If all materials are resent for review, the materials and rules that were originally consistent will be processed again, and the new processing context may cause differences in the review conclusions before and after. Therefore, the existing solution still has room for improvement in the accurate positioning of different conclusions of multiple agents and the review based on stable contexts. Summary of the Invention
[0004] This invention provides a case business review and analysis method and system based on multi-agent collaboration, which improves the consistency of different conclusions and review processes of multiple agents.
[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a case business review and analysis method based on multi-agent collaboration, the method comprising: Step S101: Obtain the case business materials and their corresponding material locations, formation times, and review rule versions; standardize the material format and divide the content into sections; write the material locations, formation times, and review rule versions into the corresponding content sections to obtain standard review data. Step S102: Based on the standard review data, arrange each content segment according to the case entity and the time of its formation, mark the supporting or conflicting relationships between the content segments, and connect the review rules with the corresponding content segments to obtain the review relationship chain; Step S103: According to the review relationship chain, divide adjacent content segments according to the review items, retain the content segments at the connection points of different review items in the adjacent division results, generate collaborative review tasks according to the order of each division result, send the collaborative review tasks to the corresponding review agent, and obtain a candidate conclusion chain carrying the material location and review rule version. Step S104: Based on the candidate conclusion chain, compare the content segments and review rules traversed by the conflicting candidate conclusions, determine the position where different content segments or different review rules first appear, retain the same chain segments before the position, and cross-send the different chain segments after the position to the corresponding review agent for re-review to obtain the review conclusion chain. Step S105: According to the review conclusion chain, check whether each review conclusion is connected to the same chain segment. If the chain segments are connected and the conclusion content is consistent, merge the review conclusions. If the chain segments are not connected or the conclusion content is inconsistent, retain the position and the different chain segments thereafter to obtain the case business review result.
[0006] Secondly, a case business review and analysis system based on multi-agent collaboration includes a data processing node, a relationship chain processing node, a task coordination node, multiple review agent nodes, a review processing node, and a result storage node. The data processing node is connected to the relationship chain processing node. The data processing node processes the case business materials and their corresponding material locations, formation times and review rule versions into standard review data. The relationship chain processing node is connected to the task coordination node. The relationship chain processing node reviews the content segments in the data according to the case entity and the formation time, marks the supporting or conflicting relationships, and connects the review rules with the corresponding content segments to form a review relationship chain. The task coordination node connects with multiple review agent nodes. The task coordination node divides the review relationship chain according to the review items, retains the common content segments of adjacent division results, sends collaborative review tasks, and receives candidate conclusion chains returned by multiple review agent nodes. The review processing node is connected to multiple review agent nodes and result storage nodes. The review processing node determines the position where conflicting candidate conclusion chains first appear with different content segments or different review rules, retains the same chain segments before the position, cross-sends the different chain segments after the position, and forms the case business review result based on the returned review conclusion chain.
[0007] The above-described solution of the present invention has at least the following beneficial effects: This invention processes case business materials and their corresponding material locations, formation times, and review rule versions into standard review data. This facilitates the preservation of material sources and version foundations during subsequent review processes. Review relationship chains are formed based on the standard review data, and content segments connecting different review items are simultaneously retained in adjacent partitioning results. This allows different review agents to share factual content located at task boundaries. Candidate conclusions returned by review agents are connected into candidate conclusion chains according to their traversed content segments and review rules, ensuring a traceable processing order in the conclusion formation process. When candidate conclusions differ, the location of the first occurrence of different content segments or review rules is determined from the candidate conclusion chain. The same chain segments before that location are retained, and only the different chain segments after that location are cross-sent to the corresponding review agent for re-review, ensuring that the review process maintains consistent material content and review rules. Case business review results are formed based on the connection between the review conclusion chain and the same chain segments, as well as the consistency of the review conclusions. This helps to locate different conclusions to specific material locations or review rules and complete the review within a relatively stable common processing context, thereby improving the consistency of multi-agent conclusion location and review processes. Attached Figure Description
[0008] Figure 1 This is a schematic diagram of the overall process of the case business review and analysis method provided in this embodiment.
[0009] Figure 2 This is a schematic diagram illustrating the shared retention relationship between the content segments of the audit items provided in this embodiment.
[0010] Figure 3 This is a schematic diagram of the process for cross-checking and verifying the results of the differential chain segments provided in this embodiment.
[0011] Figure 4 This is a simulated line graph showing the recalculation of the number of entries under different path lengths in the synthesis task of this embodiment.
[0012] Figure 5 This is a simulation heatmap showing the net reduction ratio of recalculated entries under different common chain segment ratios and additional boundary treatment ratios in this embodiment.
[0013] Figure 6 This is a simulation scatter plot showing the injected divergence positions and the divergence positions obtained by comparison in the synthesis task of this embodiment.
[0014] Figure 7 This is a simulation bar chart showing the number of branch triggers under the preset boundary conditions in this embodiment.
[0015] Figure 8 This is a simulation frequency graph showing the distribution of unilateral differential chain segment lengths under a fixed path length in this embodiment. Detailed Implementation
[0016] To make the purpose, technical solution, and implementation process of this application clearer, the following is in conjunction with... Figures 1 to 8 This paper describes a case business review and analysis method based on multi-agent collaboration. This embodiment uses fact-checking among document records, attachment records, and audio / video records in case business materials as an application scenario. The method is executed by a processor deployed in a case review and analysis platform. Review agents undertake the review tasks assigned to them, while the processor handles material organization, task sending, chain segment comparison, and result merging. In this embodiment, "first" and "second" are used to distinguish the two objects participating in cross-review. The occurrence time of the events described in the materials is stored separately, and the two times are not interchangeable. A chain segment represents a continuous part of an ordered reference path, and a connection represents the data relationship stored through explicit predecessor and successor references. Case business materials are a collection of electronic records formed for the same case business object, including readable text documents, scanned images, tables, and records with sound or image sequences. Material location is the positioning information that allows retrieval to a local area of the original document. Standard review data is a collection of data records obtained through format conversion, content segmentation, and source information encapsulation. A content segment is a basic text unit in this collection, and each content segment stores its own text and its source location.
[0017] Audit rules are executable conditions used to check specific fact categories, record integrity, or relationships between records. An audit rule version is an identifier for a snapshot of the rule content that the processor consistently uses in a single audit. Audit items are used to organize rules and tasks; in this embodiment, they are configured as four categories: program record audit, document field audit, attachment fact audit, and cross-consistency audit. An audit relationship chain is an ordered expression of relationships formed by content segments in standard audit data under case entity and formation time constraints. It stores supporting relationships, conflicting relationships, and connections between audit rules and content segments. It can be stored using directed relationship records plus an ordered position table, without forcing all relationships into a linear structure with only one successor. An audit agent is a software execution instance with the ability to receive tasks, read content segments, execute audit rules, and return structured results. In this embodiment, audit agents are set up for each of the four audit items. Each audit agent can call the same fixed version of the artificial intelligence language model but uses different audit item instructions and the range of readable materials. A candidate conclusion is an unmerged judgment returned by the audit agent for a factual issue. Its record includes the conclusion category, case entity, fact category, fact status identifier, fact value, and ordered reference path. The candidate conclusion chain is a chain of candidate conclusion records and their reference paths connected in the order of tasks. The review conclusion chain is a chain of conclusions and their reference paths formed after the common inputs are fixed and reviewed again.
[0018] The platform's storage holds the original material list, content segment list, review rule list, relationship record list, task list, and conclusion list. Material identifiers, review batch identifiers, and segment identifiers are all generated through registration and counting records. A segment identifier is composed of the material identifier within the case, the material revision number, and the order within the segment. When content text is modified, a new material revision number and corresponding segment identifier are generated. Before starting the review, the processor reads the current material list and the list of published review rules. The review batch is fixed only when all necessary materials can be opened and the selected rule list is complete. The review rule list is formed after being entered by the rule configuration personnel. The processor checks the data type, the completeness of required inputs, and the return status configuration before changing the rule status from draft to publishable.
[0019] Step S101: Formation of standard audit data, specifically including: The system acquires case-related business materials, their locations, creation times, and review rule versions. It standardizes the material format and divides them into content segments. The material location, creation time, and review rule version are written into the corresponding content segments to obtain standard review data. When the processor receives case-related business materials, it first reads the material registration information, including the case identifier, material source type, file storage location, creation time field, and revision description. It searches the original material table according to the case identifier to confirm whether the currently received material is a new revision of the same material. If the original material identifier is explicitly specified and the revision description points to an existing record, a material revision number is added. If no original material identifier is specified, it is registered as a new material. For text materials where characters can be directly read, the paragraph text is obtained according to the paragraph order saved in the document. The internal paragraph order and character start and end positions are also saved. If page numbers can be obtained from the document, they are used as part of the material location. If the source file does not have stable page numbers, the paragraph order and character range in the original file are used for location. When reading characters, the character encoding declared in the file is used first. For files declared as having a general character encoding, characters are decoded one by one according to the corresponding encoding standard. Byte ranges that cannot be decoded retain their original byte positions and are marked as unrecognizable. If there is no encoding declaration and the encoding cannot be determined through format parsing, multiple encodings are not rotated and any readable result is selected. The material is then sent to the format confirmation queue, and conversion continues after the registered encoding is determined.
[0020] In one character decoding implementation, text declared to use variable-length general character encoding is scanned byte by byte from front to back. The number of bytes occupied by the current character is determined based on the preamble of the starting byte. It is checked whether the required subsequent bytes are complete and conform to the preamble of consecutive bytes. If the check passes, the encoding flag bits in each byte are removed, and the remaining valid bits are concatenated in the original order to obtain the character encoding value. The value is then checked to see if it falls within the allowed character range and whether the minimum encoding specified for that character is used. If it passes, the character is output and its original byte range is recorded. If it fails, the error range and invalid state are retained. For text stored in fixed-width encoding units, the byte order is determined according to the file marker. Adjacent bytes are combined into encoding units in this order. Ordinary encoding units directly correspond to characters. High-order surrogate encoding units must be paired with subsequent low-order surrogate encoding units to obtain a complete character. If a pair is missing, the order is reversed, or only half an encoding unit remains at the end position, the decoding is marked as invalid. The decoding method is saved along with the file declaration. After obtaining the text characters, a copy of the original text is retained, and a standard text is generated for rule comparison. The standard text is scanned from left to right, converting full-width numbers to their corresponding half-width numbers, full-width English letters to their corresponding half-width English letters, and full-width spaces to ordinary spaces. Consecutive ordinary spaces are compressed into a single space in non-table fields. Paragraph line breaks and field separators remain unchanged. Each replacement records the correspondence between the position of the standard character and the original character range.
[0021] The full-width character conversion queries a fixed set of corresponding character records character by character. These records use the source character code as the query key and the target character code as the return value. During creation, full-width numbers, uppercase letters, and lowercase letters are arranged in their respective consecutive order, and paired with their corresponding half-width characters in the same order. Records of full-width spaces corresponding to ordinary spaces are then added separately. Each source character is checked to have only one target; any missing characters are output as is. When compressing consecutive spaces, a single retained standard space is associated with the entire original space range. For Chinese numbers, easily confused characters, negation words, and punctuation, semantic replacement is not performed during the unified material formatting stage. Value conversion is only performed when the review rules explicitly require numerical types and units can be identified. Since differences between unsubmitted and submitted text can affect the factual state, retaining the original form of these characters avoids changing the meaning of the material before formatting. Any manual revisions create a new revision record instead of directly rewriting the old content. For image materials, images of each page are obtained according to the original document's page order. The top left corner of the image is used as the origin of the coordinate system, with the horizontal direction to the right and the vertical direction downwards as the vertical direction of the coordinate system. Both horizontal and vertical coordinates are in pixels. The text recognition component receives the page image and page identifier, and returns the recognized text, the text region rectangle, and the character order within the region. A reference from the recognized text to the original image is established using the page number and the text region rectangle. The image region rectangle stores the left, top, right, and bottom boundaries. The text recognition component can use a deployed text detection and recognition model. The component first outputs candidate text line regions in the image, and then outputs the characters and their positions within each region from left to right or according to the text direction specified in the document. The processor reads the region returned by the model and does not treat the entire page of recognized text as a coordinate-free string. If the recognition component can only return line-level coordinates, the material position retains the line-level region and marks the positioning granularity as line-level. It is not permissible to claim character-level positioning without character coordinates.
[0022] In a text recognition implementation applicable to scanned pages where the text line direction has been confirmed, the text recognition component first obtains the color channel values of each pixel, sums the channel values of the same pixel and divides them by the number of channels to obtain grayscale values. The component counts the number of pixels corresponding to each grayscale value, enumerates the boundary grayscale values sequentially from low grayscale to high grayscale, accumulates the number of pixels on both sides and the sum of grayscale values at each boundary, divides the sum of grayscale values on each side by the number of pixels on that side to obtain the average grayscale on both sides, takes the square of the difference between the average grayscale values on both sides and multiplies it by the product of the pixel ratios on both sides to obtain the discrimination value of the boundary. The boundary with the largest discrimination value is selected as the foreground segmentation threshold. If the number of pixels on either side is empty, the boundary is not evaluated. If there are no valid boundaries on the entire page, a blank page or an image pending confirmation state is returned. After obtaining the foreground segmentation threshold, the component compares the grayscale value with the threshold pixel by pixel. Pixels darker than or equal to the threshold are written as text foreground, and the rest are written as background. The number of text foregrounds is accumulated line by line to obtain the horizontal row statistics. The component forms candidate line bands from consecutive non-empty lines, and then accumulates the number of foregrounds column by column within the candidate line bands to remove blank edges. For candidate line bands with cross-column connections, vertical markers, or regular lines, the original area is preserved and marked as needing area confirmation. The simplified line path is not used to forcibly process pages that do not meet the conditions. By limiting the applicable input, the text area can be obtained by explicit pixel calculation. For available candidate line bands, the component obtains the pixel height of the line band and the input height required by the recognition model. It compares the target height with the pixel height to obtain the scaling ratio, and multiplies the ratio by the pixel width to obtain the target width. When generating the target image, it first calculates the continuous coordinates of the target pixel center in the source line band based on the horizontal and vertical positions of the target pixel center relative to the target image. It reads the adjacent source pixels around the coordinates, calculates the weighted grayscale in the two lines according to the horizontal distance, and then weights the two grayscale values according to the vertical distance to obtain the target pixel grayscale. Only pixels within the boundary are used for the edges, and the edge values are repeated. When the source width and height are zero, the recognition is exited.
[0023] The above horizontal interpolation first takes the column with the continuous horizontal coordinates rounded down as the left column and the next column as the right column. The contribution ratio of the right column is obtained by subtracting the left column coordinates from the continuous horizontal coordinates. The contribution ratio of the left column is obtained by subtracting the right column contribution ratio by the unit value. The gray values of the left and right pixels are multiplied by the corresponding ratios and then added to obtain the interpolated gray values of the same source row. The contribution ratios of the upper and lower rows are obtained in the same way in the vertical direction. The interpolated gray values of the two rows are multiplied by the corresponding ratios and then added to obtain the target gray value. The continuous source coordinates are obtained by multiplying the target pixel center by the ratio of the source size to the target size and then returning to the source pixel center reference. The target width is the width after scaling, rounded up and containing at least one pixel. The actual position is reflected and the aspect ratio after rounding is saved, instead of reusing the ideal aspect ratio before rounding. When the recognition model extracts features from the line band image, it reads the input pixels and channels covered by the convolution kernel at each output position. It multiplies the input value with the weights of the same kernel position in the parameter file, sums the multiplications, and adds the bias of the output channel to obtain the channel response at that position. Negative responses are set to zero, and non-negative responses are retained. Input positions that exceed the image boundary are padded with zeros according to the model configuration. The kernel is moved sequentially along the line band width to obtain the feature map. Then, the features in the same column are averaged along the height direction to form a column feature sequence. The convolution kernel size, stride, and number of channels are all read from a fixed recognition parameter file. If the input dimension does not match the parameter file, the model refuses to execute without truncating the weights.
[0024] The component multiplies each position in the column feature sequence by the character output weight channel by channel and accumulates the results. Adding the character bias, it obtains the output score for each allowed character and blank category. The maximum score for the current position is subtracted from each score, and the natural index is taken. Each index value is divided by the sum of all index values to obtain the character output probability. The component selects the category with the highest probability. If probabilities are equal, the component selects the category with the higher character table registration order. Consecutive identical categories are merged first, and then blank categories are deleted to obtain the recognized text. Each character corresponds to the row band range covered by its selected column position. If the character position cannot be further subdivided, it is saved according to the aforementioned row-level position. The recognition parameter file is obtained after training with labeled text line samples using the same image preprocessing method. The file simultaneously saves the character table, input height, convolutional layer configuration, and output layer weights. During access, the character table order is confirmed to correspond one-to-one with the output channels using text lines that can be verified against the original text. This embodiment can directly load existing parameter files that have passed the check. If no usable parameter file is provided, or if the image type exceeds the confirmed input conditions, recognition is returned pending confirmation, and unrecognized text is not created based on the case context. When the image needs to be scaled before being input into the text recognition component, the processor saves the original image width and height as well as the recognition image width and height. First, it divides the recognition horizontal coordinate by the recognition image width to obtain the horizontal position ratio, and then multiplies the ratio by the original image width to obtain the original horizontal coordinate. The vertical coordinate is operated on in the same way as the recognition image height and the original image height. The left and top boundaries of the region are rounded to the smaller integer, and the right and bottom boundaries of the region are rounded to the larger integer.
[0025] For the rectangle obtained through coordinate transformation, check whether the left boundary is smaller than the right boundary, whether the top boundary is smaller than the bottom boundary, and whether the four boundaries fall within the original image range. If the rectangle exceeds the image by one pixel due to outward rounding, the boundary is brought back to the image edge and the correction is recorded. If the rectangle exceeds one pixel or forms a zero-width and zero-height area, the marked position is invalid and it is not added to the valid content segment that can be cited by candidate conclusions. This ensures that the cited material can be actually displayed during the review. The basic implementation of this embodiment does not arbitrarily rotate the original page image before recognition. When the original page needs orientation correction, a derived page with confirmed orientation is established through material revision, and the whole-page correspondence between the derived page and the original page is saved. For areas that cannot be stably mapped back to the original page, the processor saves both the derived page position and the original whole-page position, and marks the positioning granularity as whole-page pending confirmation, avoiding hiding undisclosed geometric transformations in the material position field. For tables, the processor reads cell text according to the page, table order, row order, and column order. It combines field names and values in the same row into independently readable text, and saves column headers as context fields with clear sources. When merged cells cover multiple rows, each referenced row retains the same merged cell position and does not register copies of inherited headers as new factual evidence. Therefore, different auditing agents can read the complete meaning of fields without counting the same original cell as multiple documents.
[0026] For audio materials, the transcription component receives the original audio file and material identifier, and returns continuous text, the start timecode, and the end timecode for each text segment. The timecode is expressed in milliseconds from the start of file playback. The processor checks that the start timecode is not earlier than zero and the end timecode is not earlier than the start timecode, while also checking that the end timecode does not exceed the file duration. Segments that do not meet the conditions are marked as invalid, and the original audio file is retained for re-transcription. By defining the time reference within the file, the review task can play back the same acoustic segment. In one executable audio input configuration, the transcription component receives mono decoded audio stored at the sampling frequency specified in the parameter file. For multi-channel input, the channel numbers are saved before conversion and transcribed separately, without mixing independent speech by direct addition. If the original audio sampling frequency does not conform to the transcription component interface, it is converted according to the following resampling process, and the input sampling frequency and output sampling frequency are saved. The target sampling frequency used is consistent with the input during transcription parameter training. The timecode always corresponds to the original file playback time, and changes in the number of sampling points must not change the original material positioning reference. Before resampling, the original sampling time and channels are saved in the decoded audio. For uncompressed audio, each signed sample value is read in the order of sampling bit width and byte specified in the file header, and then divided by the maximum amplitude value corresponding to that bit width to obtain the normalized amplitude. For compressed audio, the decoder of the corresponding file format first restores it to the same sequence of sample values. The processor checks whether the duration, number of channels and sampling frequency of the decoding report are compatible with the material registration. If decoding cannot be completed, the original file is retained and returned for transcription pending confirmation.
[0027] One executable resampling method is interpolation with a low-pass kernel. The component reads the original sample window covering each target sampling time, subtracts each original time in the window from the target time to obtain each time offset. The cutoff frequency is half of the smaller of the input and output sampling frequencies. Each cutoff frequency is multiplied by twice the time offset to obtain a dimensionless offset. When the offset is non-zero, the sine of the product of pi and the offset is calculated and then divided by the same product. When the offset is zero, a unit weight is used, multiplied by the symmetric finite window weight given in the parameter file to obtain the sampling kernel weight. The window width and window weight are saved together with the transcription model input configuration. Based on the sampling kernel weights, the component multiplies each original sample amplitude by its weight and sums them to obtain the total weighted amplitude. This total is then divided by the sum of the kernel weights to obtain the target sampling amplitude. Outside the file boundaries, the nearest valid sample is used for extension. When the sum of the kernel weights is zero, the target sampling is marked invalid and the current segment transcription is terminated. The weight window gradually moves forward with the target time. When the output sampling frequency is lower than the input sampling frequency, the cutoff frequency is used to suppress components exceeding the output's expressible frequency band, ensuring that the resampled data still corresponds to the original playback time. The resampled sequence number cannot directly replace the original timecode. Audio transcription divides the processing window according to the longest input duration allowed by the parameter file. Adjacent windows retain overlapping intervals. The overlap duration is determined based on the model's time perception range and the interface's allowed context length, and is less than the processing window duration, allowing unfinished words and phrases at the window edge to continue recognition in the next window. The processor compares segments with overlapping start and end timecodes and identical text within the overlapping area. Only one text is retained, and the source records of both windows are saved. If the text in the overlapping area is different, both are retained as pending transcription, without eliminating potential conflicts by arbitrarily selecting a transcription path.
[0028] In one implementation of the transcription model, the component sequentially extracts short-duration amplitude frames within the current audio window according to the frame length and frame shift configured by the model. For each frame, the average amplitude is first calculated, and then the average value is subtracted from each amplitude to obtain the DC-free amplitude. These are then multiplied by the analysis window value saved in the parameter file to form the windowed amplitude. If the end of the frame is less than the specified length, zeros are added and the actual sampling length is saved. The zero-padded area does not generate a separate character time code. Subsequently, the windowed amplitude is multiplied by the cosine and sine values of each sampling position at the frequency grid and accumulated to obtain the real and imaginary responses. The two responses are squared and then added to obtain the power value of the grid point. The component multiplies the power value of each frequency grid point by a fixed frequency band weight and then sums them to obtain the energy of each frequency band. The frequency band weight is formed by the low-end, center, and high-end frequencies of each frequency band in the parameter file. The weight increases from the low end to the center according to the relative distance, and decreases from the center to the high end according to the relative distance. The weight outside the frequency band is zero. The energy of each frequency band is compared with the minimum positive energy protection value. If it is lower than the protection value, the protection value is taken. Then, the natural logarithm is taken to obtain the logarithmic frequency band characteristics. The protection value, frequency band boundaries, and analysis window are all saved with the transcribed parameter version. If the configuration is missing or the upper and lower bounds of the frequency band are invalid, the program will refuse to run to avoid zero energy causing input that cannot be calculated. The transcription model reads logarithmic frequency band features sequentially over time, covers adjacent feature frames with temporal convolution kernels, multiplies each feature by its corresponding weight and accumulates them, adds a bias, and then retains the non-negative response to form the acoustic features at each output time. The output layer performs character score calculation and exponential normalization on the acoustic features, with the calculation order being the same as the aforementioned character output layer. Allowed categories are the transcription character table and blank categories. The highest probability category is selected in chronological order, consecutive identical categories are merged, and blank categories are deleted to obtain the transcribed text. Windows that fail to form valid characters return an indeterminate text state, without interpreting the indeterminate text as a lack of relevant facts.
[0029] The character timecode is obtained from the range of consecutive output times that generated the character. The component first multiplies the output time by the model's cumulative time step to convert it to the input feature frame range, and then uses the frame shift and the original audio window start time to convert it to the original file time interval. When convolution uses adjacent frames, the earliest and latest valid input frames covered by the kernel are included in the interval simultaneously. The interval start point is truncated to the file start point, and the interval end point is truncated to the file end point. Two character intervals are allowed to overlap but cannot be reversed. Deduplication in the overlapping window still needs to satisfy both time coincidence and identical text to maintain the correspondence between acoustic input and material localization. The transcription model parameters come from acoustic samples with original text annotations and time correspondences. The same frame length, frame shift, frequency band weights, and character table are used for both training and input. The parameter file explicitly registers the time step and padding method of each layer of temporal convolution. This embodiment uses the obtained valid parameter file to perform transcription. If the parameter source version is unreadable or the acoustic feature dimensions are different, a model mismatch will be returned. Therefore, simply providing the transcription service name cannot be regarded as completing the standardization of audio materials. For video materials, the processor obtains the audio track transcribed text and the on-screen text recognition text respectively. The audio track uses the original file timecode, and the on-screen text area saves the video frame timecode and the pixel rectangle within the frame. In this embodiment, a frame can be read per second for a segment containing text images. The same text in consecutive frames is not presumed to be different facts. The processor only saves the duration interval when the text in the same area is completely consistent and the time is continuous. When the text changes, the previous interval ends and a new content segment is established.
[0030] When acquiring video frames, the processor first reads the frame display timestamp and time reference from the container. For each sampling moment, it selects the first decodeable frame whose display time is no earlier than that moment. The decoder returns the complete image and the actual frame time. The processor saves the actual frame time instead of the planned sampling moment. When multiple planned sampling moments point to the same frame, only one frame record is established. If the timestamp is backdated, missing, or the frame decoding fails, the corresponding time interval is retained in the pending confirmation state. The audio track still uses its original time reference and is aligned with the video timeline through the starting offset provided by the container. When a continuous interval needs to be established for the text area of adjacent frames, the processor first checks whether the frame size and coordinate reference are the same, and then compares the region rectangle and the standard text. Only when the coordinates of all sides of the rectangle are consistent and the text is consistent will the existing interval be extended in the basic implementation. An optional implementation that allows for small positional errors first takes the larger left boundary, larger top boundary, smaller right boundary, and smaller bottom boundary of the two rectangles to obtain the overlapping rectangle. If either the overlapping width or height is not positive, the overlapping area is zero. Otherwise, multiply the width by the height to obtain the overlapping area. Add the two original areas together and subtract the overlapping area to obtain the union area. Then divide the overlapping area by the union area to obtain the positional overlap ratio. Only when the text is consistent and this ratio reaches the lower limit of the region matching configured in the model can the interval be extended, and if it equals the lower limit, it is included in the match. If the union area is zero, the match is rejected.
[0031] The lower limit of regional matching can be calibrated using confirmed frame pairs from the same and different regions. The processor calculates the position overlap ratio for each frame pair according to the aforementioned process, takes the largest ratio among frame pairs from different regions as the upper limit of error connection, and then selects the smallest ratio strictly higher than this upper limit among frame pairs from the same region as the lower limit of matching. If no such ratio exists within the same region, the approximate matching method is stopped and the process reverts to rectangular precise comparison. When the sample is empty, no approximate matching configuration is published. The calibration record and the independent verification sample are saved separately according to the video material identifier. Video frame sampling is only used to obtain visible text candidates and cannot be used to prove that there is no other screen information within the sampling interval. When the review rules require verification of events shorter than the sampling interval, the processor rereads the corresponding segments frame by frame according to the time range carried by the rule. The core review process in this embodiment uses the already obtained content segments and their material positions, and does not rely on inferring unobserved events from low-frequency sampling. After acquiring the text, the processor divides the content into segments according to paragraph boundaries, image region boundaries, or continuous semantic boundaries. For text materials, the original paragraph boundaries are used first; for image materials, the already acquired text region boundaries are used first; and for audio-visual transcribed text, the end of complete sentences and the end of spoken segments are used first. The processor subtracts the task instructions, rules, and output reserved length from the model's allowed input length to obtain the available content length. If an original paragraph exceeds this length, the processor searches for the nearest sentence-end punctuation mark from the allowed end position to divide the text. If no sentence-end punctuation mark is found, the text is divided at the allowed end position, and the continuity markers of the preceding and following content segments are saved. If the available content length is not positive, the generation task is stopped and the configuration is required to be corrected to avoid truncating the content first and then discarding the positioning basis.
[0032] When the allowed end position is obtained, the candidate text is counted using the current model's word segmentation configuration. The available content length is compared in units of words. Text is added sentence by sentence and re-segmented. If the limit is exceeded after adding the next sentence, the original character boundary of the sentence is gradually shortened and re-counted. The last complete character boundary that has not exceeded the limit is selected. The number of available words cannot be directly used as the number of characters. If even a complete character cannot be accommodated, the input capacity is insufficient. The text used for word segmentation counting is consistent with the actual text sent. In this embodiment, continuous semantic boundaries are determined using an executable rule-based method. The processor reads characters from the beginning of the current segment. When a sentence-ending punctuation mark is encountered and followed by a new speech identifier or a new field name, a boundary is formed. Sentence-ending punctuation marks inside a pair of parentheses do not form a boundary alone. The processor saves the nested parenthesis count. It increases once for each left parenthesis and decreases once when a right parenthesis is encountered and the count is non-zero. When a mismatched right parenthesis is encountered, the character is retained and the format is marked for confirmation. The new field name is obtained from the field name table corresponding to the current material type. The processor compares all characters in the candidate field names sequentially from the current character position. Only when the names are completely identical and followed by the specified field separator is it considered a hit. The new speech identifier is obtained from the speech interval boundary given by the transcription component or the confirmed speech mark. If there is no such information, only the end of the sentence and the length boundary are used. The processor does not guess the speaker change. Unclosed parentheses ensure that the remaining text is still controlled by the length limit, while the unclosed parentheses state is saved.
[0033] After a content segment is retrieved from the converted text, the processor traverses the standardized character positions covered by the segment, reads the original position correspondence of each character, and merges consecutive character intervals from the same original region into a single position interval. If a content segment spans multiple image regions or multiple timecode segments, the material position is stored as an ordered list of position intervals, without replacing the entire content segment's position set with the first position, by maintaining a one-to-many position correspondence. The formation time is read by the processor from the formation time field explicitly recorded in the material, followed by the formation time confirmed by the material registration, without automatically replacing it with the file copy time or import time. When time zone information is available, the time zone offset is read first, and then the offset is subtracted from the local time to obtain a unified time. When only the date is available without hours, minutes, and seconds, the date granularity and the corresponding interval for that day are stored. A fictitious precise second value cannot be added. When sources conflict, the formation time is marked as pending confirmation. When parsing the generated time, the processor reads integers in the order of year, month, day, hour, minute, and second fields registered in the source format. It checks the validity of the month and the current date according to the calendar, and then checks whether the hours, minutes, and seconds are within the allowed range. If there is date granularity, only the provided fields are checked. For a complete time, the processor first accumulates the number of complete days from the base date to the target date, multiplies it by the number of seconds per day, and then adds the duration corresponding to the hour, minute, and second of the current day in sequence. Finally, it subtracts the explicitly given time zone offset in seconds to obtain a unified time. Unknown time zones are not defaulted to the local time zone but are reserved for confirmation.
[0034] For content segments with missing time periods, the processor retains the missing time status and simultaneously saves the material registration order. During subsequent sorting, it places these segments after the known time content segments of the corresponding case entity, and then arranges them according to the material registration order and the order within the segment. This arrangement is only used to form a reproducible calculation order and does not indicate that the missing time material is actually later than other materials. When determining the same factual state, the chronological order cannot be established based on this calculation position. The review rule version is taken from a fixed review rule snapshot of the current review batch when the content segment is encapsulated. Each rule identifier in the snapshot corresponds to an unwriteable rule content. The content segment saves the snapshot version and a list of available rules for reference. Different rules can have different rule identifiers. When the rule content changes, a new snapshot version must be released; even if only the applicable conditions are changed, the old version will not be used. The processor writes the content segment text, paragraph identifier, material location, creation time and its validity status, audit rule version, material revision number, and character position correspondence into the same content segment record. Before writing, it checks if a record with the same name already exists by using the composite primary key. If all fields of a record are the same, the original record is returned. If the paragraph identifier is the same but the text or material location is different, the record is rejected and an identifier conflict is output. Only when all valid records are written can standard audit data be formed. Atomic writing prevents the task from reading a half-finished product where the text has been updated but the material location has not. The standard audit data also saves a record list and an exception list for the current audit batch. The record list lists the content segments according to the material registration order and the order within the segment. The exception list lists records that cannot be read, cannot be located, or whose time is pending confirmation.
[0035] Step S102 verifies the establishment of the relationship chain, specifically including: The processor reads valid content segments from the standard review data, obtaining the entity identifier, fact category, and fact content segment by segment. The entity identifier is used to distinguish the business objects being reviewed in the case. In this embodiment, the business number of the material itself is used as the priority entity identifier. If there is no business number, the confirmed name mapping table is used for querying. The name mapping table is established by the processor after receiving the confirmed object names and alias records from the case registration. Each record stores the case identifier, original name, standard entity identifier, effective material revision number, and confirmation status. The processor checks each record by case identifier and original name. Records that are completely identical and point to the same entity are merged. Records pointing to different entities are retained as conflicting records and are not published. If the query does not find a match, an unconfirmed entity record is created. If the query finds multiple matches, the object number is required to be supplemented. If a single valid match is found, the content segment is associated with the standard entity identifier. The fact category includes at least the record existence status, record formation time, number of objects, attachment corresponding status, and processing status. Each type of fact has a corresponding data type and compatibility judgment method. The processor first reads the fact categories involved in each rule from the review rule table, then finds the explicit field name for the current content segment, extracts tables and forms that can directly read fields by field name, and calls the language model used by the review agent to return candidate facts and original text ranges for narrative text.
[0036] The input to the AI language model includes the standard text of the current content segment, paragraph identifiers, a list of allowed fact categories for output, field type descriptions, and fixed format return requirements. The output includes candidate entity names, fact categories, candidate fact values, the start and end character positions of the original text, and an indeterminate flag. The processor verifies that the start and end character positions are within the current content segment, retrieves the original text for that interval, and checks whether the candidate entity names and values can be directly derived from the original text. Candidates that cannot be supported by the original text are not included in the fact record and are retained for extraction pending confirmation. The model version, word segmentation configuration version, task instruction version, and output field constraints are fixed at the start of the review batch, and the version of the actual report from the model interface is saved. The model consists of a lexical embedding layer, a causal attention calculation layer connected in the configured layer order, a feedforward transformation layer, and a lexical output layer. A lexical is a text segment and its number in the vocabulary. The parameter file saves the vocabulary, word segmentation merging order, dimensions of each layer, projection weights, and the allowed input length of the model. When the processor loads, it compares the output dimension of the previous layer with the input dimension of the next layer layer, and compares the vocabulary size with the number of output layer channels. If any of them do not match, the model is rejected from participating in the current review batch.
[0037] When forming the model input, the text is connected in a fixed order: audit item instructions, allowed fact categories, field definitions, content segment list, and return constraints. Each content segment is preceded by an independent paragraph identifier boundary and followed by an end boundary. After the word segmenter obtains the original text byte fragments, it repeatedly checks adjacent fragments according to the merging priority order in the parameter file, selecting the highest priority allowed merging pair to merge into a single token, until no allowed merging pair exists. Each token is then associated with its covered standard character range. Unrecognized bytes are backed up using the lexicon's byte backing rules while retaining their positions. When the input length exceeds the model's allowed range, the length of the usable material is obtained by deducting the audit item instructions, rule text, and the reserved return length. The material is then distributed to continuous processing windows along the defined content segment boundaries. Each window repeatedly carries the same rule and clearly identified boundary content segments. If a single content segment still exceeds the usable length, it continues to be divided into sub-intervals along its existing original text intervals, retaining the parent paragraph identifier and interval position. When the audit agent returns, it uses the parent paragraph identifier and the actual interval as a reference.
[0038] The lexical embedding layer reads the corresponding parameter row by lexical number to obtain the feature components of that lexical. It also reads the positional parameters based on the lexical's position in the current input. The components of the same dimension from two parameter rows are added together to obtain the positional lexical features. If the position exceeds the allowed range in the parameter file, an input length error is returned; if a parameter row is missing, a model configuration error is returned. Each causal attention calculation layer calculates the average of each dimension of the lexical features at the current position. The average is subtracted dimension-wise to obtain the centered features. The squares of each centered feature are averaged to obtain the variance. A positive stable variable from the parameter file is added to the variance, and the square root is taken to obtain the normalized divisor. The centered features are divided by the normalized divisor dimension-wise, multiplied by the scaling parameter of that dimension, and an offset parameter is added to obtain the normalized features. The current model call is stopped when the stable variable is non-positive or a non-finite number is encountered during calculation. The normalized features are input into the query projection, key projection, and value projection respectively. For each output dimension of each projection, the features of each input dimension are multiplied sequentially by the weights from that input dimension to the current output dimension. All products are accumulated and a bias is added to obtain the corresponding projection components. When configuring multi-head attention in the parameter file, it checks whether the total dimension of the query projection and the total dimension of the key projection can be equally divided by the number of heads. It obtains the corresponding components according to the continuous dimension range of each head in the parameter file. Each head multiplies and sums its own query component and its own key component. The attention scaling divisor uses the square root of the dimension actually involved in the calculation for the current head. Value projections are read in the same head order and concatenated after weighting. The parameter file is rejected if the number of heads is zero, the dimensions cannot be equally divided, or the key-value head correspondence is undefined. When calculating the attention weight at the current position, it enumerates the current and previously visible word positions, multiplies the current position's query projection by the same dimension component of each visible position's key projection, and sums the results to obtain the matching sum. This sum is divided by the square root of the projection dimensions involved in the summation to obtain the scaled matching value. The largest value is subtracted from each scaled matching value, and the natural exponent is taken for the difference. The sum of these exponents is then used to obtain a non-negative attention weight with a unit sum. Future positions are not included in the exponent summation. The call is stopped when the projection dimension is zero or the exponent summation is invalid. After obtaining the attention weights, for each value projection dimension, the value projection components of each visible position are multiplied by the corresponding attention weights and accumulated to obtain the weighted component of the current position. When multiple attention heads are configured, the weighted components of each head are connected in the order specified in the parameter file. The connection result is accumulated by multiplying the weights dimension by dimension through the output projection, and then added to the components of the same dimension as the original input of the layer to obtain the attention output. The aforementioned normalization is performed, and intermediate responses are obtained through the first feedforward projection. Negative intermediate responses are set to zero, and non-negative intermediate responses are retained. The original dimension is restored through the second feedforward projection and added to the attention output to form the output of the layer. This process is repeated in the hierarchical order until the last layer.
[0039] The lexical output layer reads the current position features from the last layer, multiplies each candidate lexical by its output weight dimension by the weights, accumulates the scores, and adds the corresponding biases to obtain candidate scores. It then calculates the probabilities of candidate lexicals by subtracting the maximum value, taking the natural exponent, and dividing by the sum of exponents. Within the allowed output range, it selects the lexical with the highest probability. If probabilities are equal, it selects lexicals by their lexicon numbers. The selected lexical is appended to the already output sequence, and the next position is calculated again. The process stops when a terminator is encountered and the field is complete. If the output length limit is reached but the result is still incomplete, an incomplete result is returned. The allowed output range is determined by the syntax of the returned field. The processor stores the current parsing state (field name, separator, field value, or record end). Only predefined field names are allowed in the field name position; only enumerated field values are allowed in the enumerated field value position; and only paragraph identifiers and valid interval representations provided by the current task are allowed in the original text citation position. If no allowed lexical is encountered, the current response is terminated and the original task is resent. After the model returns a candidate, the processor reads the complete characters from its original text range, retrieves the rules for the fields corresponding to the fact category, and performs extraction again. The text name uses the confirmed name table, the quantity uses numerical parsing, and the negation status uses the negation words registered in the rules and their applicable fields. If the model candidate and the value obtained again are the same, it is saved as a valid fact. If there are multiple corresponding values, the object of the negation word is unclear, or it can only be guessed from the context, the fact is returned pending confirmation.
[0040] For optional implementations requiring domain adaptation, training samples consist of confirmed content segments and their fact field annotations. Each sample retains the original text range and material location. Annotators fill in the entity, fact category, fact value, and whether it is determinable. Samples with inconsistent annotations are re-verified against the original text to form a unified label; samples that still cannot be unified are labeled as undeterminable. Training and validation samples are separated according to case identifiers. The processor concatenates the input text with the target structured output to form training samples. When predicting each target word, it provides the correct target word preceding it, uses the same causal attention layer to obtain the final layer features of the current position, fixes the parameters of the word embedding layer, attention layer, and feedforward layer, and only updates the weights and biases of the word output layer. The processor obtains the negative natural logarithm of the correct word probability for each valid target position, sums them, and divides by the number of valid target positions to obtain the average training loss. If the number of valid positions is empty, the current sample is skipped and the reason is saved. When calculating the output layer gradient, the predicted probability is read for each valid target position and each candidate word. If the current candidate word equals the correct target word, a unit value is subtracted from the probability; otherwise, the probability is retained to obtain the output error. The output error is divided by the total number of valid target positions in the current batch to obtain the average error. This average error is multiplied by the components of each dimension of the final layer feature at the current position to obtain the gradient contribution of the corresponding output weight. The contributions of all valid positions to the same weight are summed to obtain the weight gradient. The average error is directly accumulated by word to obtain the bias gradient. When updating parameters, the current output weight, the corresponding gradient, and the learning step size in the training configuration are read. The learning step size is multiplied by the gradient to obtain the weight adjustment. This adjustment is subtracted from the original weight to obtain the new weight. The bias is updated in the same way. Each batch uses the same output layer parameters before the update to complete all gradient accumulation and replaces them with the new parameters at once. If non-finite loss, gradient, or parameter issues occur, the batch update is canceled. The training configuration records the batch size, maximum number of traversals, learning step size, number of stops without continuous improvement, and upper limit of input length. The batch size is determined according to the number of final layer features that the device can store simultaneously. The processor runs candidate step sizes sequentially using the same initial output layer and the same training samples. First, it eliminates candidates with invalid parameters. Then, it prioritizes candidates with higher source location efficiency. If source location efficiency is the same, it selects the one with higher field complete consistency rate. If both are the same, it selects the one with lower validation loss. If still the same, it selects the one with the earlier configuration order. After obtaining the selected configuration, it checks it on the reserved samples that were not selected. When executing the training stop judgment, the no-improvement count is set to an initial empty count. After each round of complete traversal, the current round is compared with the current retained version in the priority order of source location efficiency, field complete consistency rate, and validation loss. If the current round is significantly better than the retained version, the retained version is replaced and the no-improvement count is cleared; otherwise, the no-improvement count is incremented. The training stops when the configured number of no-improvement counts or the maximum number of traversals is reached. The field complete consistency rate is calculated with the validation samples with defined labels as the denominator and the samples with all required fields consistent with the labels as the numerator.The effectiveness of source location is calculated by using the number of reference intervals returned by the model as the denominator and the number of intervals that can be found in the corresponding content segment and support the returned fields as the numerator. Empty reference responses do not contribute to the effective numerator. Samples that require a reference but do not return one are counted separately as incomplete fields.
[0041] When creating a fact record, the processor saves the original fact values, standardized fact values, fact categories, entity identifiers, fact status identifiers, original text intervals, and material locations. When the event number is clear, the event number determines the fact status; when only a time interval exists, the interval and corresponding object determine the fact status; when both are missing, the fact status is marked as unknown. When only a time interval exists without an event number, records are aggregated according to the case entity and fact category. The start, end, and endpoint inclusion flags of each interval are read. All identified endpoints are sorted by time, and duplicate endpoints are removed. A time slice is created for each pair of adjacent endpoints, and a separate time slice is created for each endpoint. Each slice is checked to see which original records actually cover it, forming the fact record set for that slice. Whether an interval endpoint covers another is determined by the original inclusion flag; records with undetermined endpoints remain pending confirmation. The case entity, fact category, and time slice boundaries together form the fact status identifier for that time slice. Record values are compared within the same slice. Different slices use different fact status identifiers. Original content segments can be linked to multiple slices through multiple factual associations while still retaining the same material location.
[0042] For numerical facts, the processor reads the numerical text and the immediately following unit text, and converts them according to the target unit corresponding to the fact category in the audit rules. Count types only allow non-negative integers and do not perform approximate rounding. Types with decimals, such as amounts or lengths, retain their original precision and use decimal arithmetic. If the unit is not given and cannot be uniquely inherited from the table header, it is marked as missing. Unit conversion records are registered by the rule configuration personnel according to the fact category, including the source unit, target unit, scaling factor, and whether there is an origin offset. The processor checks that the two units express the same physical quantity or count category, and searches for a unique record by source unit, target unit, and configuration version. Only when scaling is involved, the source value is multiplied by the scaling factor to obtain the target value. If there is an origin offset, the source offset adjustment is performed first according to the configuration, then multiplied by the scaling factor and the target offset is added. If a negative coefficient causes the range direction to be reversed, the upper and lower boundaries and the inclusion flag are swapped simultaneously. If there is no match or multiple matches, the unit is retained for confirmation. The parsing of decimal characters is stored separately for the integer part, decimal part, and sign. The digits are traversed from left to right, and existing accumulated values are left-shifted by decimal and added to the current digit to obtain the integer coefficient without a decimal point. The number of decimal places and the sign are recorded. When comparing two decimals, the integer coefficient of the one with fewer decimal places is padded with zeros according to the difference in the number of decimal places before comparison. When the quantity category requires integers, the decimal part is checked for emptiness or all zeros. If non-zero decimals are included, the type is marked as invalid. In the object quantity category, the audit rules require the use of integer numbers. The sign, digits, and decimal part are verified to conform to the parsing rules. The standard values are checked to ensure they are non-negative integers. If these conditions are not met, the type is marked as invalid. The generated range must be no less than the range including the lower boundary, no more than the range excluding the lower boundary, no more than the range including the upper boundary, and no more than the range excluding the upper boundary. Upper and lower boundaries can only be formed when the corresponding rules explicitly register the allowable deviation and its basis; otherwise, approximate values are retained for confirmation.
[0043] Within each entity identifier, content segments are arranged according to their formation time. When the formation time is a fixed moment, they are arranged chronologically. When the formation time is a date range, they are first arranged by the start point of the range while retaining the granularity description. If the start points are the same, they are arranged according to the material registration order, and then by the order within the segment. Different entity identifiers are organized according to the entity registration order. When a content segment involves multiple entities, multiple entity association records are saved, but they share the original segment identifier. The arrangement is achieved through stable merging. Each associated record to be arranged is regarded as an ordered small sequence. Each time, two adjacent small sequences are taken and their first records that have not yet been output are compared. The time validity status, the start point of the fixed time or date range, the material registration order, and the order within the segment are compared in turn. The record that should come first is written into the merge sequence and the reading position on that side is moved forward. When all the comparison fields are the same, the record that comes first in the original sequence is output first. When one side is empty, the remaining records on the other side are appended sequentially. The merging is repeated until there is only one sequence.
[0044] Supporting relationships indicate that two adjacent content segments under the same entity and fact category can jointly store the values given for the same fact state. Conflicting relationships indicate that these values cannot be jointly stored. When two segments describe different fact categories or different fact states, they are stored as ordered adjacent relationships but do not participate in this compatibility comparison. When the fact state is unknown, a relationship awaiting confirmation is stored. When determining whether two content segments can be written to the same fact record, the entity identifier, fact category, and fact state identifier are compared sequentially. If any determined field is different, the current compatibility comparison is exited. If there is an unknown field, it is returned to pending confirmation. All three fields simultaneously read the standard fact values and their data types. For enumerated states, they are compared word by word according to the list of allowed states. The same states can coexist, while mutually exclusive states cannot coexist. Composite states that are allowed to coexist by the rules are handled according to the compatibility items in the rule table. Unconfigured state combinations remain pending confirmation. When fact records read the same entity but come from different name table versions, it is first checked whether the two records corresponding to the names both belong to the confirmed version of the current review batch. Only when the same standard entity identifier is obtained through this version can the comparison continue. Records with the same name but not confirmed, or with the same number but pointing to different objects, retain identity conflict. The enumeration of state compatibility items is established based on the state definitions provided when the rule is published. Allowable state pairs for this fact category are enumerated sequentially. Identical states are marked as coexisting, mutually exclusive pairs explicitly defined by the rule are marked as incompatible, and other state pairs are marked as pending confirmation. Reverse state pairs are written with the same judgment. If the judgments for duplicate state pairs are inconsistent, publication is rejected. For two precise integer values, the integers are directly compared for equality; if they are the same, they can coexist; otherwise, they cannot. For a precise integer and a range value, the precise integer is compared for inclusion within the allowed range. Equality cases are handled based on the inclusion flags of the range endpoints: if within the allowed range, they can coexist; if outside the range, there is a conflict; if the inclusion flags of the endpoints are unclear, they remain pending confirmation. For two range values, the larger lower boundary is selected from the two lower boundaries, and the smaller upper boundary is selected from the two upper boundaries. If the lower boundary is smaller than the upper boundary, there are shared allowed values; if the lower boundary is larger than the upper boundary, there are no shared allowed values. If the two boundaries are equal, they can only be written together if both original ranges contain the value. If there is an unbounded range, only the given boundaries are compared. Values are marked as invalid if the format is incorrect or the upper and lower boundaries are reversed. For narrative facts, compare the extracted and verified entities, predicate categories, and state values. Do not determine whether the facts conflict by comparing whether two natural language passages are completely identical. If the narrative cannot be converted into a predefined fact category, retain the original text.
[0045] When two adjacent content segments can be written together, the support identifier, preceding segment identifier, following segment identifier, fact category, and current review rule version are written to the relationship record table. When they cannot be written together, a conflict identifier is written, and the material positions of the two content segments are simultaneously written to the conflict record. The conflict record uses the review batch, the fixed order of the two segment identifiers, and the fact category as the joint key. Repeated comparisons only return the original record. An ordered comparison sequence is established for related records of the same entity and the same fact category, and adjacent facts are compared in this sequence. When it is necessary to form a multi-source common fact, all valid records within the same fact state are read. The allowed value set of the first record is used as the current common set, and intersections are found with the allowed value sets of subsequent records in turn. Precise values are treated as single-point sets, and the range is intersected by the upper and lower boundaries and the inclusion flag. Enumerations are retained according to the same allowed state. If the intersection is empty, the current record that generated the empty intersection and the source used by the previous common set are recorded. If the intersection is unknown, it is retained pending confirmation.
[0046] Link the audit rules to the corresponding content segments, read the list of fact categories recorded in the audit rules, search for candidate content segments from the fact records by entity identifier and fact category, and verify the applicable status, material type, and time conditions of each rule one by one. If all conditions are met, write the reference relationship between the rule and the content segment; if the conditions are not met, do not link. If a fact is missing, create a missing input record and retain the rule's position. A rule in the audit rule table includes a rule identifier, the audit matter it belongs to, the audit rule version, the required fact category, the required number of inputs, comparison actions, and output categories. Comparison actions include existence checks, equality checks, time sequence checks, and value compatibility checks. The existence check reads the number of valid records for the required fact category. If the number is at least one, it returns the obtained input; if the number is zero, it returns that the input is missing. The equality check compares the standard values when both inputs are valid; otherwise, it returns that the basis is insufficient. The time sequence check reads the occurrence times of the events described by two facts. If both are definite moments, they are compared sequentially. If they are time intervals, the end of the first event interval is compared to the beginning of the second. If the end of the first event interval is earlier than or equal to the beginning of the second event interval, the "not later than" condition is met. If the intervals overlap and the order cannot be determined, the time relationship is returned for confirmation. If the audit rule requires strict earlier than, then the "equal" condition is not met. The rule record clearly saves the strict or non-strict comparison type. When an audit rule connects multiple content segments, the input is first organized according to the registration order of the input role, and then arranged according to the content position within the same input role to form an input reference group. Each input reference entry saves the paragraph identifier, rule identifier, audit rule version, input role, and content position. The position of a complete input reference group is the last content position corresponding to when all inputs are available. Rules in the same position are arranged according to the registration order. Rules with missing inputs save the missing entries.
[0047] When rule execution depends on the output of another rule, before publishing the rule version, check if the dependency forms a cycle. Select rules without incomplete predecessors, write them sequentially into the order table, and remove the corresponding predecessor relationship. Repeat this process until all rules are written. If any rules still cannot be written, reject the current configuration and enter automatic review. When executing specified dependencies, in each round, select the next rule from the rules whose predecessor results have all been obtained. The order of the round is determined by the position of the input reference group and the rule registration order. Rules that have not reached the executable state are kept pending execution. The dependent result is recorded as the rule result record with the source path and input into the successor rule. Save the original rule identifier, result status, and its input reference group. If the predecessor's return basis is insufficient, the successor rule retains this status. After the connection is completed, the review relationship chain saves the content segment record reference, fact record reference, supporting or conflicting relationships, review rule reference, and ordered position table.
[0048] Step S103, the formation of the collaborative review task and the candidate conclusion chain, specifically includes: Based on the review relationship chain, adjacent content segments are divided according to review items. Content segments connecting different review items are simultaneously retained in the adjacent segmentation results. Collaborative review tasks are generated according to the order of each segmentation result and sent to the corresponding review agent to obtain a candidate conclusion chain carrying the material location and review rule version. The review items to which the review rules belong are read sequentially from the beginning of the ordered position table of the review relationship chain. When consecutive rules belong to the same review item, they are retained in the same segmentation result. When a rule with a different review item is encountered for the first time, a segmentation position is established between the positions of the preceding and following rules. When a review item that has already appeared appears again, a new consecutive segmentation result is established without forcibly merging two non-adjacent intervals across intermediate items, so that the task order can still return to the position of the original relationship sequence. The segmentation result saves the starting rule position, ending rule position, review item, internal content segment list, and boundary content segment list. The internal content segment list is obtained from the content segments actually connected to the rule interval, and the boundary content segment list is obtained from the segmentation positions on both sides of the interval. Rules that are not connected to content segments but have missing input records are still retained in their respective intervals, and the missing input status is written into the task.
[0049] Between two adjacent partitioning results, the content segments connected by the nearest rule to the left and the nearest rule to the right of the partitioning position are read. These content segments are written to both the left and right partitioning results. When both sides contain the same content segment, each partitioning result retains only one content reference. When both sides contain multiple content segments, the actual connected content segment sets are retained according to their original content positions. No arbitrary content segment is used to replace multiple required inputs of the rules. This double-sided retention reduces the break in factual information caused by task boundaries. When forming the content segment list for each partitioning result, the processor starts from an empty list and reads the rule input reference groups according to the order of the rules within the interval. For each segment identifier, it checks if the same batch encapsulation identity already exists in the list. If not, it appends it; if it already exists, it adds the corresponding input role reference without copying the text again. The same process is then used to append the content segments at the two boundaries, saving their internal roles and boundary roles respectively. If a segment has both roles, both are retained. The output list position is determined by the first occurrence position. Figure 2 In the diagram, adjacent content segments A, B, C, and D represent consecutive content units in the original relation sequence. The internal scope of audit item one includes content segments A and B, and the internal scope of audit item two includes content segments C and D. When the division position is between content segments B and C, the processor makes both division results include content segments B and C. Copies at the two positions use the same paragraph identifier and the same material location. The connections in the diagram only indicate data retention relationships and do not indicate that copying the original document generates new evidence.
[0050] When copying content segments, the processor can either copy only the reference address of immutable records or copy identical fields to independent task records. When using reference addresses, the content segment version is also saved; when using independent task records, the original content segment is checked field by field before sending. Regardless of the storage method, the copied segment identifier, text, material location, creation time, and review rule version are consistent with those before copying. The task record uses its in-task location to distinguish whether it belongs to internal input or boundary input. When a content segment appears consecutively in three partitioning results, the processor saves the association between the content segment and each task, but only saves one authoritative copy in the content segment table. When calculating the material read volume later, it distinguishes between the number of task reads and the number of unique content segments. This distinction allows the additional sending volume caused by boundary sharing to be observed, while avoiding the assumption that multiple tasks referencing the same segment are independent support sources. The interpretation of conclusions is still based on the original material location. The task order is determined according to the starting rule position of each partition result. Supplementary tasks with the same starting position are ordered according to the registration order of the audit items. The task order increases from first to last within the current audit batch. The task table stores the audit batch identifier, task order, partition result, list of allowed rules, audit rule version, and the identifier of the receiving audit agent. The task order does not change with the order of network returns, and parallel execution will not make the arrangement of candidate conclusion chains non-repeatable.
[0051] The mapping relationship between auditing agents is established by the platform's task configuration. Each configuration records the audit item, auditing agent identifier, executable type, model version, and interface address. Before publishing the configuration, the processor sends a format check request containing only test fields to the interface to verify the returned fields and supported rule types. When multiple available instances exist for the same audit item, they are allocated from the fewest to the most current incomplete tasks. If the number is the same, they are allocated according to the instance registration order. If an instance is unavailable, the original task content is preserved and transferred to an instance with the same capability, instead of using a mismatched audit item executor. When counting the number of incomplete tasks, the processor uses the audit batch and task order as the task identity, reads the tasks that have been assigned to the current auditing agent and are in the state of waiting to receive, executing, or waiting to retry. Multiple attempt numbers for the same task are counted only once. Before allocation, the task count record of the auditing agent is locked. After comparing the counts of each available instance and selecting an instance, the task ownership is written and the count is increased. The count is deducted when a task is first converted to completed, withdrawn, or paused. Repeated returns do not deduct again. When the count is abnormal, it is re-accumulated from the task status, and negative counts are not allowed to participate in scheduling. Before sending a collaborative review task, the processor reads each content segment in the task list, checks its validity status and version to ensure consistency with the review batch, and then encapsulates the task order, content segment text, material location, fact record, support or conflict identifier, review rules, and review rule version into a single task message. The message also includes a conclusion field requiring a return and a list of paragraph identifiers that can be referenced. During task message serialization, the processor writes the review batch, task order, attempt number, content segment list, rule list, and return constraints in the order defined by the interface. The text is written using a defined character encoding, integers are represented as decimal numbers, decimals are represented as decimal strings with scale preserved, and unknown states use independent status fields instead of empty strings or zero values. Each list member is written with complete object boundaries, and delimiters in the text are escaped according to the interface escaping rules. After generation, the message is re-parsed and the original object is compared with the parsed object field by field. If the comparison is inconsistent, the message is not sent to avoid message format conversion changing the material text or location values.
[0052] During deserialization at the receiving end, the message format version and declaration length are checked. Fields are read according to object boundaries. If duplicate field names (which are not allowed to be repeated), values exceeding the field range, missing identity fields for list members, or unclosed objects are found, the entire message is rejected. Unknown but expandable fields are only saved as additional records and are not used to overwrite existing required fields. After successful parsing, the batch, paragraph identifier, and rule version verification are performed. Therefore, structure verification, source verification, and business verification are sequentially linked, and successful transmission cannot replace the latter two. This embodiment uses an interface where one request corresponds to one structured response. After receiving the request, the interface returns a task reception status, performs the verification, and returns the business result. Successful reception does not equal successful verification. The single waiting limit is configured by the platform's allowed business waiting time and the response time promised by the model interface. The processor records the sending time and the monotonic timer value. The waiting time is obtained by subtracting the sending timer value from the current timer value. If the waiting time reaches the limit and no valid response is obtained, it is recorded as a transmission timeout, and the task content is retained. The same verification batch and task order are reused, and an attempt number is added. Duplicate responses are identified by the task key. After receiving a task, the review agent first reads the fixed review rule version, then checks the required factual input according to the rule order within the task. Rules that can be directly executed based on structured facts are completed by the rule execution program. For rules that require interpretation of descriptive text, the corresponding content segment and rule description are sent to the artificial intelligence language model to propose candidate interpretations. After the review agent writes back the entities, fact categories, states, and values involved in the candidate interpretations to the rule input, it still performs the aforementioned field and source checks. If the check fails, it returns insufficient evidence or unqualified format.
[0053] For program record review, the review agent reads the process record categories and corresponding factual states required by the rules, and returns categories such as complete records, missing records, or pending confirmation based on existence and chronological order. For document field review, the review agent extracts standardized factual values from the content segments of specified fields in the document, and reads corresponding records in boundary content segments for compatibility verification, outputting whether fields are consistent, conflicting, or lack sufficient basis. Both review items retain the content segments and rule references passed through. For attachment factual review, the review agent reads the attachment identifier and the attachment identifier referenced in the main document, searches for matching according to the entity identifier and the corresponding attachment status. If a single match is found and the required factual values are compatible, the corresponding valid result is returned; if multiple matches cannot be uniquely determined, the corresponding pending confirmation result is returned. For cross-consistency review, the review agent directly reads the support or conflict identifiers already saved in the review relationship chain, checks whether their factual states are consistent with the rules, and references the boundary content segments of adjacent items when necessary. The review agent returns an ordered reference path for each conclusion. Each entry in the ordered reference path includes the path position, paragraph identifier, material position, rule identifier, review rule version, and fact field obtained for that entry. The path positions are saved in the order in which the review agent actually reads and uses them for the current conclusion when executing the rules. The conclusion formation process referred to in this embodiment is this type of verifiable explicit reference and rule execution order. It does not require reading or assumes that it can read the undisclosed reasoning state inside the artificial intelligence language model. Therefore, it can directly use structured records for subsequent comparisons.
[0054] Ordered reference paths are recorded synchronously during rule execution. The review agent reads one input reference entry for the current conclusion each time, first obtaining the position of the current end entry, writing that position as the predecessor of the new entry, then assigning the new entry position and appending it to the current path. The rule execution result is saved as a result record and linked to its complete input reference group. Candidate conclusions save the rule result identifier they adopted. Paths cannot be guessed backwards from a post-model-generated explanatory text; a path is only accepted if the runtime record and the rule result declared in the conclusion correspond. When multiple content segments share a rule, the review agent reads them sequentially according to the input role order and content position order determined by the task. All inputs must be obtained before the rule is executed and written to the result record. The path saves the actual execution order. After receiving the path, the processor checks if the order conforms to the task requirements. If an order error is found, the order verification fails and the original task is resent. The processor does not reorder path entries to eliminate differences after receiving the result, as post-ordering may change the true predecessor relationship and move the first divergence position to the wrong position. To enable the paragraph identifier and the review rule version to jointly identify the comparison object, each path entry also stores a rule identifier. In this embodiment, the review rule version points to an immutable rule snapshot, while the rule identifier points to a specific rule in the snapshot. When different rules are in the same snapshot, they cannot be identified as the same rule simply because they have the same version. The processor first checks the rule identifier and then checks the review rule version, using both as the rule reference identity to avoid different rules of the same version being mistakenly included in the same chain segment.
[0055] After receiving candidate conclusions, each citation path entry is read one by one. The corresponding segmentation results are searched according to the paragraph identifier. If the identifier exists, the material location, revision number, and original text citation range are compared. If the identifier does not exist or the material location is inconsistent, the citation verification fails. If the review agent cites a content segment that actually exists in the current case but does not belong to this task, the verification will also fail. This citation is not subject to task boundary and version constraints, preventing external materials from entering the candidate conclusion chain without authorization. Material location comparison is performed on text, images, and audio / video separately by field. For text, the material identifier, revision number, paragraph order, and original character range are compared. For images, page number, coordinate reference, and rectangle sides are compared. For audio / video, the original file identifier and start / end timecodes are compared. When there are multiple locations, the list length and the order of each location in the list are also compared. The location field should be referenced by the review agent as is from the task object; approximate locations rounded by the model are not accepted. Only after the precise location verification passes can the original text be further checked to see if it supports the current facts. Next, the rule reference identity is verified. The rule identifier and review rule version carried by the candidate conclusion are compared item by item with the list of rules allowed for the task. If the rule is not in the list, the version is inconsistent with the task, or the rule content record is missing, it is recorded as a rule verification failure. Only after both verifications pass, the processor writes the conclusion content, the referenced content segment, and the review rule into the valid candidate conclusion record and connects them according to the task order. The natural language description of the candidate conclusion can be saved separately but does not replace the structured conclusion category and fact value. If any verification fails, the processor resends the original collaborative review task to the corresponding review agent. The original task division result, review rule version, and task order remain unchanged. At the same time, a failure reason field is added to clearly indicate the invalid paragraph identifier or incorrect rule reference. The failure reason is used to correct the return format without adding new review materials. The same two verifications are performed every time it is resent. Only when both pass can a valid candidate conclusion for the task be obtained. Through this process, the source range of the candidate conclusion chain is kept within the original allocation result.
[0056] In cases of continuous verification failures, the processor reads the maximum automatic retransmission limit from the platform's recovery configuration. This limit is determined based on the total allowed waiting time for a single task, the maximum waiting time per retransmission, and the allocated computing resources, and is checked to be a finite non-negative integer at the time of publication. Each automatic retransmission increments the retransmission count. If the count reaches the maximum and verification still fails, the task is paused and written to a pending recovery status. The original task and failed responses are preserved. After the interface or material location issues are resolved, the original task is retransmitted and verification continues. Pausing is not considered a successful verification and does not generate valid candidate conclusions. If the same task receives two valid responses with different conclusions after a timeout, the processor does not overwrite the later response with the earlier one. Instead, it saves both responses separately and marks them as inconsistent within the task. It uses the same reference path comparison process. When both responses have the same conclusion and path, the valid response with the earlier attempt number is recorded as the primary response, and the other response is registered as a duplicate return. This avoids network factors changing the number of fact judgments or masking the instability of model output. When connecting candidate conclusion chains, the processor first establishes task nodes according to the task order, and then arranges candidate conclusions within the task nodes according to the fact category registration order, entity registration order, and fact status identifier. Each candidate conclusion still retains its own path; conclusions without conflicts also retain their paths for subsequent tracing. Unfinished task positions are kept in a vacant state and are not directly spliced with the next task to form a completed chain. Based on this, step S104 can distinguish between normal differences and task missing, and avoid calculating false first divergence positions on incomplete chains.
[0057] When determining whether candidate conclusions conflict, comparison groups are established based on case entity, fact category, and fact status identifier. The structured categories of each conclusion are then compared with the standardized fact values. Mutually exclusive categories or fact values that cannot be co-existed form conflicting candidate conclusion pairs. Conclusions on different issues in different review matters do not conflict due to differences in wording. If the evidence for the same issue is insufficient to determine the result, it is treated as a pending confirmation group. By limiting the comparison groups, subsequent cross-reviews have clear target facts. When establishing conclusion comparison groups, each case entity, fact category, and fact status identifier is checked to ensure they are all definite values. Group records are searched according to the fixed order of these fields. Candidate conclusions with identical fields are added to the same group; if any field differs, a new group is created. Before adding a candidate conclusion, the candidate conclusion identifier is checked for duplication. If duplicates are found, the number of members is not increased. If unknown grouping fields exist, the conclusion is moved to a pending confirmation group. All unknown fact statuses should not be placed in the same comparison group to prevent different unknown events from being incorrectly assigned for cross-review. When there are more than two inconsistent conclusions within the same comparison group, the processor enumerates the conclusion pairs in order of task sequence. Each conclusion pair is assigned a review number based on a fixed order of the two conclusion identifiers. All reviews are based on the original candidate conclusion chain. In this embodiment, other candidate conclusions are not overwritten immediately after the first pair of reviews are completed, to avoid the input of the later review task being asynchronously modified by the previous task. After the above processing, the candidate conclusion chain carries the review items, execution status, valid candidate conclusions, ordered reference paths, material locations, and review rule versions for each task, and saves conflicting candidate conclusion pairs and combinations to be confirmed. Step S104 only determines the divergence position for conflicting candidate conclusion pairs with complete and valid reference paths. Combinations where the task is incomplete, the source is invalid, or the factual status is unknown remain to be confirmed.
[0058] Step S104, determining the divergence location and cross-checking, specifically includes: Based on the candidate conclusion chain, compare the content segments and review rules traversed by conflicting candidate conclusions to determine the position where different content segments or review rules first appear. Retain the same chain segments before the position and cross-send the different chain segments after the position to the corresponding review agent for re-review, resulting in a review conclusion chain. Read a conflicting candidate conclusion pair, retrieve the verified ordered reference path from their respective candidate conclusion records, and use the starting entries of the two paths as the comparison starting point. Before comparison, check whether the two paths belong to the same review batch and the same fact comparison group, and check whether the content segments and rules referenced by the path entries are still readable. If any check fails, the candidate conclusion pair is converted into a pending confirmation combination. Do not establish common chain segments between paths that are across batches or have broken sources, ensuring that common inputs have the same bill of materials basis. Simultaneously read entries from the starting ends of the two paths, first comparing the segment identifier, then comparing the rule identifier and review rule version of the entry. If all three fields are the same, continue comparing the next position; if any field is different, stop the forward comparison and determine the current position as the divergence position. The divergence position stores the entry sequence number, material position, rule reference, and difference type for each of the two paths. The synchronous comparison first sets the read positions on both sides as the first position of their respective paths, sets the number of confirmed identical entries as an empty count, and retrieves the current entries on both sides one by one. When the three reference identity fields are the same, the pair of entries are appended to the common front list and moved forward one by one respectively. If they are different for the first time, the current actual positions on both sides are written and the comparison stops. If one side has reached the end while the other side has not, the difference in the end status is written. If both sides end at the same time, the comparison loop returns that the entire path is the same. The comparison loop reads at most the longer path length plus one end status, and does not perform cross-position search to find longer non-contiguous common parts.
[0059] Paragraph identifiers that are the same but differ in material revision number, original text, or material location constitute data integrity anomalies and are not treated as ordinary material disagreements. The processor marks the entry as having corrupted citation identity and stops automatic review. Similarly, reviewing rules with the same version but whose content records have been overwritten also constitutes rule integrity anomalies, requiring restoration of an immutable snapshot before continuing. By separating corrupted identity from normal differences, common segments are prevented from being fixed on identifiers that have lost their stable meaning. Consecutive entries preceding the first occurrence of different entries form common segments, which are the longest common prefixes of two paths starting from their starting positions. Identical entries that reappear after a disagreement are not spliced across disagreements into common segments. Since the interpretation of later entries may depend on the already disagreed preceding material, even if the paragraph identifiers of later entries are coincidentally the same, they are still included in the review of differing segments. In this embodiment, the divergence position is considered the starting point of the difference chain segment. All chain segments describing the divergence position, during execution, begin from the first differing entry, including that entry and all subsequent referenced entries up to the candidate conclusion. If only the successors of the first differing entry are exchanged, omitting the first differing entry, the review will be unable to read the direct source of the difference. When extracting the first and second difference chain segments, an interval including the starting point is used; identical chain segments do not include the divergence position. In cases where one path ends first while another path still has subsequent entries, the processor sets a path end position after the end of the shorter path. The actual entry at that position in the longer path is compared with the path end state to determine a path length divergence. The difference chain segment of the shorter path can be empty but still carries the candidate conclusion and end state. The difference chain segment of the longer path contains the remaining entries. During cross-review, the review agent clearly knows that the other party has not continued to cite materials, rather than receiving a message of incomplete format.
[0060] When two paths are identical in all entries and lengths but conflict in their conclusions, and there is no point of divergence that can be determined based on differences in content segments or review rules, the processor records this combination as a conflicting conclusion along the same path, retaining the complete common reference path and two candidate conclusions. This combination can be re-requested for review after fixing all identical inputs, but it cannot claim to have completed the location through the first different content segments, nor can it automatically create a material divergence point. If it is still inconsistent after the re-request, it will enter the pending confirmation record. If the two paths differ in the first entry, the same chain segment is empty, and the processor saves a common starting point record. The record includes the review batch identifier, fact comparison group, and empty front status, but this record does not impersonate a material chain segment that has already been consistent. During the review, the complete two difference chains are sent. The result merging requires checking the consistency of all review paths and conclusions in addition to checking the common starting point. Empty same chain segments do not lead to a narrowing of the review scope, and avoid describing a situation without common materials as having sufficient common evidence. The initial divergence position reflects the first difference in the observable reference sequence, but does not directly prove that this position is the sole cause of the semantic conflict. The processor stores the difference type in the record to be confirmed as material reference difference, rule reference difference, or path length difference. If the review proves that the facts described in the two material segments are compatible, the processor can reach a consistent review conclusion, but still retains the original divergence record for retrospection. After obtaining the same chain segment, its entry and the referenced content segment are set to read-only status in the current review task. The read-only status means that the review agent can read these records but cannot rewrite their text, material position, rule content, or entry order through the review interface. The processor stores an immutable copy in the storage layer, only provides read permissions in the task interface, and the review return message only allows writing new difference chain segment references and conclusions.
[0061] When set to read-only status, the common entry list and all its references are written to an independent review snapshot. The snapshot key consists of the review batch and review number. After all entry references have passed the check, the snapshot status is changed from "Preparing" to "Readable". The review interface only accepts requests to read common input with this snapshot key and does not provide write actions that overwrite its content. Update requests return rejection and the request task is recorded. The difference segment and review result are written to another appendable area. Before the task ends, the snapshot reference count is retained. Common records are not reclaimed until the reference count reaches zero. The read-only requirement is implemented into executable write and lifecycle rules. The same segment simultaneously saves its end position and complete entry list. The end position is used to connect new difference segments, and the complete entry list is used for item-by-item verification. The processor does not save only one common segment name that can be arbitrarily filled in by the review agent. At the end of the review, the processor rereads the original list, checks the paragraph identifier, rule reference identity, and entry order. Only if all are the same is the review conclusion accepted and connected to the fixed same segment. By verifying actual records, the read-only requirement is ensured to be more than just a textual requirement in the task prompt. Starting from the divergence point of the first candidate conclusion, its content segment, material location, review rule version, and rule reference identity are obtained, forming the first difference chain segment in the original path order. The second difference chain segment is formed from the corresponding position of the second candidate conclusion. The first and second difference chains segment respectively store the original review agent identifier, the original candidate conclusion identifier, the end of the preceding identical chain segment, and the chain segment validity status. During cross-transmission, the first difference chain segment is sent to the review agent that formed the second difference chain segment, and the second difference chain segment is sent to the review agent that formed the first difference chain segment. The same read-only identical chain segment is written into both tasks as common input. Each re-review task also carries the current fact comparison group, the list of allowed review rules, and the required return connection field, enabling the review agents to re-examine the difference basis used by the other party around the same fact, thereby separating the source of difference from the existing division of labor of the executing agent.
[0062] Two auditing agents are responsible for different auditing matters. In this embodiment, they are configured to jointly possess the ability to check factual value compatibility and execute rules. The cross-review task carries the specific audit rules to be executed and the allowed rule types. If the receiving auditing agent cannot execute a certain provision in the other party's difference chain segment, the interface returns a capability mismatch and the review remains incomplete. The processor does not unilaterally replace the rules or make the auditing agent guess the execution method. The review can only be resumed after the task configuration supplements the same rule execution capability. After receiving the difference chain segment, the auditing agent reads the structured fact records and original material fragments that have been verified as known conditions in the common input, and then reads the received difference chain segment entries in sequence. It extracts fact fields according to the input order of the rule records and performs comparisons. The auditing agent reports that the received chain segment has insufficient source, inapplicable rules, or factual conflicts. When a problem is found in the common input, a common input confirmation flag is generated, and the processor keeps the entire group unmerged. The common input for re-review primarily consists of fixed materials and rule bases. Unverified natural language inferences from the original candidate conclusions are not treated as immutable facts. Only fields obtained from source verification and rule determination within the same chain segment are passed in as stable input. If the original path contains a free description of the model but has no traceable source, that description is not included in the valid chain. When the review agent returns the review conclusion, it saves the conclusion category, fact value, fact state, the actually used difference chain segment entries, and the end reference of the common chain segment. It also saves the original candidate conclusion identifier and the review task identifier. The review conclusion may differ from the original candidate conclusion; the processor writes it as a new record without overwriting the original record. Figure 3 The two cross connections in the code represent this source exchange, and the two returned results can always return to the original difference chain segment and the receiving auditing agent, respectively.
[0063] During the review phase, updated materials cannot be read from outside the task to fill in discrepancies. If new materials are indeed needed, the review agent returns the supplementary material requirements and the missing fact categories. The processor keeps the current review pending confirmation and establishes a new review batch after the supplementary materials are obtained. The new batch re-executes the standard review data and review relationship chain formation, while the results of the old batch are retained. Batch isolation ensures that identical segments remain the same version throughout the review process, rather than quietly changing the common context during the review. The re-review also performs in-task references and rule version verification. When the review return references content segments other than the received discrepancies and common inputs, the processor does not directly connect the results but resends the original review task to the review agent. If consecutive failures reach the aforementioned automatic resend limit, the review is paused, and the failure location and interface reason are saved. Even if the other party has returned a consistent category, it cannot be regarded as a consistent result between the two parties. When the received difference chain segment is empty, the review agent only reviews whether the candidate conclusion has sufficient supporting sources based on the common input and the path end state. If the rule requires additional facts but the input is missing, the return is insufficient. If the common input is sufficient to execute all necessary provisions, a definite conclusion can be returned, but the end of the common chain segment must be explicitly referenced. Non-existent content segments cannot be generated to meet chain length requirements. The review conclusion chain is established in the order of fixed common chain segments, returned difference chain segments, and review conclusions. The connection is recorded item by item using the predecessor position field. The predecessor of the first difference entry must point to the end of the common chain segment, and subsequent entries must point to the previous entry of the current review task. When the returned record does not declare a predecessor, declares a non-existent predecessor, or forms a loop, the review response is retained and marked as unconnected. The unqualified result is not transformed into a qualified path by automatically filling in missing edges.
[0064] In one implementation that saves on repetitive reads, two review tasks share a read-only copy of the same chain segment in memory. Each review agent can still read the complete text and rules of the common input. The processor only avoids the repeated copying of common records and the re-extraction of deterministic fields. If the language model interface requires the entire context to be re-entered each time, the processor still needs to send the common input to the model. This part of the text transmission cannot be counted as saved. Therefore, subsequent simulations separately calculate the amount of recalculation of difference items and the amount of reading the complete context. When caching common input, the review batch, common chain segment snapshot key, rule version, and data representation version are used as cache keys. Before writing to the cache, a complete list of entries and validity flags are saved. After the cache is hit, the key and list length are compared. The paragraph identity and rule identity are checked item by item. If all are the same, the saved common record is returned. If they are different, it is read again from the immutable snapshot. Cache eviction only deletes reconstructable copies and not authoritative snapshots. When the model version or data representation version changes, the old model intermediate features are not reused.
[0065] Step S105, reviewing the connection verification and generating the case business review results, specifically includes: Based on the review conclusion chain, verify whether each review conclusion connects to the same chain segment. If they connect to the same chain segment and the conclusion content is consistent, merge the review conclusions. If they do not connect to the same chain segment or the conclusion content is inconsistent, retain the position and the different chain segments that follow, thus obtaining the case business review result. Read the end path position, paragraph identifier, rule identifier, and review rule version from each review conclusion. Search backwards along the predecessor position, first verifying that the current item belongs to the corresponding review task, then reading the predecessor item, repeating until the end of the same chain segment is reached. During the backward search, add the already read item positions to the access list. When the same position is encountered again, mark it as a loop; if the predecessor does not exist, mark it as a broken chain. Do not skip items and continue searching for a common chain segment with the same name when an anomaly occurs. After finding the end of the same chain segment in reverse lookup, the processor continues to check the complete list of entries for the same chain segment, comparing its segment identifier, material location, rule reference identity, and original order to see if they are the same as the read-only copy. If the review conclusion only mentions the same chain segment name without forming a valid predecessor connection, the chain segment connection result is still unconnected. Only when the path is continuous, all references are valid, and the common record remains unchanged, the chain segment connection result indicates that it has been connected. The maximum number of steps in reverse lookup is the actual number of entries in the current review conclusion chain plus one. This upper limit is determined by a finite number of entries and does not depend on an empirical threshold. If the upper limit is exceeded but the common chain segment start point has not been reached, it is considered a connection anomaly. If the current chain has no entries and the common chain segment is not empty, it is directly judged as unconnected. When the common chain segment is empty, the completeness of the common start point record and the difference chain segment is checked. Empty chain scenarios are not incorrectly classified as connection failures, nor are the checks of difference materials and rule references exempted.
[0066] The reverse lookup uses an access list to identify nodes by verifying task identifiers and entry positions. For each predecessor obtained, it first checks if the joint identity has already been accessed, then checks if the snapshot or difference region to which the node belongs is within the current task's allowed range. After reaching the end of a common chain segment, it continues to verify back to the common starting point according to the original list order. If a segment in the path is referenced multiple times by different rules, it will not be incorrectly judged as a cycle simply because the segment identifier is the same; a cycle is only judged if the same entry position appears again. Therefore, reused materials and erroneous loop connections have different calculation criteria. In addition to indicating whether a chain segment is connected or not, the chain segment connection result also stores the reason for the inability to complete the verification, such as the content segment being temporarily unreadable, the rule snapshot being missing, or the task not returning. These reasons are expressed in the result as pending confirmation. When a temporary read fails, the processor can reread according to the original task key. After the read is restored, a complete verification is re-executed. Before restoration, the unknown is not treated as already connected; by retaining the unknown state, storage failures prevent two similar conclusions from being incorrectly merged. After the connection verification is successful, the processor compares the case entity, fact category, fact status identifier, conclusion category, and standard fact value of the review conclusion. The first three fields are used to confirm that the two results still belong to the same comparison group, and the last two fields are used to determine that the conclusion content is consistent. The comparison of standard fact values follows the type conversion result of step S102. String enumerations are compared according to the specified values, and precise quantities are compared according to integers. For fact values with ranges, the upper and lower boundaries and inclusion flags must be the same for the conclusion content to be consistent. If the ranges only have intersection, it does not mean that the results are the same.
[0067] If two review conclusions have different natural language descriptions but identical conclusion categories and normative fact values, and both pass the connection verification, the two descriptions are retained as source interpretations, and a single consistent review conclusion is generated. If two descriptions have the same text but different structured fact values, they are not merged. This is because the actual judgment object is determined by structured fields, and language models are not used to judge seemingly similar meanings to replace precise result consistency verification, reducing the possibility of introducing uncertain interpretations again during the summarization stage. When generating a consistent review conclusion, the processor creates a new result identifier and writes the case identifier, review batch identifier, fact comparison group, unified conclusion category, unified fact value, two review conclusion identifiers, common link segment reference, and difference link segment reference. The positions of the materials from both parties are saved as an ordered list after deduplication according to paragraph identifiers and original positions. The materials that caused the original conflict are not deleted so that users can still return to the formation process when viewing the consistent result. When merging the material location list, each location is appended in the original order of the first review path, and then the locations are read in the original order of the second review path. The existing list is searched item by item using the material identifier, revision number, location type, and complete location fields. If all fields are the same, only the source association of the second path is added; if any field is different, it is appended to another location. If an image region contains another region or two time intervals overlap, they are not treated as equal. The result retains the original paths of both review conclusions. If any review conclusion does not connect to the same chain segment, or if the two conclusions have inconsistent categories or fact values, the processor creates a pending confirmation record. This record includes the divergence location, the first difference chain segment, the second difference chain segment, the two review conclusions, the connection verification status, and the reason for not being able to merge. The reasons for not being able to merge are saved separately for non-connection, different conclusion categories, different fact values, missing sources, and common input pending confirmation. When multiple reasons exist simultaneously, all are listed, and a single general conflict flag does not cover the specific fault location.
[0068] For conflicts in conclusions along the same path, the disagreement position field in the pending confirmation record stores a clear state that there are no material or rule disagreements, while also storing the complete identical path and different conclusions, without writing a zero position to impersonate a chain head disagreement; for cases without common chain segments, the pending confirmation record stores an empty common chain segment and two complete difference paths. When multiple review conclusion pairs exist in the same comparison group, the processor first checks whether all required review pairs have been returned, then checks the connection results and conclusion content pair by pair. Only when all valid review conclusions point to the same conclusion category and fact value is a consistent audit conclusion for the group generated; when some conclusion pairs are consistent while others are still conflicting, the processor saves the partial state of the completed review pairs but does not publish the entire group's consistency, and for those without... For matters involving conflicting candidate conclusion pairs, the processor retains the valid candidate conclusions and their source verification status, marking them as audit conclusions that have not triggered cross-review in the case business review results. This marking is distinguished from the consistent audit conclusions formed through cross-review by both parties. The pending confirmation combinations and incomplete tasks are also listed separately. When generating the case business review results, the processor reads the consistent audit conclusions, audit conclusions that have not triggered cross-review, pending confirmation records, and incomplete tasks in the original task order, writes them into the result list and batch status. When all necessary tasks are completed and there are no pending confirmation records, the batch status is "audit completed." When there are pending confirmation records, it is "audit completed but pending confirmation." When there are tasks that have not been returned or have failed verification, it is "partially completed."
[0069] The results display program opens the original material based on its location. Text materials are positioned within the original character range, image materials display the page number corresponding to the image and mark the reference area with a rectangle, and audio and video materials play from the start timecode to the end timecode. When a content segment has multiple material locations, they are presented sequentially according to the original location list. If a location is invalid or the original file is temporarily unreadable, a positioning failure status is displayed and the result record is retained. It does not automatically replace with adjacent pages or similar time segments. When displaying the location of image materials, the display program reads the width and height of the original image and the width and height of the currently available display area, obtains the ratio of the display width to the original image width and the ratio of the display height to the original image height, selects the smaller one as the proportional display coefficient, multiplies the coordinates of each side of the original rectangle by this coefficient, and adds the horizontal and vertical starting point offset of the image in the display area to obtain the display rectangle. When the user clicks on the display area to retrieve the original text, the same offset is first subtracted and then divided by the same coefficient. If the display coefficient is zero or the image has not yet been obtained, positioning is not performed. Therefore, scaling the display does not modify the saved original material location. The review rules display reads a fixed snapshot based on the rule identifier and the review rule version. It does not overwrite the old conclusion basis from the current latest rule table. When the rule has been updated, it displays the old version identifier and the status of the new version. When the user initiates a re-review based on the new version of the rule, a new review batch is generated and the old and new results are linked. The original results and raw material locations can still be read.
[0070] Attached Figure: Simulation Configuration and Data Description The simulation uses a processor to compare the results line by line. Figure 4 Five single-path lengths (40, 80, 120, 160, and 200 entries) are used, generating 30 sets of dual-path tasks for each, totaling 150 sets. In each set, the first path creates an entry with a location identity, and the second path copies the first path and replaces the paragraph identifier at a predetermined branching position. The common segment percentage is generated by multiplying the trial number by 17, dividing by 81, adding 10 to get the target percentage, multiplying by the path length, and rounding to the nearest integer to produce a distribution ratio within the range of 10% to 90%. Based on the number of common entries, the first entry after that is set as the first branching entry. The full-chain comparison method reprocesses all entries on both paths once, with the total number of processing steps being twice the length of a single path. The difference segment method subtracts the actual number of common entries from the single path length to obtain the number of difference entries on one side, and adds the two sides together to get the number of recalculated entries. Figure 4 The whole-chain control lengths were 80, 160, 240, 320, and 400, respectively, while the differential chain segment lengths were 41.93, 83.07, 124.13, 165.27, and 205.60, respectively.
[0071] Figure 5 Considering the additional processing for retaining boundary content segments, the length of a single path is fixed at 200 entries. The common chain segment ratio is configured as 10%, 30%, 50%, 70%, and 90%, while the additional boundary processing ratio is configured as 0%, 5%, 10%, 15%, and 20%. The number of unilaterally differing entries is obtained by subtracting the number of common entries from the path length. The total number of unilaterally reprocessed entries is obtained by adding the number of unilaterally reprocessed boundary entries. The sum of the two totals is compared with the full chain comparison. This total is subtracted from the full chain comparison entries, divided by the full chain comparison entries, and converted to a percentage to obtain the net reduction rate of recalculated entries, ranging from -10% to 90%. Figure 6 from Figure 4 Thirty groups with a single path length of 120 were selected. The horizontal coordinate represents the injection branch position, and the vertical coordinate represents the branch position obtained by the processor through comparison. All 30 results are located on the position consistency reference line. Figure 7 The simulation employs 240 sets of boundary conditions, categorized into eight types: material differences, rule differences, persistent conflicts, connection breaks, different conclusions along the same path, empty common links, short chain omissions, and version mismatch. Each type generates 30 sets, with a basic path length of 20 entries. Differences are injected at the 9th entry; rule differences replace the rule identifier at that position; empty common links replace the first entry; short chain omissions cause one side to end after the 8th entry; and version mismatch is triggered by a discrepancy between the returned version and the task version, resulting in a validation rejection. The program actually outputs 90 sets of consistent connections, 120 sets reserved for confirmation, and 30 sets of validation rejections. Figure 8Using 30 groups of tasks with a single path length of 120, the lengths of the differential chain segments on one side of each group were counted. The statistical intervals were 0-19, 20-39, 40-59, 60-79, 80-99, and 100-120. The calculation results were 2, 6, 6, 6, 6, and 4 groups, respectively.
[0072] All five simulation diagrams use the same entry identity definition and divergence comparison rule. Figure 4 , Figure 6 , Figure 8 Share all or part of the records from the 150 generated task groups. Figure 5 It is a deterministic parameter scan of the boundary cost. Figure 7 This is a branch validation based on predefined conditions. All averages and frequencies can be retrieved from the path length, the number of common entries, and the branch conditions.
[0073] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A case business review and analysis method based on multi-agent collaboration, characterized in that, The method includes: Step S101: Obtain the case business materials and their corresponding material locations, formation times, and review rule versions; standardize the material format and divide the content into sections; write the material locations, formation times, and review rule versions into the corresponding content sections to obtain standard review data. Step S102: Based on the standard review data, arrange each content segment according to the case entity and the time of its formation, mark the supporting or conflicting relationships between the content segments, and connect the review rules with the corresponding content segments to obtain the review relationship chain; Step S103: According to the review relationship chain, divide adjacent content segments according to the review items, retain the content segments at the connection points of different review items in the adjacent division results, generate collaborative review tasks according to the order of each division result, send the collaborative review tasks to the corresponding review agent, and obtain a candidate conclusion chain carrying the material location and review rule version. Step S104: Based on the candidate conclusion chain, compare the content segments and review rules traversed by the conflicting candidate conclusions, determine the position where different content segments or different review rules first appear, retain the same chain segments before the position, and cross-send the different chain segments after the position to the corresponding review agent for re-review to obtain the review conclusion chain. Step S105: According to the review conclusion chain, check whether each review conclusion is connected to the same chain segment. If the chain segments are connected and the conclusion content is consistent, merge the review conclusions. If the chain segments are not connected or the conclusion content is inconsistent, retain the position and the different chain segments thereafter to obtain the case business review result.
2. The case business review and analysis method based on multi-agent collaboration according to claim 1, characterized in that, Obtain case-related business materials along with their locations, creation dates, and review rule versions; standardize the material format and divide them into content segments; write the material locations, creation dates, and review rule versions into the corresponding content segments to obtain standard review data, including: Convert text materials into paragraph text, convert image materials into recognized text with image region coordinates, and convert audio and video materials into transcribed text with timecode; The converted text is divided according to paragraph boundaries, image region boundaries, or continuous semantic boundaries; the material identifier is combined with page number, image region coordinates, or audio / video timecode to form the material location; Each content segment, along with its corresponding material location, creation time, and review rule version, is encapsulated into a single data record to obtain standard review data.
3. The case business review and analysis method based on multi-agent collaboration according to claim 1, characterized in that, Based on standard review data, the content segments are arranged according to the case entity and the time of its formation. Supporting or conflicting relationships between the content segments are marked, and review rules are connected to the corresponding content segments to obtain the review relationship chain, including: Extract entity identifiers, fact categories, and fact content from each content segment; arrange content segments with the same entity identifiers according to their creation time. A support relationship is marked when the factual content of adjacent content segments can be true at the same time; a conflict relationship is marked when the factual content of adjacent content segments cannot be true at the same time. Find the corresponding content segment according to the fact category recorded in the review rules, and connect the found content segment to the review rules; obtain the review relationship chain according to the order of the content segments and the connection order between the content segments and the review rules.
4. The case business review and analysis method based on multi-agent collaboration according to claim 3, characterized in that, A supporting relationship is marked when the factual content of adjacent content segments can be simultaneously true; a conflicting relationship is marked when the factual content of adjacent content segments cannot be simultaneously true, including: Compare the entity identifier, fact category, formation time, and fact value of adjacent content segments; when the entity identifier and fact category are the same, and the formation time corresponds to the same fact state, determine whether the fact values can be written into the same fact record. When both can be written together, a support flag is written between adjacent content segments; when both cannot be written together, a conflict flag is written between adjacent content segments, and the corresponding material location is also written with the conflict flag.
5. The case business review and analysis method based on multi-agent collaboration according to claim 1, characterized in that, Based on the review relationship chain, adjacent content segments are divided according to the review items. Content segments connecting different review items are simultaneously retained in the adjacent segmentation results. Collaborative review tasks are generated according to the order of the segmentation results and sent to the corresponding review agents. This yields a candidate conclusion chain carrying the material location and review rule version, including: Determine the audit items to which each audit rule belongs; divide the audit relationship chain where adjacent audit rules belong to different audit items; Copy the content segments on both sides of the division position to the adjacent division result, and retain the same paragraph identifiers before and after copying; form the task order according to the order of the review rules contained in each division result; The partitioning results, task order, and identifiers of identical paragraphs are sent to the corresponding review agent; the candidate conclusions returned by the review agent are connected according to the task order to obtain a candidate conclusion chain.
6. The case business review and analysis method based on multi-agent collaboration according to claim 5, characterized in that, The candidate conclusions returned by the review agent are concatenated in the order of the tasks to obtain a candidate conclusion chain, including: Receive the conclusion returned by the review agent, the location of the cited content, and the version of the review rules used; Verify that the referenced content segment exists in the corresponding partitioning result, and verify that the version of the audit rules used is the same as the corresponding collaborative audit task; If both checks pass, connect the conclusion content, the referenced content segments, and the audit rules used, according to the task order; If any check fails, the original collaborative review task is resent to the corresponding review agent until both checks pass, thus obtaining a candidate conclusion chain.
7. The case business review and analysis method based on multi-agent collaboration according to claim 1, characterized in that, Based on the candidate conclusion chain, compare the content segments and review rules traversed by conflicting candidate conclusions to determine the position where different content segments or different review rules first appear. Retain the same chain segments before the position, and cross-send the different chain segments after the position to the corresponding review agent for re-review, resulting in a review conclusion chain, including: Compare the paragraph identifiers and review rule versions traversed by conflicting candidate conclusions sequentially from the beginning of the candidate conclusion chain; The first occurrence of a different paragraph identifier or a different version of the review rule is identified as a branch point; the chain segment with the same paragraph identifier and review rule version before the branch point is set to read-only; Exchange the chain segments of the conflicting candidate conclusions after the divergence position and send them to the corresponding review agent for re-review; connect the re-reviewed conclusions according to the chain segments in read-only state to obtain the review conclusion chain.
8. The case business review and analysis method based on multi-agent collaboration according to claim 7, characterized in that, Exchange the chain segments following the divergence point of conflicting candidate conclusions and send them to the corresponding review agent for re-review, including: Extract the content segments, material locations, and review rule versions from the point of divergence to the candidate conclusions of conflicting candidate conclusions to obtain the first and second difference chain segments; The first difference chain segment is sent to the auditing agent that formed the second difference chain segment, and the second difference chain segment is sent to the auditing agent that formed the first difference chain segment. Write the read-only chain segment as common input to two re-audit tasks; receive the review conclusions returned by the two audit agents based on the common input and the received differential chain segments.
9. The case business review and analysis method based on multi-agent collaboration according to claim 1, characterized in that, Based on the review conclusion chain, verify whether each review conclusion is connected to the same segment. If they are connected to the same segment and the conclusion content is consistent, merge the review conclusions. If they are not connected to the same segment or the conclusion content is inconsistent, retain the position and the different segments that follow, to obtain the case business review result, including: Read the paragraph identifier and audit rule version carried by each review conclusion; search for the same chain segment in reverse along the review conclusion chain to obtain the chain segment connection result; When the chain connection result indicates that the chain is connected and the conclusion category and fact value of each review conclusion are the same, the corresponding review conclusions will be merged into a consistent audit conclusion. When the chain segment connection result indicates no connection, or when the conclusion category or fact value is different, write the divergence position, different chain segments and review conclusion into the pending confirmation record; The case business review result is generated based on the consistent review conclusion and the records to be confirmed.
10. A case business review and analysis system based on multi-agent collaboration, wherein the system implements the method as described in any one of claims 1 to 9, characterized in that, This includes data processing nodes, relationship chain processing nodes, task coordination nodes, multiple auditing intelligent agent nodes, review processing nodes, and result storage nodes; The data processing node is connected to the relationship chain processing node. The data processing node processes the case business materials and their corresponding material locations, formation times and review rule versions into standard review data. The relationship chain processing node is connected to the task coordination node. The relationship chain processing node reviews the content segments in the data according to the case entity and the formation time, marks the supporting or conflicting relationships, and connects the review rules with the corresponding content segments to form a review relationship chain. The task coordination node connects with multiple review agent nodes. The task coordination node divides the review relationship chain according to the review items, retains the common content segments of adjacent division results, sends collaborative review tasks, and receives candidate conclusion chains returned by multiple review agent nodes. The review processing node is connected to multiple review agent nodes and result storage nodes. The review processing node determines the position where conflicting candidate conclusion chains first appear with different content segments or different review rules, retains the same chain segments before the position, cross-sends the different chain segments after the position, and forms the case business review result based on the returned review conclusion chain.
