Bidder verification method, device and medium based on timing attenuation and synchronization detection

Abstract

The application belongs to the technical field of Internet, and discloses a tenderer verification method and device based on time sequence attenuation and synchronous detection and a medium, which are applied to a bidding and tendering evaluation separation and calibration link. The method comprises the following steps: collecting historical written statement texts of a tenderer across multiple years and multiple government scenarios, labeling time stamps to construct a statement benchmark set with time sequence anchors; adopting a sliding window to compare the texts across periods, triggering digital person customized statement question and answer when the threshold is exceeded; collecting question and answer audio and video, extracting voice pauses, micro expressions, and gaze trajectories to generate emotion fluctuation curves; calculating semantic differences and configuring time sequence attenuation to form a time sequence curve; positioning the synchronous peak interval of the emotion curve and the time sequence curve of the corrected semantic difference, and counting the peak density, and determining that the statement is untrustworthy and outputting a quantitative untrustworthy score when the threshold is exceeded. The application weakens the interference of distant historical statements through time sequence attenuation, improves the accuracy of untrustworthy determination through two-dimensional detection of emotions and semantics, and effectively serves the calibration decision.

Description

Technical Field

[0001] This application relates to the field of Internet technology, and in particular to a bidder verification method, device and medium based on time decay and synchronization detection. Background Technology

[0002] The separation of evaluation and award processes treats the "bid evaluation by the bid evaluation committee" and the "bid award by the tendering party" as two relatively independent stages. The bid evaluation committee is responsible for recommending candidates to the tendering party, and the tendering party forms a bid award committee to select the best bidder from among the candidates. This model primarily addresses issues such as bid rigging and collusion, while also reinforcing the primary responsibility of the tendering party.

[0003] In existing technologies, the verification work of the bid evaluation committee mainly relies on bid documents and expert scoring data. This information source is relatively singular, making it difficult to identify inconsistencies and risks of dishonesty in the bidder's statements across different times and scenarios. Taking the medical field as an example, the procurement of medical supplies, equipment, and information systems directly relates to medical safety and patient interests. If a supplier provides false statements or conceals dishonest information during the bid evaluation process, the consequences are more severe than in general procurement projects. Therefore, existing technologies have proposed relevant bid evaluation assistance techniques to identify such content. However, the inventors have found that existing bid evaluation assistance techniques have several problems. For example, over many years, bidders participate in different bidding projects and submit materials to different government regulatory platforms, generating a large number of written statements regarding their qualifications, performance, financial status, and contract fulfillment capabilities. These statements are scattered across different systems and time points, and existing technologies have not yet established a mechanism for chronologically collecting and comparing the consistency of historical statements from multiple government regulatory scenarios. Furthermore, existing technologies lack effective verification methods to determine whether the same bidder's statements on the same key matter are consistent across different years, and whether there is deliberate concealment or false statement. For example, while some regions have introduced a presentation and Q&A session into the bidding process, the questions used are based on generic templates. This makes it impossible to generate targeted inquiries based on discrepancies in the bidder's past statements. Furthermore, behavioral information such as pauses in speech, micro-expressions, and changes in eye contact during the Q&A session currently relies primarily on the reviewers' subjective perception, lacking automated data collection and quantitative analysis solutions. Additionally, existing systems primarily assess bidders' integrity through qualitative descriptions, lacking the technical means to quantify information such as the verification results of statements from different periods and behavioral characteristics during the Q&A session into an actionable credit score and dynamically update it to the bid-winning risk profile.

[0004] Therefore, there is an urgent need to provide a technical solution that can systematically verify the consistency of bidders' statements across periods, conduct multimodal quantitative analysis of the bid presentation and defense process, and quantify the verification results into a credit score to drive the dynamic updating of the bid-winning risk profile. This solution aims to address the problems of single information sources, lack of verification methods, and lack of quantitative assessment mechanisms for credit risk in existing bid-winning auxiliary technologies. Summary of the Invention

[0005] The main objective of this invention is to provide a bidder verification method, device, and medium based on time-series decay and synchronization detection, aiming to solve the technical problems of existing benchmarking assistance technologies, such as single information sources, lack of verification methods, and lack of quantitative assessment mechanisms for the risk of dishonesty.

[0006] To achieve the aforementioned objectives, the first aspect of this invention proposes a bidder verification method based on time-series decay and synchronization detection, comprising: Collect historical written statements retained by bidders in preset scenarios, and mark key matters in the historical written statements with segmented timestamps to construct a set of statement benchmarks with time-series anchors; Compare the differences in the statements of similar events in the statement benchmark set at different historical times, and trigger a digital human-customized presentation and Q&A process for the bidder when the difference exceeds a preset threshold; The audio and video data streams of the bidder's interaction with the digital human during the customized bidding presentation and Q&A process are obtained, and the real-time emotion fluctuation curve of the bidder is constructed based on the voice data and video data in the audio and video data streams. The audio and video data stream is converted into question-and-answer text with timestamps. The question-and-answer text is then sliced ​​according to a preset duration, and the semantic similarity difference between each slice and the corresponding time-series baseline text is calculated to form a semantic difference time series. Next, the time length difference between the timestamp of each slice and the timestamp of the corresponding baseline text is calculated, and a corresponding time-series decay coefficient is configured based on the time-series difference to weighted correct the semantic differences of each slice, forming a weighted corrected semantic difference time series curve. The time-series decay coefficient is negatively correlated with the magnitude of the time-series difference. The synchronous peak intervals of the emotional fluctuation curve and the semantic difference time series curve are matched, and the peak density within each synchronous peak interval is calculated. When the peak density is greater than the density value threshold, it is determined that the bidder has committed a breach of trust, and the corresponding quantitative breach of trust score is output. The quantitative breach of trust score is used to update the bidder's bid-winning risk profile.

[0007] A second aspect of the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described bidder verification method based on timing decay and synchronization detection.

[0008] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described bidder verification method based on timing decay and synchronization detection.

[0009] This application uses a time decay coefficient to weight and correct semantic similarity differences, thereby weakening the impact of long-standing historical statements on current creditworthiness assessments. In existing technologies, time decay factors are typically used only for allocating weights to contextual records; this application uses it to weight and correct the semantic similarity differences themselves. This coefficient is negatively correlated with the time span—the longer the historical statement is, the smaller its impact on current creditworthiness assessments. This mechanism can reduce misjudgments of current bidders' creditworthiness due to omissions or inaccuracies in reporting from years ago, making the verification results closer to the bidder's actual integrity.

[0010] Furthermore, this application achieves dual-dimensional cross-validation of statement dishonesty by synchronously detecting peaks and density statistics of emotion fluctuation curves and weighted semantic differences. In existing technologies, multimodal emotion recognition and semantic analysis are usually performed independently. This application performs synchronous correlation analysis between the two in the time dimension. That is, when semantic deviation and emotional abnormality simultaneously peak within the same time interval and exceed the density limit, it is judged as a high-confidence statement dishonesty. This dual-validation mechanism can reduce the misjudgment rate of single-dimensional analysis and transform subjective judgment into quantifiable judgment criteria.

[0011] Furthermore, this application utilizes a sliding window cross-period text difference comparison and targeted triggering mechanism to screen and verify the consistency of bidders' statements. In existing technologies, sliding windows are mostly used for keyword matching to trigger general question-and-answer sessions. This application transforms it into a cross-period text difference comparison tool. The window length can be adaptively adjusted according to the project type, and the fluctuation threshold dynamically scales with the distance of historical time. Moreover, it only triggers digital human-based responses for bidders with cross-period statement anomalies, eliminating the need for batch presentations to all candidates. This mechanism improves the verification efficiency of the bid awarding process and reduces invalid, large-scale responses.

[0012] This application also provides a systematic data foundation for verifying the consistency of statements across different periods by collecting historical written statements from bidders across multiple years and government scenarios and constructing a benchmark set with time-series anchors. Existing benchmarking assistance technologies lack a mechanism for time-series aggregation of historical statements from multiple government scenarios. This application organizes the scattered statements of bidders across different times and scenarios into structured benchmark data that can be searched along a timeline, making the judgment of "whether the statements of the same bidder on the same key matter are consistent in different years" have a calculable and traceable technical basis.

[0013] This application further provides quantifiable auxiliary basis for bid award decisions under the evaluation-determination separation model by outputting quantitative dishonesty scores and using them to update the bidder's risk profile. Existing systems mostly use qualitative descriptions to judge the integrity of bidders. This application quantifies the results of cross-period statement verification and multimodal behavior analysis into specific dishonesty scores, writes them into the bidder's risk profile, and pushes them to the bid award evaluation module. This facilitates the bid award committee in making decisions based on quantitative data, improving the scientific rigor and traceability of the selection process under the evaluation-determination separation model. Attached Figure Description

[0014] Figure 1 A schematic diagram of a bidder verification method based on time decay and synchronization detection according to an embodiment of the invention; Figure 2 This is a schematic block diagram of a computer device according to an embodiment of the invention.

[0015] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0017] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of features, integers, steps, operations, elements, modules, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, modules, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any modules and all combinations of one or more associated listed items.

[0018] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0019] Reference Figure 1 This invention provides a bidder verification method based on time-series decay and synchronization detection, comprising the following steps: S101: Collect the historical written statements retained by the bidders in the preset scenario, and mark the key contents of the historical written statements with segmented timestamps to construct a statement benchmark set with time-series anchors.

[0020] In this step, various historical written statements left by bidders in preset scenarios are collected to analyze whether there are inconsistencies in the bidders' descriptions of the same matter, thereby determining whether there are any integrity issues with the bidders. To ensure the accuracy and efficiency of data analysis, the information in the historical written statements is not used. Instead, only the key content is extracted and marked with segmented timestamps to form a statement benchmark set with time-series anchors. In this way, the scattered statements of bidders at different times and in different scenarios are integrated into structured benchmark data that can be retrieved along a timeline.

[0021] The aforementioned historical written statements can be written records submitted or retained by the bidder in past bidding activities, annual reports, performance acceptance, regulatory inspections, and other government-related scenarios. Key information refers to content that reflects the bidder's integrity, such as core elements like company qualifications, performance, financial status, staffing, and performance capabilities. Segmented timestamps mark each independent statement with a specific time marker, allowing for the setting of time ranges based on granular requirements; in this embodiment, "day" is used. It should be noted that before adding timestamps, relative time expressions in each statement (e.g., "last three years," "last five years") are converted to absolute start and end dates (e.g., "January 1, 2022 to December 31, 2024") to ensure comparability of the actual time ranges indicated by statements from different periods. The time-series anchor point is the time-marked baseline record point of the statement.

[0022] When collecting historical written statements, the following paths are generally used: (1) Through the standardized interface of the electronic archive system of the public resource trading platform, the bidder's unified social credit code is used as the query key to obtain the historical bid documents archived by the bidder in the platform's trading projects. This electronic archive system is made in accordance with the national standard "Specifications for Archiving Electronic Documents of Bidding and Tendering". It collects bidding documents, tender documents, bid evaluation reports and other full-process materials in the form of "one project, one archive". The data can be retrieved online through the interface of the electronic archive system of the trading platform. (2) Through the official API interface of the National Enterprise Credit Information Publicity System, the annual report data of the enterprise over the years can be obtained after authorization by the bidder. (3) Through the government data sharing and exchange platform, in accordance with the "Regulations on Government Data Sharing" and related data sharing approval procedures, apply to obtain high-level government data such as on-site inspection records and historical project performance acceptance materials. After approval and authorization by the data provider, the data can be collected through the platform interface. (4) If some historical documents are only saved in paper form, the structured text can be extracted by scanning and digitizing them through the OCR recognition module. If data for a certain type of government service scenario cannot be obtained due to reasons such as access restrictions, lack of data digitization, or lack of response from the data provider, the system will downgrade the data according to the credibility level: the highest level of data that can be obtained will be used as the time-series benchmark anchor point for the matter; if no usable historical benchmark data can be found for the matter, the cross-period consistency verification of the matter will be skipped, and the verification report will be marked as "data missing - not verified", which will not be used as the basis for the judgment of dishonesty.

[0023] In one specific implementation, a medical device procurement project of a tertiary hospital enters the bidding stage. The system collects historical written statements from a winning bidder, Company A, over the past five years in different scenarios from channels such as public resource trading platforms and enterprise credit information disclosure systems. Specifically, this includes statements from Company A in its bid documents from three different years regarding "performance in similar projects over the past three years," statements regarding "total operating revenue" in two annual corporate reports, and a written explanation regarding "delivery cycle" during a medical equipment performance acceptance process. The system categorizes these texts according to key items such as "performance," "revenue," and "delivery cycle," and marks each statement with a specific date, thus forming a time-stamped set of statements for the bidder. For example, Company A stated in its 2021 bid document that it had "completed 5 similar projects in the past three years," and in its 2023 annual report that it had "completed 7 medical equipment supply projects from 2020 to 2022." Both of these statements are timestamped and added to the benchmark set.

[0024] S102: Compare the differences in the statements of similar events in the statement benchmark set at different historical times, and when the difference exceeds a preset threshold, trigger a digital human-customized presentation and Q&A process for the bidder.

[0025] This step compares the bidders' descriptions of the same matter at different historical times to see if there are discrepancies. If discrepancies exist, and the magnitude of the discrepancy exceeds a preset threshold, the digital human's presentation and Q&A process is triggered. Because a discrepancy exceeding the preset threshold indicates a possible dishonesty by the bidder, the system's digital human Q&A process is activated. In this embodiment, the Q&A text in the digital human-customized presentation and Q&A process is the text corresponding to the questions about the discrepancies exceeding the preset threshold, rather than the preset standard questions used in existing technologies.

[0026] To compare the differences in the statements of similar events at different historical times within the stated benchmark set, this embodiment employs a sliding window mechanism to perform cross-period textual difference comparison on the stated benchmark set. Simply put, the sliding window mechanism involves a variable-length time window sliding across a sequence of time anchor points, each time highlighting several adjacent anchor points for comparison, and performing semantic-level difference calculations on the stated text at different time points within the window.

[0027] In one specific implementation, let's take Company A's statement regarding "project leader stability" as an example. The system uses a three-year window, placing Company A's three relevant statements from 2023, 2024, and 2025 into the same window for cross-period comparison. The 2023 tender document stated, "The company's project leader team has remained unchanged for the past three years, and core members are stable"; the 2024 annual report stated, "Project leader Zhang Moumou left the company in October 2023, and Li Moumou was newly hired to take over in January 2024"; the 2025 tender document stated, "The project leader team has remained stable for the past three years, and core members have remained unchanged." After semantic comparison, it was found that the "no changes" in 2025 and the "departed and taken over" in 2024 clearly do not match, with a calculated difference of 0.58, exceeding the preset threshold of 0.35. Based on this, the system determined that Company A had a discrepancy in its cross-period statement, triggering the digital human's response. No discrepancy was detected in the statements from Companies B and C within the same project, so no trigger was triggered. By using a sliding window to compare across periods, the system can automatically screen for consistency in the bidders' statements. Instead of initiating responses to everyone, it precisely targets those bidders with disagreements and then triggers responses accordingly. This usually makes the bid awarding process more efficient.

[0028] S103: Obtain the audio and video data stream of the bidder's interaction with the digital human during the customized bidding presentation and Q&A process, and construct the bidder's real-time emotion fluctuation curve based on the voice data and video data in the audio and video data stream.

[0029] In this step, the audio and video data stream includes video and audio data of the bidder during the presentation. The video data includes the bidder's facial expressions, body language, and the trajectory of their gaze over time, while the audio data includes pauses and changes in tone. Based on this data, the bidder's emotions can be determined, and a real-time emotion fluctuation curve can be generated based on the time sequence. This real-time emotion fluctuation curve reflects the bidder's emotional changes and provides quantifiable analytical data. This gives reviewers data to support their judgments, unlike existing technologies that rely entirely on subjective feelings to determine the bidder's integrity.

[0030] In one specific embodiment, an authorized representative of Company A answered system-generated questions via a digital human interface, while the system simultaneously collected video and audio data during the answering process. When answering the question, "Please explain the specifics of the five performance projects submitted during the 2021 bidding process," the system detected in real time that the representative paused for more than two seconds three times while stating the project names, facial key point detection showed slight asymmetrical eyebrow movements, and the gaze trajectory showed frequent deviations from the center of the screen. After fusing these multimodal features, the system generated an emotion fluctuation curve that varied over time, with a noticeable peak occurring during the answering of this question.

[0031] S104: The audio and video data stream is converted into question-and-answer text with timestamps. The question-and-answer text is sliced ​​according to a preset duration, and the semantic similarity difference between each slice and the corresponding time-series reference text is calculated to form a semantic difference time series. Then, the time length difference between the timestamp of each slice and the timestamp of the corresponding reference text is calculated, and a corresponding time-series attenuation coefficient is configured based on the time length difference to perform weighted correction on the semantic difference of each slice, forming a weighted corrected semantic difference time series curve. The time-series attenuation coefficient is negatively correlated with the magnitude of the time length difference.

[0032] In this step, audio and video data are converted into time-stamped question-and-answer text. Automatic speech recognition technology can be used to transcribe the audio sentence by sentence, retaining the start and end times of each sentence or word, forming a text stream with word-level or sentence-level time stamps. Time slicing refers to dividing the answer text into several continuous time segments according to a preset fixed duration (e.g., 1 second) or complete semantic segments. Slice-by-slice calculation involves calculating the semantic similarity between the text content within each time slice and each time-series benchmark text, thus obtaining the semantic difference value corresponding to each slice. These semantic difference values ​​arranged along the time axis constitute a semantic difference time series. Based on this, the semantic difference values ​​of each slice are weighted and corrected by a corresponding time-series decay coefficient, resulting in a corrected semantic difference value sequence arranged along the time axis, i.e., the weighted corrected semantic difference time series curve. This curve is used for point-by-point or interval-by-interval synchronous peak comparison with the emotion fluctuation curve on the same time axis. Semantic similarity difference refers to the degree of semantic consistency between the current defense answer text and a certain historical benchmark statement text; it is a quantified value. The time-series decay coefficient is a weighting coefficient that decreases as the time span of historical statements increases. It is used to control the impact of historical statements from different periods on the current creditworthiness determination. This coefficient is negatively correlated with the time span; the longer the time interval, the smaller the coefficient, and the weaker its influence on the final determination result.

[0033] In one specific embodiment, Company A's authorized representative answered during the digital human defense that "the five performance projects in 2021 were indeed medical equipment supply projects." The system semantically compared this answer with the original statement in Company A's 2021 tender documents, calculating an original semantic difference value of 0.18. Simultaneously, the system also compared this answer with Company A's 2023 annual report statement that "seven medical equipment supply projects were completed," with an original semantic difference value of 0.51. The 2021 baseline is now more than four years old, and the system assigns a time decay factor of 0.15; the 2023 baseline is approximately two years old, and the assigned decay factor is 0.60. The aforementioned decay factors of 0.15 and 0.60 are specific configurations under this example scenario; in actual implementation, other equivalent numerical configurations can be adopted based on the same decay rules. After weighted adjustments, the adjusted difference for the 2021 benchmark is 0.18 × 0.15 = 0.027, and the adjusted difference for the 2023 benchmark is 0.51 × 0.60 = 0.306. The former, due to its longer timeframe, has a much lower contribution to the determination of dishonesty; the latter, being more recent, retains more of its indicative role in determining dishonesty.

[0034] The time decay coefficient is introduced in this step mainly because historical statements from a long time ago may have inaccurate records or the personnel who filled them out may have changed. Therefore, their indicative significance for current breaches of trust should be weaker than that of recent statements.

[0035] S105: Match the synchronous peak intervals of the emotion fluctuation curve and the semantic difference time series curve, and count the peak density within each synchronous peak interval. When the peak density is greater than the density value threshold, determine that the bidder has committed a breach of trust, and output the corresponding quantitative breach of trust score. The quantitative breach of trust score is used to update the bidder's bid-winning risk profile.

[0036] This step involves checking whether the density of peaks in the emotional fluctuation curve and the semantic difference time series curve is too high within the same time interval. If it exceeds the density threshold, it indicates that the bidder has integrity issues. This is because the continuous occurrence of emotional fluctuation inflection points and semantic difference inflection points within the same time period suggests that the bidder's statements are problematic. This problem mainly stems from discrepancies between the statements and the facts. By cross-validating the information from the two dimensions of "inconsistent statements" and "abnormal emotions during the defense," a relatively accurate result can be obtained. In this embodiment, the quantitative dishonesty score can be calculated using the formula: Score = min(100, max(0, (w_P·P / P_threshold + w_S·S_peak+ w_E·E_peak)×100)), where: Score: the final output quantitative dishonesty score, ranging from 0 to 100; w_P: the weighting coefficient of the peak density index, ranging from 0 to 1; P: the actual measured peak density, in %; P_threshold: the preset peak density threshold, in %; w_S: the weighting coefficient of the semantic difference peak, ranging from 0 to 1; S_peak: the weighted and corrected semantic difference peak, dimensionless, ranging from 0 to 1; w_E: the weighting coefficient of the emotional fluctuation peak, ranging from 0 to 1; E_peak: the emotional fluctuation peak, dimensionless, ranging from 0 to 1; weight constraint: w_P + w_S + w_E = 1; min(100, max(0,...)) The purpose of this function is to truncate the calculation result within the parentheses to between 0 and 100, taking 0 for values ​​below 0 and 100 for values ​​above 100, ensuring that the final score falls within the range of 0 to 100.

[0037] During Company A's response process, the system detected that when answering the question "Why were the performance metrics updated from 5 to 7 in the 2023 annual report?", the system divided the effective response time (from the 2nd to the 45th second) into 43 time slices, each 1 second long, and calculated the sentiment fluctuation value and weighted semantic difference value for each slice. Between the 10th and 12th seconds after the response began, the sentiment fluctuation curve reached a significant peak (sentiment fluctuation value reached 0.72, baseline mean was 0.25), and the weighted semantic difference value also peaked in the same time interval (corrected difference value was 0.365, mean was 0.05 in other time periods). The two peaks highly overlapped on the timeline, with overlapping slices N_sync=2, total effective response slices N_total=43, and peak density P=2 / 43×100%≈4.65%. Since the density did not exceed the system's preset density threshold (8%), the system judged it as a mild statement bias and output a quantitative credit score of 35 points (out of 100) according to the scoring formula. This score was written into Company A's benchmarking risk profile as a reference for the benchmarking committee's decision-making.

[0038] This embodiment aggregates bidders' historical statements across different periods into time-series benchmark data, uses a sliding window to screen for discrepancies in statements and triggers targeted digital human-based responses, integrates multimodal behavioral features to construct an emotion fluctuation curve, introduces a time-series decay coefficient to correct semantic differences, and then determines dishonesty through synchronous peak density detection of emotion and semantics. This forms a complete technical solution for verifying bidders' dishonesty in cross-period statements. This solution does not require batch responses to all candidates; it only performs targeted verification on bidders with discrepancies in their statements, which is generally more efficient. Non-verbal behavioral information during the responses is quantified into calculable emotion data, providing a quantitative basis for subjective judgment. The dual verification of semantic bias and emotional abnormalities also helps reduce the misjudgment rate of single-dimensional analysis. The awarding committee can refer to the quantitative dishonesty score output by the system to conduct focused reviews of bidders with a high risk of dishonesty.

[0039] In one embodiment, the above-described construction of a set of statement benchmarks carrying time-series anchors includes: S1011: Based on the scenario of obtaining the historical written statement text, classify and rank the historical written statement text according to its credibility. S1012: The description of the same key business matter is fragmented and a unique segment timestamp is configured for each independent description to form an independent time-series anchor point. Multiple time-series progressive benchmark anchor point records are generated for the same business matter to construct a multi-level time-series statement benchmark set.

[0040] In this embodiment, the credibility of historical written statements retained in different scenarios varies. For example, the credibility of various common scenarios, from highest to lowest, is: regulatory on-site inspection records > enterprise annual public disclosure reports > historical project performance acceptance materials > legal representative statements for bidding, etc. By ranking and marking the credibility of historical written statements, higher-level statements have priority over lower-level statements in subsequent cross-period comparisons; that is, statements with higher credibility are more valuable. In this embodiment, comprehensive statements containing multiple matters are also broken down into several independent descriptive fragments based on key business matters—a process known as fragmentation. Then, multiple benchmark records of the same matter formed at different times and arranged chronologically are formed into a chronologically progressive benchmark anchor record. Thus, through hierarchical classification and fragmentation, historical statements from different sources and times are organized into a clearly structured, hierarchical, and traceable benchmark set.

[0041] Let's continue with Company A. The historical written statements of Company A obtained by the system include the following: (1) The verification record formed by the municipal drug regulatory department during an on-site inspection of Company A in May 2024, which states that "the company's performance qualification rate for medical device supply projects in 2023 was 98.2%"; (2) The 2023 corporate publicity report submitted by Company A to the market supervision department in March 2024, which states that "total operating revenue was RMB 128 million"; (3) The performance acceptance materials of Company A in a medical equipment procurement project in December 2023, which states that "the equipment installation and commissioning were completed and the acceptance was qualified"; (4) The legal representative statement of Company A in this bidding project in January 2025, which states that "there have been no major violations in the past three years".

[0042] The system categorized the above texts according to their level of credibility: on-site inspection records by regulators are ranked first, annual public notices of enterprises are ranked second, performance acceptance materials are ranked third, and declarations by legal representatives of bidders are ranked fourth.

[0043] Regarding the key matter of "corporate performance", the system extracted independent descriptive fragments from the above texts: (1) "The performance qualification rate of medical device supply projects in 2023 was 98.2%" was extracted from the verification record, with a timestamp of May 2024; (2) "Total operating revenue of RMB 128 million in 2023" was extracted from the company's annual report, with a timestamp of March 2024; (3) "Equipment installation and commissioning completed, and acceptance qualified" was extracted from the performance acceptance materials, with a timestamp of December 2023; (4) "12 medical equipment supply projects completed in the past three years" was extracted from the current tender declaration, with a timestamp of January 2025. These four records are arranged in chronological order to form multiple time-series progressive benchmark anchor records for the "corporate performance" matter of Company A.

[0044] In one embodiment, the comparison of the differences in the statements of similar events in the statement benchmark set at different historical times, and the triggering of a customized digital human-based presentation and Q&A process for the bidder when the difference exceeds a preset threshold, includes: S2011: Based on the time-series anchor points, a variable-length sliding window is used to slice the data in the statement benchmark set. The window length of the sliding window is adaptively adjusted according to the type of bidding project and the operating years of the bidder. S2012: Compare the statements of adjacent time anchors within the sliding window to obtain the differences in the statements of the same type of event at different historical times, and adjust the size of the preset threshold based on the time distance between the historical anchor and the current time. The longer the time distance between the historical anchor and the current time, the larger the range of adjustment allowed for the preset threshold. S2013: When the difference exceeds the preset threshold, a special verification instruction is generated based on the bidder's main identifier, triggering a digital human customized presentation and Q&A process for the bidder. The digital human customized presentation and Q&A process is only for bidders with abnormal statements.

[0045] In this embodiment, the size of the sliding window is not fixed but dynamically determined based on the type of bidding project and the bidder's years of operation. Similarly, the preset threshold is not static; the further back in time the current time is from the historical statement, the larger the allowed preset threshold. The system can flexibly configure verification standards based on project type and bidder characteristics. Companies with shorter operating histories have fewer historical statements, so the window is narrower; older historical statements allow for a larger deviation, preventing errors from years ago from being amplified into current suspicions of dishonesty. Furthermore, triggering only bidders with abnormalities avoids the inefficiency of requiring all bidders to respond.

[0046] For example, a medical group launched a procurement project for a batch of imaging equipment, and the system set the window length to five years. Among the three shortlisted candidates, Company A was established in 2010, and the system assigned it a five-year window length; Company B was established in 2020, and the system assigned it a three-year window length; Company C was established in 2022, and the system assigned it a two-year window length.

[0047] When the system performed a cross-period comparison of Company A's "similar project performance" statements, discrepancies were found between the 2021 and 2025 statements. The 2021 statement is now over four years old, and the system assigned a preset threshold of 0.50 to this historical anchor point; the 2023 statement is approximately two years old, and the preset threshold is 0.30. Calculated, the difference between the 2021 and current statements is 0.42, which does not exceed the allowed threshold of 0.50; however, the difference between the 2023 and current statements is 0.38, exceeding the allowed threshold of 0.30. Based on this, the system determined that there was an anomaly in the cross-period statements, linked it to Company A's entity identifier, generated a special verification instruction, and triggered the digital human defense process. However, no anomalies were detected in the bidding companies B and C, and the digital human defense process was not triggered.

[0048] In one embodiment, the aforementioned digital human customized statement question-and-answer process includes: S2111: Extract the key item fragments within the sliding window whose difference exceeds the preset threshold, and generate a hierarchical special verification question sequence based on the abnormal content of the bidder's statement; S2112: Prioritize the abnormal statement types according to their severity. Prioritize pushing core business type issues, then push related supporting inquiry issues to complete the hierarchical and progressive question-and-answer interaction. Bind the original time-series benchmark anchor number to each verification issue to achieve full-link traceability and retention of divergent issues.

[0049] In this embodiment, if the difference in a key matter segment exceeds a preset threshold, it indicates that the segment contains anomalous content. The system will generate corresponding questions and answers for these anomalous statements. These questions and answers are strategically structured and hierarchically ordered, organized into a list of questions at different levels based on the severity and logical relationship of the matter type. For example, core questions are those with high severity, generally reflecting the bidder's integrity, while supporting questions are related to these core questions. The original time-series anchor point traceability number in this embodiment is a unique identifier for each anchor point record, used during verification to trace the origin of the original statement corresponding to the discrepancy. The response in this embodiment is no longer a generic, template-based Q&A, but focuses on the truly contradictory key points in the bidder's statement. The hierarchical, progressive approach, asking about core contradictions first and then supporting details, makes the response process more organized. After the traceability number is bound, each verification question is traceable.

[0050] Returning to Company A's case, the system detected discrepancies in Company A's statements regarding "performance in similar projects" across different periods. The identified point of contention was the "number of projects"—"5" in 2021, "7" in 2023, and "12" in this 2025 bid. The system automatically generated a tiered sequence of verification questions: The first level (the core point of contention): "Please explain the specific names, contract amounts, and client companies for the five medical equipment supply projects that your company submitted in the 2021 bidding." The second layer (related supporting questions): "What is the correspondence between the seven items reported in the 2023 company annual report and the aforementioned five items? What are the two newly added items?" The third layer (cross-validation question): "Of the 12 projects submitted in this tender in 2025, which were added compared to the 7 projects in 2023? Please provide the corresponding contracts or acceptance certificates." The digital human sequentially pushes questions to the authorized representative of Company A. Each question is linked to a corresponding original baseline anchor traceability number; for example, the first-level question is linked to the storage number of the "Performance Status" section in Company A's 2021 tender document. During subsequent review, the awarding committee can directly retrieve the original statement text for verification using the traceability number.

[0051] In one embodiment, the acquisition of audio and video data streams of the bidder's interaction with the digital human during the customized bidding presentation and Q&A process, and the construction of the bidder's real-time emotion fluctuation curve based on the voice and video data in the audio and video data streams, includes: S1031: Synchronously collect audio and video data streams and uniformly configure millisecond-level timestamps, mapping voice pause distribution data, facial micro-expression key point coordinates, and gaze tracking trajectory coordinates to a unified temporal coordinate axis; S1032: Perform weighted enhancement processing on time slices of abnormal fluctuations in multimodal data synchronization, filter out invalid time segments with no response or blank pauses, and retain only multimodal feature data that effectively corresponds to the content of the question and answer, generating a time-aligned, continuous and stable bidder sentiment fluctuation curve.

[0052] In this embodiment, each frame of audio and video data is labeled with a time stamp accurate to the millisecond, ensuring precise alignment of data from different modalities on the same timeline. Then, the data from the three modalities—voice, facial expression, and gaze—are mapped to the same temporal reference frame. To simplify computation and conserve resources, time segments exhibiting abnormal fluctuations across multiple modalities are given a higher analytical weight. For example, speech pause duration, micro-expression displacement, and gaze deviation are Z-score normalized based on an individual baseline, and then weighted and fused according to preset weights to generate an emotion fluctuation curve. Other time segments can be considered invalid time segments, such as pauses in thought, waiting for a question, or gaps in silence. The modal data during these periods does not reflect the response to a specific question and are therefore discarded.

[0053] In one specific embodiment, during the digital human's presentation at Company A, the system simultaneously captured audio and video data using a camera and microphone, with each frame marked with a millisecond-level timestamp. After the digital human pushed the question, "Please explain the specific details of those 5 projects in 2021," the system started recording from second 0. Company A's authorized representative began answering from second 2 and finished answering at second 45. The system marked and removed the time segments from second 0 to second 2 (the time waiting for the response) and from second 45 to 50 (the time after the response was completed) as invalid time sequences.

[0054] During the effective response period (seconds 2 to 45), the system mapped voice pause data, facial micro-expression key point coordinates, and gaze trajectory coordinates onto a unified time axis. In the time slice from second 8 to second 12, abnormal fluctuations occurred simultaneously in the three modalities of voice pause, abnormal micro-expression, and gaze deviation. The system performed weighted enhancement processing on this slice. The final generated emotion fluctuation curve for Company A showed a significant peak in the second 8 to second period.

[0055] In one embodiment, the semantic similarity difference between each slice and the corresponding temporal reference text is calculated to form a semantic difference time series; then, the time length difference between the timestamp of each slice and the timestamp of the corresponding reference text is calculated, and a corresponding temporal decay coefficient is configured based on the time length difference to weight and correct the semantic difference of each slice, forming a weighted and corrected semantic difference time series curve, including: S1041: For each slice carrying an independent timestamp, calculate the original difference in semantic similarity between the corresponding baseline text and the current digital human question-and-answer text. S1042: Match the temporal decay coefficient of the corresponding gradient based on the time length difference of each temporal reference anchor point relative to the current question and answer, and multiply the original difference of each semantic similarity with the corresponding temporal decay coefficient to obtain the weighted corrected semantic difference corresponding to each reference anchor point, and draw the semantic difference temporal curve.

[0056] In this embodiment, without considering the time factor, the semantic difference (raw semantic similarity difference) between the current defense response text and a certain historical benchmark statement text is calculated. The value is between 0 and 1, where 0 represents complete similarity and 1 represents complete irrelevance. Specifically, the semantic similarity difference can be calculated using a pre-trained semantic encoding model to calculate cosine similarity, with a semantic difference value of 1. Cosine similarity; for numerical statements, add structured fact verification, extracting subject, event, value, unit, statistical interval, negation relation, and time. Inconsistent values ​​are directly marked as factual contradictions, and the difference value is set to 1. The time-series decay coefficient can be expressed as α(t) = ε + (1 ε)e λt In this formula, t represents the time span (in years), λ is the decay rate adjustment parameter (preferably 0.2~0.8), and ε is the minimum retention weight (0<ε<0.05). This formula ensures that α=1 at t=0, and the coefficient is never greater than 1, thus allowing the long-term benchmark to retain the minimum reference weight. The larger the time span, the more exponentially the decay coefficient decreases, but it will not decrease to zero.

[0057] Let's take Company A's "project manager stability" as an example again. The system performs semantic comparisons between the authorized representative's response text and three historical benchmark anchor points. In this example, ε=0.01, λ=0.5, and the time-series decay function is α(t)=ε+(1 ε)e λt .

[0058] The first benchmark: The tender document from March 2023 stated that "the company's project management team has remained unchanged for the past three years, and the core members are stable." This is approximately two years ago, so α1 = 0.01 + 0.99 × e^(-0.5 × 2) ≈ 0.374. During this defense, Company A's representative answered, "The team indeed did not change in 2023." The original semantic difference value S1 = 0.12, and the weighted corrected difference value S1′ = 0.12 × 0.374 ≈ 0.045.

[0059] The second benchmark: The company's annual report in May 2024 stated that "Project Manager Zhang left the company in October 2023, and Li was newly hired to take over in January 2024," which is approximately one year ago. Therefore, α2 = 0.01 + 0.99 × e^(-0.5 × 1) ≈ 0.610. During this defense, Company A's representative answered that "the person in charge has not changed in the past three years." The original semantic difference value S2 = 0.73, and the weighted adjusted difference value S2′ = 0.73 × 0.610 ≈ 0.445.

[0060] The third benchmark: The tender document from January 2025 stated that "the project leader team has been stable for the past three years, with no changes in core members." This is approximately six months ago. Therefore, α3 = 0.01 + 0.99 × e^(-0.5 × 0.5) ≈ 0.781. This benchmark is from the same period as the one used in this defense. The original semantic difference value S3 = 0.05, and the weighted correction difference S3′ = 0.05 × 0.781 ≈ 0.039.

[0061] After time-series decay correction, the original difference value of the second benchmark decreased from 0.73 to 0.445, but it is still the largest source of difference among the three benchmarks. This benchmark clearly records the departure in October 2023, which does not match the statement of "no changes in the past three years" in this tender. Moreover, the time is relatively recent, making it the most indicative of the breach of trust.

[0062] Using an exponential decay model for time-series decay can quantify and control the impact of historical statement biases. Recent statements are more indicative of current defaults than long-term statements. The added ε ensures that the long-term benchmark does not completely fail, retaining some reference value for long-term trend analysis.

[0063] As an alternative implementation, a multi-step temporal decay gradient rule can be configured to generate risk scores differentially, specifically including: Based on the time of this digital human defense, a first time span, a second time span, and a third time span are set, and corresponding first weight attenuation coefficient, second attenuation coefficient, and third attenuation coefficient are assigned and configured; wherein, the first time span is shorter than the second time span, the second time span is shorter than the third time span, the first weight attenuation coefficient is greater than the second attenuation coefficient, and the second attenuation coefficient is greater than the third attenuation coefficient. The weighted adjustment difference of all benchmark anchor points under the same business matter is accumulated in segments according to the time gradient to form short-term risk score and medium- and long-term risk score respectively, so as to distinguish the two types of dishonesty risk situations: deliberate falsification of statements by bidders and unintentional omissions in historical reporting.

[0064] The multi-tiered time-series decay gradient rule divides the time span into several intervals, each with a different decay coefficient. Short-term risk scores are calculated based only on the recent (within one year) benchmark, reflecting the reliability of the bidder's recent statements. Medium- and long-term risk scores are calculated based on benchmarks older than one year, reflecting the long-term consistency trend of the bidder's historical statements. Deliberate falsification occurs when a bidder knowingly and intentionally states inaccurate information, typically manifesting as a sudden and significant deviation between recent and historical statements. Unintentional omissions in historical reporting are due to non-deliberate reasons such as inaccurate records, personnel changes, or missing information, typically showing a gradual drift in statements over many years.

[0065] Let's take Company A's "project manager stability" as an example. The system uses the time of this defense (June 2025) as a benchmark and applies a three-tiered attenuation coefficient: First tier (within 1 year): Attenuation coefficient α1 = 0.85; Second tier (1 to 3 years): attenuation coefficient α2 = 0.50; Third tier (3 years or more): Attenuation coefficient α3 = 0.15.

[0066] Company A has four relevant benchmarks for this matter: January 2025 (0.5 years, α=0.85, this bid states "no changes in the past three years"), May 2024 (1.1 years, α=0.50, the company's annual report discloses "Zhang Moumou left the company in October 2023"), March 2023 (2.3 years, α=0.50, the previous bid stated "no changes in the team"), and July 2022 (2.9 years, α=0.50, the previous annual report stated "the team is stable"). There is also a benchmark from June 2021 (4 years, α=0.15), which states "Project leader Zhang Moumou joined the company in 2020".

[0067] The original semantic difference values ​​for each benchmark are calculated as follows: January 2025 0.05, May 2024 0.73, March 2023 0.08, July 2022 0.10, and June 2021 0.04.

[0068] The weighted adjustments are: 0.043, 0.365, 0.040, 0.050, and 0.006. The system adds up the adjustment differences between the first and second tiers, resulting in a short-term risk score of 0.498; and adds up the scores for the third tier, resulting in a medium- to long-term risk score of 0.006.

[0069] The short-term risk score of 0.498 is significantly higher than the medium-to-long-term risk score of 0.006, and the benchmark from May 2024 (approximately one year ago) contributed a major difference of 0.365. This benchmark clearly states that Mr. Zhang left the company in October 2023, directly contradicting the statement in this bid that there have been "no changes in the past three years." Based on this, the system determines that the discrepancy in Company A's statements is a recent deliberate concealment, not an unintentional oversight from many years ago.

[0070] After applying the multi-step decay gradient rule, the system can distinguish between two different types of statement discrepancies: a high short-term risk score but a low medium-to-long-term risk score indicates a sudden and significant deviation between recent and historical statements, which is more likely to be deliberate fabrication; a high and roughly equal short-term and medium-to-long-term risk score indicates a long-term problem with statement consistency, which is more likely to be a systemic reporting oversight or management disorder. It is important to note that the aforementioned continuous decay function scheme and the multi-step decay gradient rule are interchangeable; in practical applications, one should be used at a time, not both simultaneously.

[0071] In one embodiment, the above-mentioned matching of the synchronous peak intervals of the emotion fluctuation curve and the semantic difference time series curve, and the statistical calculation of the peak density within each synchronous peak interval, and when the peak density is greater than the density value threshold, it is determined that the bidder has committed a breach of trust, and the corresponding quantitative breach of trust score is output, including: S1051: Divide the effective response duration into equal-length time slices according to the left-closed and right-open intervals, and count the number of slices N_sync in which the semantic difference peak and the emotional fluctuation peak coincide synchronously within each time slice. Combined with the total number of effective response slices N_total, calculate the peak density according to P=N_sync / N_total×100%. S1052: When the peak density exceeds the density value threshold, it is determined that the bidder has a subjective risk of dishonesty; when only a single dimension has an abnormal peak, it is determined to be a mild statement deviation, and a differentiated quantitative dishonesty score is output according to different abnormal situations.

[0072] In this embodiment, the effective response time is the actual time period used by the bidder to answer questions, excluding pauses and waiting. The semantic difference peak is the local maximum value reached by the semantic similarity difference at a certain point in time or within a certain time period. The emotional fluctuation peak is the local maximum value reached by the emotional fluctuation curve at a certain point in time or within a certain time period. Peak density is the proportion of synchronously overlapping slices to the total number of effective response slices in that time period; synchronous overlap means that the semantic difference peak and the emotional fluctuation peak appear in the same time slice on the time axis.

[0073] During Company A's defense, the system divided the effective response time (seconds 2 to 45, a total of 43 seconds) into 43 slices, each 1 second long: In the interval from the 10th to the 12th second (i.e., time slices [10,11) and [11,12), a total of 2 slices): the semantic difference value reaches its peak (0.365), and the emotion fluctuation curve also reaches its peak (0.72) at the same time, with 2 overlapping slices; Seconds 18 to 19: Emotional peak (0.58), low semantic value (0.05), no overlap; Seconds 28 to 29: Semantic peak (0.22), emotional stability (0.18), no overlap; The other 40 slices: none had peak values.

[0074] N_sync=2, N_total=43, P=2 / 43×100%≈4.65%.

[0075] The density threshold is 8%, which is within the acceptable range. It is judged as a slight deviation and is scored out as 35 points according to the scoring formula (the risk profile is updated to 35 points, and the original 42 points are approximately 40 points after comprehensive evaluation).

[0076] Suppose that another bidder, Company D, also has a statement discrepancy, but the sentiment fluctuation curve does not show synchronous anomalies during the peak period of semantic discrepancy (P=2.3%, not exceeding the threshold), the system will only judge it as "mild statement deviation", give it 35 points, and not trigger a high-risk warning.

[0077] By calculating the degree of synchronization between "semantic deviation" and "emotional abnormality" over time as a comparable density indicator, an objective and quantifiable basis for judging the risk of breach of trust can be obtained. Two-dimensional synchronous abnormalities constitute a high-confidence signal of breach of trust, while single-dimensional abnormalities are only considered as slight deviations. This makes it less likely to misjudge simple emotional tension or misremembering as malicious breach of trust.

[0078] In one embodiment, the above-mentioned bidder verification method based on time-series decay and synchronization detection further includes a risk profile iterative update step: The generated quantitative credit score is written into the integrity dimension field of the bidder's multi-dimensional risk profile to complete the iterative update of the profile data; When the credit score reaches a high-risk level, a special credit risk label is automatically marked in the risk profile, risk warning information is pushed to the benchmarking and quantitative assessment module, and the data of this multimodal verification process and the basis for credit judgment are linked and archived to the bidder's main credit file to achieve permanent traceability and retention.

[0079] A multi-dimensional risk profile is a data model that comprehensively depicts a bidder's situation from several dimensions, including qualifications, credit, performance, and risk. The integrity dimension field is a data field within the profile specifically recording the bidder's stated integrity status. Specific dishonesty risk tags are identifiers used to mark specific acts of dishonesty within the profile. The bid-setting quantitative assessment module is a subsystem within the bid-setting decision system responsible for conducting comprehensive quantitative comparisons of candidates.

[0080] In Company A's case, after the system output a quantitative credit default score of 35, it did the following: It added the 35 points to the "Integrity Dimension" field of Company A's risk profile. Previously, Company A's integrity dimension score was 42 points; after the update, it was adjusted to approximately 40 points based on the weighted historical average. A special credit default risk label, "Performance Statement Credit Default - 202506," was automatically generated in Company A's risk profile. This label included the default event, the default level, and the basis for the judgment. A risk warning was pushed to the benchmarking and quantitative assessment module. When the benchmarking committee opened the benchmarking comparison interface, a red risk indicator appeared on Company A's candidate entry; clicking on it revealed a detailed credit default judgment report. Finally, the complete verification process data was linked and archived in Company A's main credit file, forming a permanently retained credit default verification record.

[0081] After the risk profile is iteratively updated, the results of the bidder's credit verification are no longer a one-time conclusion, but a continuously accumulating and dynamically updated credit record. The risks discovered in this verification will affect the bidder's award evaluation in future projects. High-risk items are automatically marked and early warnings are pushed out, allowing the award committee to obtain key risk information in a timely manner. The permanent archiving of all data throughout the process also provides complete data support for subsequent regulatory audits and credit retrospectives.

[0082] In one embodiment, a pre-targeted risk screening step is included before constructing the stated benchmark set: The bidders' digital footprint, bid text data, and time-series bid behavior data are collected respectively, and corresponding digital footprint anomaly index, text semantic anomaly index, and bid behavior anomaly index are generated. By integrating three types of abnormal indices and combining them with a knowledge graph of corporate equity relationships, the comprehensive risk level of bid rigging and collusion among bidders is calculated. Only bidders whose comprehensive risk level meets the high-risk threshold are included in the aforementioned multimodal verification process for cross-period statements of dishonesty. The text semantic abnormality index is calculated by integrating AI-generated batch bid document recognition features.

[0083] Bidding digital footprint refers to the digital traces left by bidders during the electronic bidding process, including equipment information, network information, and file metadata. Bidding text data refers to the text content of the bidding documents. Time-series bidding behavior data represents the patterns of various operations performed by bidders over time. The digital footprint anomaly index is a quantified risk score that combines various digital footprint anomalies. The text semantic anomaly index is a quantified risk score that combines abnormal features of the bidding document text, including the identification of AI-generated traces. The bidding behavior anomaly index is a quantified risk score that combines anomalies in bidding time-series behavior. The comprehensive risk level of bid rigging is a comprehensive risk assessment result derived by integrating the three anomaly indices and considering corporate relationships. AI-generated batch bidding document identification features are indicators used to determine whether bidding documents are generated in batches by large language models, such as text semantic entropy value and sentence complexity distribution.

[0084] For example, in a medical consumables procurement project, six bidders were shortlisted. Before performing cross-period statement of dishonesty, the system first conducted a preliminary targeted risk screening: Regarding the digital footprint anomaly index, the system collected the digital footprints of six bidders' bid documents. The comparison revealed that Company E and Company F's bid documents shared the same document creator (both "admin"), and their creation times were similar (only 3 minutes apart). Furthermore, Company E's CPU serial number and Company F's hard drive serial number were correlated. The system calculated an anomaly index of 0.82 for Company E and 0.79 for Company F.

[0085] Regarding the text semantic anomaly index, the system performed semantic analysis on the technical proposals of six bidders. Several paragraphs in the proposals from Company E and Company F contained highly similar wording (similarity reaching 0.91). The AI ​​generation detection model analysis showed that the semantic entropy values ​​of the two proposals were 0.31 and 0.33 respectively, and their sentence complexity distributions were also highly consistent, leading to the conclusion that they were highly likely to have been generated by AI in batches. The system assigned a text semantic anomaly index of 0.88 to Company E and 0.85 to Company F.

[0086] Regarding the abnormal bidding behavior index, the system analyzed the submission times of six bidders' bids. Company E and Company F submitted their bids 2 minutes and 18 seconds and 2 minutes and 5 seconds before the deadline, respectively, very close together, and their bids also showed a regular difference. The system assigned Company E an abnormal bidding behavior index of 0.76 and Company F an index of 0.72.

[0087] When calculating the overall risk level, the system weighted and fused three types of abnormal indices (digital footprint weight 0.35, text semantic weight 0.35, and behavioral weight 0.30), and then combined this with the corporate equity relationship knowledge graph (the supervisors of companies E and F are the same person). The calculated overall risk level for company E was 0.82, and for company F it was 0.79, both exceeding the high-risk threshold of 0.70. The overall risk levels of the other four bidders were all below 0.40. The system only sent companies E and F into the multimodal verification process for cross-period statements of dishonesty; the other four were not triggered.

[0088] By implementing targeted risk screening beforehand, the system quickly identifies bidders suspected of bid rigging using multidimensional data before proceeding with the time-consuming multimodal verification. This concentrates verification resources on high-risk targets, significantly improving the overall efficiency of bid evaluation and decision support. The AI-powered batch generation of bid documents and feature recognition is primarily designed to address the new risk of AI-generated bid documents being used in batches after the widespread adoption of large language models.

[0089] In one embodiment, the aforementioned digital footprint of bidding includes at least one of the following: hardware fingerprint of the equipment used to prepare the bid document, network trace of the bid upload, metadata of the electronic bid document, and encryption key number of the pricing software. The hardware fingerprint includes MAC address, CPU serial number, and hard disk serial number; the network trace includes public IP address, local area network IP address, routing information, and DNS resolution records; the document metadata includes document creator, saver, creation time, and last modification time.

[0090] Hardware fingerprints are unique identifiers for computer devices at the hardware level. A MAC address is the physical address of a network device and is globally unique. A CPU serial number is a unique identifier for the central processing unit. A hard drive serial number is a unique identifier for a storage device. Network traces are network path information left during the upload of tender documents. A public IP address is the public network address assigned to the bidder when accessing the internet. A local area network (LAN) IP address is the bidder's private address within its internal network. Routing information is the network path that data packets take from the bidder's end to the receiving end. DNS resolution records are query records left during the Domain Name System (DNS) resolution process. Document metadata is file attribute information recorded in electronic documents. A pricing software encryption key number is a unique authorization key number for the pricing software used by the bidder when compiling the bill of quantities.

[0091] The digital footprints of Company E and Company F are as follows: Company E: MAC address "AA:BB:CC:11:22:33", CPU serial number "BFEBFBFF000906EA", hard drive serial number "S3Z8NB0M123456"; public IP "202.96.128.15", local area network IP "192.168.1.101"; document creator "admin", creation time "2025-06-10 14:23:05"; encryption key "LOCK-2024-8876".

[0092] Company F: MAC address "AA:BB:CC:11:22:34", CPU serial number "BFEBFBFF000906EB", hard drive serial number "S3Z8NB0M123457"; public IP "202.96.128.15", local area network IP "192.168.1.102"; document creator "admin", creation time "2025-06-10 14:26:18"; encryption key "LOCK-2024-8877".

[0093] These digital footprints contain multiple overlaps that form a relatively high-confidence chain of evidence. The same public IP address indicates that the two tender documents were transmitted from the same network exit; the same document creator and similar creation time suggest that the two documents may have been created by the same person using the same or the same batch of equipment; and the consecutively numbered encryption lock numbers indicate that the pricing software licenses used by the two companies came from the same batch.

[0094] In one embodiment, after updating the bidder's risk profile, the process also includes full-process compliance documentation and smart contract auditing steps: S106: Write the key data hash values ​​of the time-series statement benchmark set, digital human question-and-answer audio and video data, multimodal emotion feature data, weighted and corrected semantic difference time-series data, quantitative credit score, and updated bidder multidimensional risk profile data into the blockchain. The original data is stored in the off-chain encrypted storage system, recording the data hash, timestamp, storage address, access permissions, and signature information. S107: Deploy a dedicated smart contract on the blockchain to adapt to the evaluation and separate benchmarking scenario, and solidify the rules for compliance of the number of members of the benchmarking committee, voting ratio rules, conflict of interest avoidance rules, and publicity time limit rules into the smart contract; S108: Automatically verify the compliance of all data on the blockchain through smart contracts. If the verification is abnormal, the benchmarking process will be suspended and a rectification warning will be pushed. After the verification is passed, a compliance audit certificate will be automatically generated. Relying on the immutability of blockchain, a full-chain audit traceability system that connects risk screening, multimodal integrity verification, and benchmarking decision-making will be formed, and a regulatory standard audit report will be automatically generated and pushed to the public resource transaction supervision system.

[0095] Blockchain-based evidence storage involves writing key data into a distributed ledger on the blockchain in the form of hash values, relying on the blockchain's immutability to ensure data authenticity and integrity. Dedicated smart contracts are pre-written, automatically executed code deployed on the blockchain according to the legally mandated procedures for the evaluation and selection process. Compliance rule solidification involves encapsulating legally mandated rules such as the required number of members on the selection committee, voting ratios, conflict of interest avoidance, and public disclosure deadlines into executable logic within smart contracts. A full-chain audit and traceability system ensures that the entire data chain, from screening for bid-rigging risks, cross-period statements, and verification of breaches of trust to the final selection decision, can be traced and verified through the blockchain.

[0096] In the bidding process for Company A's project, after the system completed all verifications and output a credit score, it performed the following evidence storage and auditing operations: Regarding on-chain evidence storage, the system packages and generates hash values ​​from the time-series statement benchmark set of Company A and other candidates, the complete audio and video files of digital human Q&A, the multimodal emotional feature data and emotional fluctuation curves of each bidder, the weighted and corrected semantic difference data of each benchmark anchor point, the quantitative credit score and judgment basis of each bidder, and the updated multidimensional risk profile of each bidder, and writes them into the blockchain.

[0097] In the aspect of automatic verification of smart contracts, after the bid evaluation committee completes voting, the system uploads the voting result data to the blockchain. The dedicated smart contract deployed on the blockchain automatically performs the following verifications: whether the number of members of the bid evaluation committee is an odd number greater than 5 — the current number is 7, which passes; whether the votes obtained by the winning bidder exceed half of the total — Company A gets 4 votes, which passes; whether there is any person who should have recused but failed to do so — the smart contract invokes the equity association knowledge graph for verification, which passes; whether the public announcement time limit for the bid evaluation result meets the requirements — the system sets 5 working days, which passes.

[0098] In the case of abnormal interception, assume that in another project, the voting result of the bid evaluation committee is 3 votes in favor, 3 votes against, and 1 abstention, and no one obtains more than half of the votes. When the smart contract verifies the rule of "obtaining more than half of the votes" and finds that the rule is not satisfied, it automatically suspends the bid evaluation process, and pushes a rectification warning to the project manager stating that "the bid evaluation vote does not reach the statutory majority, please reorganize the voting", and the process will not continue until the re-voting is passed.

[0099] In the aspect of audit traceability, after the project bid evaluation is completed, the system automatically generates an audit report that covers the entire chain of risk screening, multi-modal integrity verification, and bid evaluation decision-making based on the blockchain's chained hash structure, including operation records, verification results, and abnormal handling records of each link, and pushes it to the public resource transaction supervision system for future reference in the regulatory standard format.

[0100] With the blockchain evidence storage, the key data of the entire bid evaluation process can be permanently stored in an untampered manner. If the data of any link is modified, the integrity of the hash chain will be destroyed, which can be detected in time. After the dedicated smart contract is deployed, the statutory rules of the bid evaluation procedure are changed from manual verification to automatic execution. Non-compliant operations will be blocked when they occur, without the need for post-remediation. Once the full-link audit traceability system is established, the regulatory authority can conduct penetrating verification on the entire bid evaluation process at any time.

[0101] In one embodiment, after step S105 and before the blockchain evidence storage step, the method further comprises: S105A: Collect the quantitative scoring data of each member of the bid evaluation committee on the bidders, construct a bid evaluation expert scoring matrix, and perform abnormal monitoring of expert scoring; The abnormal monitoring of bid evaluation expert scoring comprises: (1) Obtain the independent scores of each member of the bid evaluation committee on each bidder in each evaluation dimension (including technical solution, commercial conditions, performance capacity, bid quotation, etc.), and construct an m×n scoring matrix, where m is the number of members of the bid evaluation committee, and n is the number of bidders; (2) Calculate the mean value μ and standard deviation σ of all expert scores under each evaluation dimension, and calculate the standardized score Z=(x-μ) / σ for each score value; (3) Arrange the standardized scores of each expert in ascending order, calculate the first quartile Q1 and the third quartile Q3, and determine the quantile interval IQR = Q3 - Q1; set the outlier judgment boundary as [Q1 - 1.5 × IQR, Q3 + 1.5 × IQR], and judge the original scores corresponding to the standardized scores that exceed the boundary as outlier scores. (4) Calculate the number of abnormal scores and the magnitude of abnormal deviation for each expert, and generate an expert scoring abnormality index; when the number of abnormal scores for an expert exceeds the preset threshold or the magnitude of abnormal deviation exceeds the preset range, mark the expert as an "abnormal scoring expert" and trigger the calibration scoring review process; (5) Cross-correlation analysis of the monitoring results of abnormal expert scoring and the results of the verification of the dishonesty of the bidder: When a bidder is judged to have committed dishonest behavior, and the number of experts in the bid evaluation committee who gave the company abnormally high scores is ≥2, a warning of "suspected collusion in bidding - expert interest association" is generated and pushed to the bid evaluation supervision module.

[0102] Using a Z-score combined with the interquartile range (IQR) statistical discrete model, abnormally high or low expert scores can be objectively identified, eliminating personal bias or human interference. Linking "bidder misconduct" and "abnormally high expert scores" for analysis allows the system to automatically trigger an alert if a dishonest bidder receives abnormally high scores from multiple experts simultaneously. This helps identify situations where experts are being "targeted" or where there is bribery. The system also triggers a review mechanism and automatically pushes the alert to the monitoring module, transforming the bidding decision from relying on subjective feelings to data quantification and double verification.

[0103] Reference Figure 2 The present invention also provides a computer device, the internal structure of which can be as follows: Figure 2 As shown. The computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores bidder information, etc. The network interface is used to communicate with external terminals via a network connection. Furthermore, the computer device may also include input devices and a display screen. When the computer program is executed by the processor, it implements the bidder verification method based on timing decay and synchronization detection described in any of the above embodiments. Those skilled in the art will understand that... Figure 2The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer equipment on which the present application is applied.

[0104] One embodiment of this application also provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the bidder verification method based on timing decay and synchronization detection described in any of the above embodiments. It is understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0105] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in this application and in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-speed SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0106] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0107] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A bidder verification method based on time-series decay and synchronization detection, characterized in that, include: Collect historical written statements retained by bidders in preset scenarios, and mark key matters in the historical written statements with segmented timestamps to construct a set of statement benchmarks with time-series anchors; Compare the differences in the statements of similar events in the statement benchmark set at different historical times, and trigger a digital human-customized presentation and Q&A process for the bidder when the difference exceeds a preset threshold; The audio and video data streams of the bidder's interaction with the digital human during the customized bidding presentation and Q&A process are obtained, and the real-time emotion fluctuation curve of the bidder is constructed based on the voice data and video data in the audio and video data streams. The audio and video data stream is converted into question and answer text with timestamps. The question and answer text is sliced ​​according to a preset duration, and the semantic similarity difference between each slice and the corresponding time-series baseline text is calculated to form a semantic difference time series. Then, the time length difference between the timestamp of each slice and the timestamp of the corresponding reference text is calculated, and a corresponding temporal decay coefficient is configured based on the time length difference to perform weighted correction on the semantic differences of each slice, forming a weighted correction semantic difference temporal curve; the temporal decay coefficient is negatively correlated with the magnitude of the time length difference; The synchronous peak intervals of the emotional fluctuation curve and the semantic difference time series curve are matched, and the peak density within each synchronous peak interval is calculated. When the peak density is greater than the density value threshold, it is determined that the bidder has committed a breach of trust, and the corresponding quantitative breach of trust score is output. The quantitative breach of trust score is used to update the bidder's bid-winning risk profile.

2. The verification method according to claim 1, characterized in that, The construction of the statement benchmark set with temporal anchors includes: Based on the scenario of obtaining the historical written statements, the historical written statements are ranked and sorted according to their credibility. The statements of the same key business matter are fragmented and each independent statement is assigned a unique segment timestamp to form an independent time-series anchor point. Multiple time-series progressive benchmark anchor point records are generated for the same business matter, and a multi-level time-series statement benchmark set is constructed.

3. The verification method according to claim 1, characterized in that, The step of comparing the differences in the statements of similar events in the statement benchmark set at different historical times, and triggering a customized digital human-based presentation and Q&A process for the bidder when the difference exceeds a preset threshold, includes: Based on the time-series anchor points, a variable-length sliding window is used to slice the data in the statement benchmark set. The window length of the sliding window is adaptively adjusted according to the type of bidding project and the bidder's years of operation. The statements of adjacent time-series anchor points within the sliding window are compared to obtain the differences in the statements of similar events at different historical times. The preset threshold is adjusted based on the time distance between the historical anchor point and the current time. The longer the time distance between the historical anchor point and the current time, the larger the range of adjustment allowed for the preset threshold. When the difference exceeds the preset threshold, a special verification instruction is generated based on the bidder's main identifier, triggering a digital human-customized presentation and Q&A process for the bidder. The digital human-customized presentation and Q&A process is only for bidders whose statements are abnormal.

4. The verification method according to claim 3, characterized in that, The digital human customized statement question-and-answer process includes: Extract the key information segments within the sliding window whose differences exceed the preset threshold, and generate a hierarchical special verification question sequence based on the abnormal content of the bidder's statement; The issues are prioritized according to the severity of the abnormality, with core business-related issues being pushed first, followed by related supporting questions. This creates a hierarchical and progressive question-and-answer interaction. Each verification question is then linked to an original time-series benchmark anchor number to achieve full-chain traceability and retention of conflicting issues.

5. The verification method according to claim 1, characterized in that, The process involves calculating the semantic similarity difference between each slice and its corresponding time-series reference text to form a semantic difference time series. Then, the time length difference between the timestamp of each slice and the timestamp of the corresponding reference text is calculated. Based on this time length difference, a corresponding time-series decay coefficient is configured to weight and correct the semantic differences of each slice, resulting in a weighted and corrected semantic difference time series curve. This includes: For each slice carrying an independent timestamp, calculate the original difference in semantic similarity between the corresponding baseline text and the current digital human question-and-answer text; Based on the time length difference between each time-series benchmark anchor point and the current question and answer, the corresponding gradient temporal decay coefficient is matched, and the original difference of each semantic similarity is multiplied by the corresponding temporal decay coefficient to obtain the weighted corrected semantic difference corresponding to each benchmark anchor point, thus forming the semantic difference temporal curve.

6. The verification method according to claim 1, characterized in that, It also includes configuring multi-step temporal decay gradient rules and generating risk scores in a differentiated manner, specifically: Based on the time of this digital human defense, a first time span, a second time span, and a third time span are set, and corresponding first weight attenuation coefficient, second attenuation coefficient, and third attenuation coefficient are assigned and configured; wherein, the first time span is shorter than the second time span, the second time span is shorter than the third time span, the first weight attenuation coefficient is greater than the second attenuation coefficient, and the second attenuation coefficient is greater than the third attenuation coefficient. The weighted adjustment difference of all benchmark anchor points under the same business matter is accumulated in segments according to the time gradient to form short-term risk score and medium- and long-term risk score respectively, so as to distinguish the two types of dishonesty risk situations: deliberate falsification of statements by bidders and unintentional omissions in historical reporting.

7. The verification method according to claim 1, characterized in that, The process involves matching the synchronous peak intervals of the emotion fluctuation curve and the semantic difference time-series curve, calculating the peak density within each synchronous peak interval, and determining that the bidder has committed a breach of trust when the peak density exceeds a density threshold. A corresponding quantitative breach of trust score is then output, including: The effective response duration is evenly divided into time slices of equal length according to the left-closed and right-open intervals. The number of slices N_sync in each time slice where the peak of semantic difference and the peak of emotional fluctuation coincide synchronously are counted. Combined with the total number of effective response slices N_total, the peak density is calculated as P=N_sync / N_total×100%. When the peak density exceeds the density value threshold, the bidder is deemed to have a subjective risk of dishonesty; when only a single dimension shows an abnormal peak, it is determined to be a mild statement deviation, and a differentiated quantitative dishonesty score is output according to different abnormal situations.

8. The verification method according to claim 1, characterized in that, After updating the risk profile of bidders, the process also includes full-process compliance documentation and smart contract auditing steps: The key data hash values ​​of the time-series statement benchmark set, digital human question-and-answer audio and video data, multimodal emotion feature data, weighted and corrected semantic difference time-series data, quantitative credit score, and updated bidder multidimensional risk profile data are written into the blockchain. The original data is stored in the off-chain encrypted storage system, recording the data hash, timestamp, storage address, access permissions, and signature information. Dedicated smart contracts adapted to the evaluation and separate benchmarking scenario are deployed on the blockchain to solidify the compliance rules for the number of members of the benchmarking committee, voting ratio rules, conflict of interest avoidance rules, and public disclosure time limit rules into the smart contracts. The system automatically verifies all data uploaded to the blockchain through smart contracts. When an error occurs, the calibration process is paused and a rectification warning is sent. After the verification is passed, a compliance audit certificate is automatically generated. Relying on the immutability of blockchain, a full-chain audit traceability system is formed that connects risk screening, multimodal integrity verification, and calibration decision-making. The system also automatically generates regulatory audit reports and pushes them to the public resource transaction supervision system.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the bidder verification method based on timing decay and synchronization detection as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the bidder verification method based on timing decay and synchronization detection as described in any one of claims 1 to 8.

Images