An AI-dispute neutral judgment intelligent agent system without interest relations

Abstract

The application belongs to the technical field of cross of artificial intelligence intelligent decision and credible electronic evidence, and discloses an AI dispute neutral judgment intelligent Agent system without interest correlation. The application aims to solve the core problems of existing AI dispute judgment subject interest binding, rule human tampering, judgment logic and AI technology characteristics disconnection, and disposal evidence chain not closed loop. The system is overall deployed on an independent physically isolated server, and comprises a desensitization evidence data read-only call module, a judgment rule bottom solidification encryption module, an AI technology fault special judgment module, a full autonomous non-intervention decision Agent module, a judicial adaptation judgment report generation module, and a judgment result closed loop binding module. Through a full-link mechanism of one-way read-only call desensitization evidence data, hardware-level solidified judgment rules, full-process autonomous judgment without manual intervention, and judgment result and credible time stamp evidence forced binding archiving, the third-party objective neutral judgment of AI technology type disputes is realized. The application can effectively guarantee the judgment neutrality and result accuracy, form a complete and tamper-proof dispute disposal evidence chain, and the judgment result has the basis for judicial acceptance, and is suitable for the disposal needs of consumer-level, enterprise-level and judicial-level multi-scene AI disputes.

Description

Technical Field

[0001] This invention belongs to the cross-technical fields of artificial intelligence intelligent decision-making, trusted computing, neutral judgment of digital disputes, and electronic evidence solidification. Specifically, it involves a dedicated neutral judgment intelligent agent architecture for AI technology faults deployed in an independent, physically isolated environment. Relying on a full-link technical mechanism of one-way read-only retrieval of anonymized evidence data, underlying hardware-solidified judgment rules, fully autonomous decision-making without human intervention, judicial adaptation report generation, and closed-loop evidence storage of judgment results, this system is specifically designed for third-party objective judgment of digital disputes involving AI model output errors, defects in the model's native algorithm, and non-subjective program malfunctions. This system can seamlessly connect with external judicial-grade trusted evidence storage systems, forming a complete technical closed loop of "original data storage - neutral technical judgment - result solidification and archiving." It adapts to the judgment needs of multiple scenarios, including consumer-level AI service disputes, enterprise-level AI interface fault disputes, judicial electronic evidence presentation, and industry compliance supervision, providing core technical support for the construction of a trust system in the AI ​​industry. Background Technology

[0002] With the large-scale commercial deployment of generative AI, intelligent interactive systems, AI automation services, and industrial-grade AI models across industries, user disputes, enterprise service controversies, and compliance regulatory demands arising from technical issues such as inaccurate AI model outputs, program logic vulnerabilities, inherent algorithmic defects, and abnormal operating environments are growing exponentially. AI-related disputes possess unique characteristics such as strong technical attributes, high thresholds for fault diagnosis, ease of tampering with raw data, and blurred boundaries of responsibility. Their adjudication requirements are far higher than those for traditional internet service disputes. However, existing AI dispute adjudication technologies generally follow the framework for handling general internet disputes, exhibiting four unavoidable fundamental technical flaws that fail to meet the core requirements of neutrality, credibility, technological adaptability, and judicial acceptance: 1. Determining the alignment of interests among stakeholders, neutrality lacks structural safeguards. In existing AI-related disputes, the adjudicating bodies are all the AI ​​service operators themselves. This means that the adjudicating body highly overlaps with the service provider and other stakeholders with commercial interests, creating a natural conflict of interest and a tendency to mitigate liability. The adjudication rules, processes, and results are all controlled by the operator's backend, essentially operating on a "self-adjudication" model. This lack of third-party neutrality inherently leads to biased judgments, rendering them unsuitable as valid evidence in judicial proceedings and difficult to gain public trust from users and the industry. 2. The decision rules are editable and tamper-proof, and the algorithm logic has no fixed constraints. In conventional intelligent judgment systems or manual review processes, judgment rules are stored in an editable business server area in the backend. Operators can arbitrarily modify judgment standards, adjust threshold parameters, and interfere with the judgment process according to business needs and liability avoidance requirements. The algorithm logic is opaque, the judgment process leaves no trace, and rule modifications are not recorded. Neutrality relies solely on the operator's business commitment, without any rigid technical guarantees. This makes it extremely easy for problems such as rules being changed frequently, inconsistent judgments for similar cases, and human intervention in judgment results to occur. 3. The judgment logic is out of sync with the characteristics of AI technology, resulting in insufficient accuracy. Existing dispute resolution schemes all follow the logic used for general civil disputes and internet service disputes, failing to establish a technically-based judgment system for specific scenarios such as non-subjective technical vulnerabilities in AI, model algorithm defects, and program malfunctions. This makes it impossible to accurately distinguish between four different liability scenarios: native AI technical failures, external environmental interference, human error, and malicious user complaints. The judgment process relies heavily on subjective human experience or general semantic recognition, failing to make quantitative judgments based on underlying technical data such as model operation logs, program error codes, and abnormal parameters. Consequently, the judgment results are severely out of sync with the operational characteristics of AI technology, and their reasonableness and accuracy fail to meet the requirements for handling technology-related disputes. 4. The judgment process and the evidence preservation system are disconnected, resulting in an incomplete chain of evidence. The existing judgment scheme and front-end data storage system are independent of each other. The judgment basis can be replaced, tampered with, and added afterward, and there is no encrypted linkage or verification mechanism. The judgment result is stored separately on the operation server and is not forcibly bound and solidified with the original evidence data and trusted timestamps. It is extremely easy to be modified, deleted, or overturned afterward, and cannot form a complete, tamper-proof, and traceable chain of evidence for dispute resolution. It has no evidentiary effect in judicial evidence and compliance audit scenarios. The aforementioned technical deficiencies directly result in insufficient credibility, low accuracy, and lack of judicial validity of the existing AI dispute resolution system. This makes it unable to meet the dispute resolution needs under the rapid development of the AI ​​industry and has become a core technical bottleneck restricting the construction of a trust system in the AI ​​industry and the implementation of a diversified digital dispute resolution mechanism. Summary of the Invention

[0003] (a) Purpose of the invention To address the core shortcomings of existing AI dispute resolution technologies, such as the binding of interests between stakeholders, the tampering of rules, insufficient technological adaptability, and the lack of a closed chain of evidence, this invention constructs a purely technical, third-party neutral intelligent agent system for dispute resolution. This system achieves comprehensive technological breakthroughs across five dimensions: deployment architecture, data entry, underlying rules, decision execution, and result archiving. The core objectives of this invention include: 1. By using independent physical deployment and a one-way data link, the interests of the judgment subject, the parties involved in the dispute, and the business operators are completely separated at the architectural level, achieving a truly neutral third-party judgment and eliminating any form of human intervention; 2. The judgment rules are solidified by hardware encryption chips, making the judgment logic unwriteable, unmodifiable, and unbypassable. This ensures the constancy and neutrality of the judgment standard from the bottom layer of the rules, eliminating all non-technical factors related to commercial interests. 3. Build a dedicated AI technology fault judgment logic, and make judgments based on quantitative technical indicators such as model operation logs, error parameters, and deviation rates. This will accurately distinguish multiple responsibility scenarios, adapt to the operating characteristics of AI technology, and improve the accuracy of judgments. 4. Achieve encrypted linkage between the judgment process and the external trusted evidence storage system, with one-way read-only retrieval of judgment basis, mandatory binding of judgment results to evidence storage certificates and return for solidification and archiving, forming a complete and tamper-proof evidence chain for dispute resolution; 5. Generates standardized judgment reports that comply with judicial evidence standards, with built-in anti-counterfeiting and traceability marks, enhancing the legal credibility of the judgment results and filling the technological gap in the industry for a dedicated neutral judgment system for AI-based disputes. (II) Technical Solution A neutral AI-powered dispute resolution agent system, devoid of any conflict of interest, is deployed across an independent, physically isolated server cluster. It establishes only encrypted one-way communication links with commercial operation servers and external trusted evidence storage systems, with no bidirectional data exchange channels or manual intervention ports. The system comprises six core units linked sequentially in a fixed data flow: a module for reading-only access to anonymized evidence storage data, a module for underlying encrypted determination rules, a module specifically for AI-driven fault diagnosis, a fully autonomous, non-intervention-based decision-making agent module, a judicial adaptation determination report generation module, and a closed-loop binding module for determination results. Each module contains dedicated functional sub-units, and communication between modules is unidirectional, achieved only through pre-defined encrypted interfaces, forming a closed-loop autonomous determination link. The specific technical features of each module are as follows: 1. Read-only retrieval module for anonymized and stored evidence data This module is the system's only data access port, and it has built-in data access subunit, desensitization and verification subunit, and data caching subunit. It only establishes an SSL encrypted one-way read-only communication link with an external trusted evidence storage system. - Data access subunit: Only supports one-way flow of evidence storage data into the judgment system. It prohibits the judgment end from modifying, writing back, or deleting evidence storage data. There is no additional data collection channel, eliminating the risk of data tampering from the entry point. - Desensitization and verification subunit: Performs secondary desensitization and verification on the imported evidence data package to ensure that sensitive fields such as user personal identity information, account information, and business operation data have been completely removed, and only non-privacy technical fields related to AI technical faults are retained, including model running logs, error output messages, program abnormal parameters, service node identifiers, and trusted timestamp encoding, thereby avoiding privacy leaks and interference with commercial interests from the data level; - Data caching subunit: Temporarily encrypts and caches the verified and anonymized evidence data. After the judgment is completed, the cached data is automatically cleared and not stored for a long time to avoid the risk of data leakage. 2. The underlying encryption module for determining the rules. It incorporates a built-in rule burning subunit, a hardware encryption chip, and a rule verification subunit, providing the system with a constant and tamper-proof judgment rule base. - Rule Burning Subunit: Adopting the OTP one-time programmable solidified programming mode, the standardized judgment rules for AI technology faults, the judgment threshold for non-subjective errors, and the logic for dividing dispute responsibilities are written into an independent hardware encryption chip and the underlying code area of ​​the algorithm; once the rules are burned, they cannot be modified, bypassed, or replaced in any way, including through the backend, remote, or local, and there is no rule editing interface; - Hardware encryption chip: It adopts a national cryptographic-level security encryption chip, which runs independently of the system's main processor. The rules are stored in the encrypted partition inside the chip, which cannot be exported or read, and can only be called and matched by the judgment module. - Rule verification subunit: Before each rule is called for judgment, the rule integrity hash verification is automatically performed to ensure that the rule has not been tampered with or replaced. The judgment process can only be started after the verification is passed. The rule base is set only for purely technical issues such as AI model defects, program bugs, output logic errors, and abnormal operating environments, completely eliminating non-technical factors such as commercial interests, platform preferences, and liability reduction, thus ensuring the neutrality and objectivity of the judgment from the rule level. 3. AI technology fault diagnosis module It incorporates a feature extraction subunit, a scene classification subunit, and a quantitative matching subunit, forming a core unit specifically for judging AI-based disputes. - Feature extraction subunit: Automatically extracts technical feature fields from de-identified and stored evidence data, including model error codes, runtime abnormal log fragments, program abnormal parameters, output content deviation features, and service node status data, forming a structured judgment feature set; - Scene Classification Subunit: Based on feature sets, the dispute scene categories are automatically identified and accurately divided into four types of liability scenarios: AI native defects, external environmental interference, malicious human operation, and user operation error. Each scenario corresponds to an independent judgment rule branch. - Quantization Matching Subunit: All judgments are based on quantization technical parameters, abandoning subjective judgment logic; for AI native defect scenarios, three core judgment indicators are set: model error code matching degree, log anomaly ratio, and output semantic deviation rate. Each indicator corresponds to a fixed threshold standard, and quantization matching is completed through weighted calculation to output judgment confidence. 4. Fully Autonomous, Intervention-Free Decision-Making Agent Module As the core execution unit of the system, it has built-in task scheduling subunit, autonomous decision-making subunit, and log encryption subunit. It adopts an autonomous closed-loop decision-making algorithm and requires no human intervention throughout the process. - Task scheduling subunit: Automatically receives judgment instructions, and executes tasks in sequence from data retrieval, feature extraction, rule matching, and report generation, without the need for manual triggering or process control; - Autonomous Decision-Making Subunit: Based on quantitative matching results and fixed responsibility division rules, it automatically completes the qualitative analysis of AI technology faults, the determination of responsibility attribution, and the classification of fault levels, and outputs the final judgment conclusion; the system does not set any manual intervention interface, back-end review port, result correction permission, or process termination entry point, and all decisions are executed autonomously by the intelligent agent; - Log encryption subunit: Real-time hash encryption is performed on each operation to leave a trace. The log records include the operation sequence, call rule number, matching parameters, intermediate results, and final conclusion. The log data is tamper-proof, cannot be deleted, and cannot be erased, and is fully traceable. 5. Judicial Compatibility Assessment Report Generation Module It includes built-in document generation sub-units, anti-counterfeiting generation sub-units, and format specification sub-units, automatically generating standardized judgment documents that comply with judicial evidence standards. - Document generation sub-unit: Automatically generates a structured judgment report based on the judgment conclusion. The report content includes the hash fingerprint of the evidence data, the trusted timestamp number, the AI ​​fault technical characteristics, the quantitative judgment logic, the rule matching basis, the final judgment conclusion, and the fault level classification. - Anti-counterfeiting generation sub-unit: Generates a unique anti-counterfeiting hash identifier for each report. This identifier is bidirectionally bound to the hash of the original evidence data and the hash of the judgment process log. It can be verified and traced through the official trusted timestamp platform to eliminate the risk of report forgery and tampering. - Format Specification Sub-unit: Strictly adheres to the electronic litigation document format specifications when outputting reports, supports standard PDF encrypted format, and can be directly used as technical basis for dispute mediation, judicial litigation, and compliance auditing. 6. Judgment result closed-loop binding module It incorporates a data coupling subunit, a one-way feedback subunit, and an archiving trigger subunit to achieve integrated solidification of the judgment result and the original evidence. - Data Coupling Subunit: The generated judgment report, anti-counterfeiting mark, and judgment process log are forcibly encrypted and coupled with the original evidence data hash fingerprint and trusted timestamp to form an inseparable integrated dispute evidence package, and no single item can be modified individually; - One-way backhaul subunit: Through an encrypted one-way link, the bound evidence package is synchronously backhauled to the hardware read-only storage partition of the external trusted evidence storage system. The backhaul process only supports data outflow and does not support reverse data interaction. - Archiving Trigger Subunit: After the evidence package is returned, the hardware read-only lock command of the evidence storage system is automatically triggered. Once the judgment result is archived, it cannot be modified, deleted, or revoked, realizing the integration and permanent solidification of evidence storage data and judgment results, forming a complete and tamper-proof evidence chain for dispute resolution. Corresponding to the above system, the present invention also provides a method for determining neutrality in AI disputes without any conflict of interest, comprising the following automated execution steps: S1. One-way retrieval of desensitized evidence data: When the system receives the dispute judgment instruction, the desensitized evidence data read-only retrieval module retrieves the corresponding AI interactive desensitized evidence data package through a one-way read-only link, performs secondary desensitization verification, and there is no two-way data interaction throughout the process. S2. Fault Feature Extraction and Rule Quantization Matching: The AI ​​technology fault-specific judgment module extracts technical features such as model operation logs, error parameters, and output deviation rates from the evidence data, and performs quantitative matching with the underlying fixed judgment rules to complete the scene classification. S3. Fully autonomous, intervention-free decision-making: The fully autonomous, intervention-free decision-making agent automatically completes the qualitative analysis of AI technology faults, the attribution of responsibility, and the determination of fault level based on the quantitative matching results. There is no human intervention throughout the process, and each step of the operation is encrypted and recorded in real time. S4. Automatic generation of judicial adaptation determination report: The judicial adaptation determination report generation module generates a standardized determination document with a unique anti-counterfeiting mark, marks the evidence traceability mark and quantitative technical determination basis, and complies with the judicial evidence document norms. S5. Closed-loop solidification and archiving of judgment results: The judgment results are forcibly encrypted and bound to the original evidence hash and trusted timestamp to form an integrated evidence package. It is then transmitted back to the external trusted evidence storage system hardware read-only partition for archiving via a one-way link, triggering permanent read-only locking. The entire process log is synchronously encrypted and retained. (III) Beneficial Effects 1. Architectural neutrality ensures complete severing of any vested interests. Through an overall architecture featuring independent physical server deployment, one-way read-only data links, and no human intervention ports, the intervention path of commercial operators is completely cut off from four levels: deployment environment, data entry, underlying rules, and execution process. This achieves truly neutral third-party judgment, fundamentally solving the industry pain point of "self-review and self-judgment" on the platform. The neutrality of the judgment is supported by rigid technology, rather than by commercial promises. 2. Hardware-level rules are fixed, and judgment criteria are constant and uniform. The system uses an OTP one-time programmable hardware encryption chip to solidify the judgment rules. Once written, the rules are permanently unmodifiable, unbypassable, and unreplaceable, completely eliminating the possibility of human adjustment of standards or opaque operations. Similar cases are judged in the same way, and the standards are constant. The consistency and credibility of the judgment results are far higher than those of conventional judgment systems that use software storage and can be edited. 3. Dedicated technology for accurate AI scenario adaptation. A dedicated quantitative judgment logic is built for AI technology failures. The judgment is completed based on pure technical parameters such as error codes, log characteristics, and deviation rates. It accurately distinguishes multiple responsibility scenarios such as AI native defects, external interference, and human operation, completely eliminating subjective evaluation factors. The judgment results are highly adapted to the operating characteristics of AI technology, and the accuracy is significantly better than general dispute judgment solutions. 4. The entire process is a closed-loop evidence preservation system, and the results have a basis for judicial acceptance. The judgment is based on one-way read-only retrieval from a trusted evidence storage system. The entire judgment process is encrypted and traceable. The judgment result is forcibly bound to the original evidence storage hash and trusted timestamp and sent back to the hardware read-only partition for solidification, forming a complete and tamper-proof evidence chain of "original data - judgment process - judgment result". It meets the judicial acceptance standards for electronic evidence and can be directly used as evidence. 5. Fully autonomous agent execution, high efficiency and stability. The entire process is autonomously scheduled and executed by the intelligent agent, without the need for manual review or intervention. The judgment time has been shortened from 24 hours in the traditional manual mode to seconds. There is no subjective error, no emotional bias, and no fatigue threshold. It can support high-concurrency batch dispute judgment. The system's operational stability and processing efficiency are far superior to manual review and semi-automatic judgment modes. 6. Modular and independent deployment, adaptable to all scenarios. The system adopts a modular design, which can be deployed and run independently, or seamlessly integrated with external evidence storage, compensation, and isolation systems. It supports lightweight single-node deployment and high-concurrency cluster deployment, and can be adapted to the usage needs of multiple entities such as consumer-grade AI platforms, enterprise-grade AI service providers, judicial mediation institutions, and industry regulatory departments. It has a wide range of commercial application scenarios and high technology reuse value. Attached Figure Description Figure 1This is a schematic diagram of the overall module architecture and data flow of the AI-powered dispute neutrality determination intelligent agent system for non-interested parties as described in this invention. Detailed Implementation The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. All embodiments are non-limiting applications of the technical solution of the present invention. Equivalent substitutions, detail optimizations, and scenario adaptations made by those skilled in the art based on the principles of the present invention are all within the protection scope of the present invention. Example 1: User-Generated AI Output Error Dispute Resolution Scenario This embodiment applies to a neutral judgment scenario where a dispute arises due to a factual error in the model's output when an ordinary individual user uses a text-generating AI service. The specific execution process is as follows: When users believe that the generative AI content service contains clear factual errors that affect their rights, they can submit a dispute appeal and request a neutral judgment through the platform. The system backend automatically generates a judgment instruction and pushes it to this intelligent agent judgment system. In Phase S1: The de-identified evidence storage data read-only retrieval module receives the judgment instruction and retrieves the corresponding de-identified evidence storage data packet from the external trusted evidence storage system through an SSL encrypted one-way read-only link; the de-identification verification subunit performs a second verification to confirm that sensitive information such as user mobile phone number, account ID, and personal privacy content has been removed from the data packet, and only technical fields such as model output message, running log, error feature parameters, and trusted timestamp encoding are retained. After the verification is passed, it is stored in a temporary encrypted cache. Phase S2: The AI ​​technology fault determination module is activated. The feature extraction subunit extracts structured parameters such as model error features, abnormal fragments in the running log, and semantic deviation rate between the output content and factual standards from the cached data. The scene classification subunit initially determines that it is an AI native defect scenario. The quantization matching subunit calls the underlying solidified rules to perform weighted quantization calculations on three indicators: error code matching degree, log abnormality ratio, and output deviation rate, and obtains a matching confidence of 92%, which meets the judgment threshold of AI native technology fault. S3 Phase: The fully autonomous, non-intervention-free decision-making agent receives the quantitative matching results and, based on the fixed responsibility division rules, autonomously determines that the dispute is due to a fault in the AI's native technology, with the fault level being a general fault. The log encryption subunit performs SHA256 hash encryption and traces for each step of the operation, recording the entire process information, including instruction trigger time, data retrieval time, feature parameters, matching results, and judgment conclusions. The logs are tamper-proof throughout the entire process. S4 stage: The judicial adaptation determination report generation module automatically generates a standardized determination report, which includes the evidence storage hash value, trusted timestamp number, fault technical characteristics, quantitative determination logic, rule basis, and final determination conclusion; the anti-counterfeiting generation subunit generates a unique anti-counterfeiting hash identifier, which is bidirectionally bound to the evidence storage hash and process log; the report output is in encrypted PDF format, which conforms to the judicial evidence document standard. Phase S5: The closed-loop binding module for the judgment result forcibly encrypts and couples the judgment report, anti-counterfeiting label, process log, original evidence hash, and trusted timestamp to form an integrated dispute evidence package. This package is then transmitted back to the hardware read-only storage partition of the external trusted evidence storage system via a one-way encrypted link. The archiving trigger subunit synchronously triggers a hardware read-only lock command, permanently fixing the evidence package and preventing modification. Both users and the platform can verify the judgment report and evidence certificates through a dedicated port. The entire process involves no manual intervention from the operator, ensuring the judgment result is neutral, objective, and possesses complete evidentiary validity. Example 2: Batch Fault Dispute Judgment Scenario for Enterprise-Level AI Interfaces in B2B This embodiment is applied to a scenario where service failure disputes between enterprise customers and AI service providers are batch-determined when a batch of enterprise-level AI open interfaces experience operational anomalies. The specific execution process is as follows: A corporate client experienced widespread call anomalies with its purchased AI open API service, with batch business requests returning incorrect results. The client believed this was a technical fault on the part of the AI ​​service provider and initiated a batch of dispute complaints, requesting neutral judgment on a total of 127 abnormal call records. This system is deployed on an independent third-party server cluster, physically isolated from both clients' business systems. After receiving the batch judgment instruction, the task scheduling subunit starts concurrent judgment scheduling, supporting hundreds of concurrent processing sessions per batch. The de-identified evidence data read-only retrieval module retrieves the de-identified evidence data of 127 call records in batches, performs de-identification verification on each record, and ensures that only technical data such as interface call parameters, returned error codes, and server operation logs are retained, while removing enterprise commercial data and business-sensitive information. The AI-powered fault diagnosis module performs feature extraction and quantitative matching on each record of batch data. Among them, 119 records were matched with error codes and log features corresponding to server-side program bugs, and the output deviation rate met the severe fault threshold. The remaining 8 records were matched with features indicating call failures caused by abnormal enterprise network environment, which belong to the external environment interference scenario. The fully autonomous, non-intervention-based decision-making agent outputs judgment conclusions for each record: 119 records were determined to be due to technical faults inherent to the AI ​​service provider, with the fault level classified as severe; 8 records were determined to be due to external environmental interference and were not the responsibility of the AI ​​service provider. The entire process involved no human review or intervention, with the total time for judging 127 records in a single batch less than 3 seconds, demonstrating a processing efficiency far exceeding that of manual review. Batch judgment reports are generated one by one according to the session dimension, and each report is accompanied by an independent anti-counterfeiting mark. After all judgment results are bound with the corresponding evidence storage data, they are sent back in batches to the read-only partition of the trusted evidence storage system for solidification and archiving. Enterprises and service providers can verify the judgment results in batches through a unified verification port. The judgment standards are unified, the process is transparent, and the results are traceable, providing objective and neutral technical evidence for subsequent compensation negotiations and service rectification. Example 3: Application of Judgment Report Verification in Judicial Mediation Scenarios This embodiment is applied to the scenario of verifying the authenticity and legality of the judgment report issued by this system during the judicial mediation of AI disputes, demonstrating the judicial effect of the system's judgment results: In a judicial mediation process for an AI service dispute, the plaintiff submitted a neutral assessment report generated by the system as core evidence. The defendant objected to the report's authenticity and neutrality. The judicial mediators conducted a full-chain verification of the report through an official verification channel. The first step is to scan the anti-counterfeiting mark on the report, retrieve the corresponding evidence data package, and re-execute the double hash verification to confirm that the evidence data has not been tampered with since its generation and that the data source is the read-only partition of the trusted evidence storage system hardware. The second step is to verify the trusted timestamp encoding, confirm that the timestamp comes from the national legal timekeeping source, is forcibly bound to the data fingerprint, and that the data generation time is authentic and valid. The third step is to retrieve the encrypted logs of the judgment process to verify that the judgment process is executed autonomously by the intelligent agent without any human intervention. The judgment rule call hash is consistent with the hardware-fixed rule hash, proving that the judgment rule has not been tampered with. The fourth step is to verify the deployment qualifications and physical isolation certificate of the judgment system, confirm that the judgment system has no interest relationship with either the plaintiff or the defendant, the deployment environment is independent, the data link is one-way read-only, and it has the attributes of a neutral third party. After full-chain verification, the report was determined to meet the requirements for authenticity and legality of electronic evidence in the "Provisions of the Supreme People's Court on Several Issues Concerning Evidence in Civil Litigation". It can be accepted as valid technical evidence in judicial mediation, providing core credibility support for the rapid resolution of disputes. Example 4: Multimodal AI-based text-based image copyright infringement dispute determination scenario This embodiment is applied to AI-generated image services, specifically in dispute resolution scenarios where users believe the generated image poses a copyright risk or is significantly inconsistent with the prompts. This demonstrates the system's adaptability to multimodal AI scenarios. A user who used the AI-generated image service to generate commercial images believed that the generated images were significantly inconsistent with the prompts and posed a risk of copyright infringement, and filed a dispute complaint requesting a neutral judgment. Our system retrieved the corresponding anonymized and stored evidence data, including the user's prompt text, generated image files, model generation parameters, and generation logs. This data has been anonymized, removing user personal information and commercial identifiers. The feature extraction subunit extracts text prompt word features, image content features, generation parameter features, and log anomaly features, respectively; the quantization matching subunit uses an image-text similarity algorithm to calculate the semantic matching degree between the generated image and the prompt word, and at the same time compares the material call records and anomaly parameters in the generated log. Quantitative matching revealed that the semantic match between the generated image and the prompt words was only 38%, far below the acceptable threshold. Furthermore, the generated logs showed abnormalities in the model's attention mechanism. This was determined to be an output deviation caused by a defect in the AI ​​model's original logic, which falls under the technical responsibility of the AI ​​service provider. The system generates a judgment report containing quantifiable data on image-text matching and log anomaly evidence, along with a unique anti-counterfeiting identifier. This report is then archived and solidified after being bound to the image and text evidence data. This judgment report can be directly used as technical evidence for users applying for service compensation and for intellectual property disputes. The judgment process is objective, the data is quantifiable, and the results are traceable, effectively solving the industry pain points of difficulty and lack of standards in multimodal AI dispute judgment.

Claims

1. A neutral AI-powered dispute resolution agent system with no conflict of interest, deployed entirely on an independent, physically isolated server, and establishing only encrypted one-way communication links with commercial operation servers and external trusted evidence storage systems, characterized in that... This includes a read-only retrieval module for desensitized and stored evidence data, a module for underlying solidification and encryption of judgment rules, a dedicated judgment module for AI technology faults, a fully autonomous and non-intervention-based decision-making agent module, a judicial adaptation judgment report generation module, and a judgment result closed-loop binding module. The desensitized evidence storage data read-only retrieval module is the system's only data access port. It establishes a one-way read-only communication link only with the external trusted evidence storage system. It can only retrieve AI interactive evidence storage data packets that have been desensitized and have removed user privacy information and business operation data. It has no data modification, deletion, or tampering permissions and no additional data collection channels. The underlying encryption module of the judgment rules writes the standardized judgment rules for AI technology faults, the judgment threshold for non-subjective errors, and the logic for dividing dispute responsibilities into an independent hardware encryption chip and the underlying code area of ​​the algorithm. It adopts an unerasable and fixed programming mode. The rule base is only set for technical issues such as AI model defects, program bugs, and output logic errors, and excludes non-technical factors such as commercial interests and platform preferences. The AI ​​technology fault determination module has a built-in AI technology fault feature recognition algorithm. By extracting model operation logs, error output features, abnormal program parameters, and output deviation rate quantification indicators from the evidence data, it accurately identifies the types of non-subjective AI technical faults and distinguishes three judgment scenarios: AI native defects, external environmental interference, and malicious human operation. The fully autonomous, non-intervention-free decision-making agent module is the core execution unit of the system. It adopts an autonomous closed-loop decision-making algorithm to automatically complete the entire process of evidence data retrieval, rule matching, fault determination, and result generation. The system does not set up manual intervention interfaces, backend audit ports, or result correction permissions, and the execution logs are encrypted and stored in real time. The judicial adaptation determination report generation module automatically generates a standardized determination report based on the determination results. The report content includes the hash fingerprint of the evidence storage data, the trusted timestamp number, the AI ​​fault determination basis, the quantitative determination logic, and the final determination conclusion. The report is generated in an encrypted format and comes with a unique anti-counterfeiting mark. The closed-loop binding module of the judgment result will forcibly encrypt and bind the generated judgment report with the hash fingerprint and trusted timestamp of the original evidence storage data, and synchronously send it back to the hardware read-only storage partition of the external trusted evidence storage system for archiving. Once the judgment result is uploaded, it cannot be modified, deleted, or revoked.

2. A method for determining neutrality in AI disputes without conflict of interest, characterized in that, The AI-powered dispute neutrality assessment intelligent agent system for non-interested parties, as described in claim 1, includes the following automated execution steps: S1. One-way retrieval of desensitized evidence data: When the system receives a dispute judgment instruction, it retrieves the corresponding AI interactive desensitized evidence data package through the read-only retrieval module of desensitized evidence data, with no two-way data interaction throughout the process. S2. Fault Feature Extraction and Rule Quantization Matching: The retrieved evidence data is transmitted to the AI ​​technology fault-specific judgment module to extract the technical features of model operation logs, error parameters, and output deviation rate, and then quantitatively match them with the underlying fixed judgment rules. S3. Fully Autonomous and Intervention-Free Decision Making: The fully autonomous and intervention-free decision-making agent automatically completes the identification of AI technology faults and the determination of responsibility based on the matching results, without human intervention throughout the entire process; S4. Automatic generation of judicial adaptation determination report: The judicial adaptation determination report generation module generates standardized determination documents with anti-counterfeiting marks, and marks the evidence traceability mark and quantitative technical determination basis; S5. Closed-loop solidification and archiving of judgment results: After the judgment results are encrypted and bound, they are sent back to the external trusted evidence storage system for hardware read-only partition archiving, and the entire process operation log is synchronously encrypted and retained.

3. The AI-powered dispute neutrality determination intelligent agent system according to claim 1, characterized in that, The de-identified evidence storage data read-only retrieval module adopts an SSL encrypted one-way communication link, which only supports the one-way flow of evidence storage data into the judgment system, and prohibits the judgment end from modifying or writing back the evidence storage data; the de-identification processing scope covers user personal identity information, account information, and business operation data, and only retains non-privacy technology fields related to AI technology failures.

4. The AI-powered dispute neutrality determination intelligent agent system according to claim 1, characterized in that, Once the rules of the underlying encryption module are written, they cannot be modified, bypassed, or replaced by any means, whether in the background, remotely, or locally. The rule base only retains technical judgment factors and does not contain any judgment logic that is driven by commercial interests or exempts the platform from liability.

5. The AI-powered dispute neutrality determination intelligent agent system according to claim 1, characterized in that, The AI ​​technology fault determination module uses quantitative technical parameters as the basis for judgment, abandoning subjective judgment logic; independent judgment thresholds are set for three types of judgment scenarios, including AI native defect scenarios matching model error codes, log abnormal features, and output logic deviations.

6. The AI-powered dispute neutrality determination intelligent agent system according to claim 1, characterized in that, The fully autonomous, non-intervention-based decision-making agent module does not have any manual review, result correction, or process termination entry points; the execution log is stored in real time using hash encryption, and the log records include the timing of each operation, calling rules, and matching parameters, and cannot be tampered with or deleted.

7. The AI-powered dispute neutrality determination intelligent agent system according to claim 1, characterized in that, The judgment report output by the judicial adaptation judgment report generation module conforms to the format specifications of judicial evidence documents, and the anti-counterfeiting mark corresponds one-to-one with the hash fingerprint of the evidence storage data, which can be verified and traced through the official trusted timestamp platform.

8. The method for determining neutrality in AI disputes without conflict of interest as described in claim 2, characterized in that, In step S5, the judgment report is forcibly coupled and bound to the hash fingerprint and trusted timestamp of the original evidence data. The three form an inseparable integrated evidence certificate. After being sent back for archiving, the hardware read-only lock is triggered synchronously, permanently prohibiting modification and deletion operations.

Images