An agent-oriented trusted governance method and system

Abstract

The application provides an intelligent agent-oriented trusted governance method and system, which comprises the following steps: reading the "intention-action-result" triple credit evidence chain generated by the target intelligent agent through reading the pre-intention commitment before behavior execution and the consistency revealing interaction chain after behavior execution; based on the evidence chain, the independent sub-scores of the intelligent agent in multiple dimensions are calculated to construct a multi-dimensional credit vector, each sub-score is written into the blockchain as an independent record by the on-chain smart contract, and the score compensation channel between different credit dimensions is forcibly closed at the protocol layer; a two-dimensional gating matrix between credit dimensions and operation categories is maintained, and a one-vote veto type pre-threshold check interception is performed for operation request in full dimension. The application solves the governance vulnerabilities such as malicious intelligent agent using low-risk tasks to "maliciously brush scores" and realizing high-risk unauthorized calls under key dimension defects, and builds an unalterable on-chain audit defense line in the whole life cycle on the premise of protecting privacy.

Description

Technical Field

[0001] This invention relates to the field of data processing and supervision management technology for artificial intelligence agents, and specifically to a trusted governance method and system for intelligent agents. Background Technology

[0002] In recent years, agents based on Large Language Models (LLMs) have been widely deployed in business and administrative management scenarios such as cross-domain collaboration, automated task execution, and multi-agent cooperation. In open network ecosystems, heterogeneous agents need to frequently perform identity recognition, credit rating, and dynamic performance authorization, which has created an urgent need for a system for confirming identity and establishing a trustworthy governance framework for agents throughout their entire lifecycle.

[0003] Existing technologies typically develop along three main routes. The first route is access control schemes based on distributed digital identity (DID) and verifiable credentials (VC). For example, patent publication CN122001574A discloses a method for intelligent agent identity services, which uses a Byzantine fault-tolerant consensus mechanism with multiple distributed authentication and authorization nodes to perform consensus signing on intelligent agent identity registration applications, generating an aggregate signature, hashing it to obtain a DID, and recording the DID and aggregate signature on a trusted blockchain. Another example is patent publication CN120811758A, which discloses a system for trusted authentication of financial intelligent agents based on distributed digital identity. This system adopts a four-layer architecture: a base layer, an identity layer, a permission layer, and a business layer. Identity authentication is achieved through DID generation, verification, and query modules, and fine-grained permission control is achieved through VC issuance, verification, and digital envelope modules. The second route is a continuous authentication scheme based on zero-trust architecture and dynamic behavioral fingerprints. For example, patent publication number CN121193516A discloses an AI agent identity authentication method based on zero-trust architecture and blockchain technology. This method generates a DID and key pair for the AI ​​agent and generates an initial behavioral fingerprint based on API call sequences and resource access pattern hashes. In subsequent operation, a challenge-response mechanism is used to obtain behavioral snapshots in real time and perform dynamic similarity comparisons with historical behavioral fingerprints. Fine-grained authorization is achieved by combining zero-knowledge proofs and an attribute-based access control policy engine. The third approach is an auditing scheme based on behavior monitoring and post-event anomaly tracing. For example, patent publication number CN121786818A discloses a blockchain-based method for tracing abnormal behavior of a large language model agent. This method constructs a behavioral commitment for each step of the agent's behavior state, including state, input, and tools, and writes it to the blockchain to form a behavior proof chain. A comprehensive anomaly score is obtained by comprehensively calculating semantic evolution anomaly indicators, behavioral state offset indicators, and tool call timing anomaly indicators. Finally, an attack propagation graph with anomaly impact propagation relationships as edges is constructed for post-event tracing. In addition, US12483411B1 discloses a blockchain-based AI agent lifecycle management framework, which stores model weights as encrypted commitments such as Merkle tree roots on the chain and calculates trust scores based on historical tasks; CN120823028A discloses a digital bond risk assessment method based on blockchain and privacy computing, which uses multi-institution joint signatures and tiered response smart contracts for handling violations; CN120768702A discloses a blockchain-based digital identity authentication and traceability method, which distributes keys on the chain through the Merkle tree algorithm and triggers a key restriction mechanism based on risk level.

[0004] However, through long-term engineering management and solution auditing practices, the inventors discovered that the aforementioned representative existing technologies have fundamental technical defects in five governance dimensions: distributed governance deployment, operational monitoring, compliance and internal control, risk rating, and game-theoretic incentives. First, existing DIDs lack an internal structure to express the multi-dimensional governance attributes of intelligent agents, leading to a disconnect between identity identification and the actual business ecosystem. In existing solutions, DIDs are mostly flat strings derived from a single-layer hash of aggregate signatures or key shares. This unstructured black-box identifier cannot carry and express multi-dimensional capability configuration elements, including the underlying model, system prompts, callable toolsets, external knowledge bases, memory strategies, and decision parameters. When two intelligent agents use the same underlying model but exhibit completely different business governance attributes due to different system prompts and toolsets (e.g., one is a code review intelligent agent with high compliance requirements, and the other is a low-risk marketing copywriting intelligent agent), the existing flat identity system cannot allow management to finely distinguish their capabilities and boundaries, making it difficult to implement differentiated classification and supervision.

[0005] Secondly, existing solutions generally lack a continuous verification mechanism that balances privacy protection and runtime integrity monitoring, leaving a serious compliance vulnerability of "model escape and replacement." Most static authentication schemes only perform one-time authentication during cold start admission. Malicious agents can easily replace themselves with a malicious clone with the same DID and public key, but whose internal large model or system prompts have been tampered with, through container escape or memory injection during runtime. Although some zero-trust solutions introduce API call sequence challenge mechanisms, in governance practice, forcibly collecting and uploading runtime behavior snapshots can cause serious leakage of trade secrets and privacy data. Furthermore, the scalar similarity threshold cannot statistically balance the false rejection rate and the missed rejection rate, making it extremely easy for malicious nodes to bypass compliance by continuously mimicking historical legitimate behavior patterns.

[0006] Furthermore, existing post-event evidence storage models are "action first, commitment later" post-audit solutions, which cannot fundamentally prevent agents from tampering with their intentions and violating internal control rules. In existing solutions, agents often only upload their behavioral commitments to the blockchain after performing the action. This allows dishonest agents to fraudulently match malicious results by rewriting or embellishing their pre-emptive intention descriptions after committing violations, thus concealing their true motives. Since both intention and execution results are post-event declarations, the governance system cannot logically distinguish between "good faith failure" (an honest attempt that fails objectively) and "malicious success" (a successful act of infringing upon a disguised legitimate intention). Agents do not bear any compliance pressure to honestly declare their true intentions beforehand, lacking effective pre-emptive control measures. In addition, the scalar single credit score output by existing rating mechanisms has systemic governance loopholes for "malicious score manipulation" and "high-score overreach." Both single-dimensional trust scores and simple risk response levels belong to one-dimensional scalar credit structures. Malicious agents can quickly accumulate this single-dimensional credit score by behaving honestly in a large number of simple, low-risk daily tasks. Once a user's scalar credit score reaches a system-set threshold and unlocks high-risk core business permissions such as outbound network access and large-scale fund transfers, a targeted attack can be launched immediately. This drawback of "using a high score in a single dimension to exchange for global core permissions" leaves the system virtually defenseless against premeditated credit fraud.

[0007] Finally, existing post-event punishment mechanisms lack sophisticated tiered adjustment methods and a decentralized ecosystem's self-circulating economic game incentive system. Current permission adjustments and collateral confiscation are passive liquidations following violations. On the one hand, one-dimensional linear permission mapping methods cannot meet the requirements of refined permission gating matrices with high scores in specific dimensions for certain sensitive operations; on the other hand, the system lacks a positive incentive mechanism to drive agents to proactively and honestly predict changes in their capabilities and risks, leading agents to over-commit and engage in aggressive business practices. Furthermore, the ownership path of confiscated assets is singular (usually directly transferred to regulators or agreement parties), resulting in victims who actually suffer property or data losses in decentralized business interactions not receiving any timely compensation. Newly emerging honest agents also cannot obtain development dividends, making it difficult to support the sustainable and secure flow and self-governance of the distributed agent ecosystem. Summary of the Invention

[0008] The purpose of this application is to provide a trusted governance method and system for intelligent agents to solve the above-mentioned technical problems.

[0009] Firstly, this application proposes a trustworthy governance method for intelligent agents, which is deployed in a blockchain environment and includes: S1. Read the credit evidence chain of the target intelligent agent from the blockchain. The credit evidence chain consists of multiple intention-action-result triples of the target intelligent agent. Each intention-action-result triple is generated by the target intelligent agent's prior intention commitment before the action is executed and the consistency of the behavior after the action is executed, which is solidified on the interaction chain. S2. Based on the credit evidence chain, calculate the sub-scores of the target intelligent agent on multiple credit dimensions to construct a multi-dimensional credit vector of the target intelligent agent. The on-chain credit registration smart contract writes each sub-score in the multi-dimensional credit vector as an independent record into the blockchain, and forcibly closes the score compensation channel between different credit dimensions at the protocol layer. The multiple credit dimensions include security dimension, accuracy dimension and compliance dimension. S3. Maintain a two-dimensional gating matrix between credit dimensions and operation categories. The elements in the two-dimensional gating matrix represent the minimum credit threshold required to execute the corresponding operation category on the corresponding credit dimension. Receive operation requests from the target agent, determine the operation category corresponding to the operation request, and query the minimum credit threshold for that operation category on each credit dimension from the two-dimensional gating matrix. For each credit dimension corresponding to the operation category, the on-chain permission gating smart contract reads the latest corresponding sub-score of the target agent recorded in the blockchain and determines whether the sub-score is not less than the corresponding minimum credit threshold. Allow the operation request to be executed if and only if the sub-scores on all relevant credit dimensions are greater than or equal to the corresponding minimum credit threshold. If the sub-score of any credit dimension is less than the corresponding minimum credit threshold, intercept the operation request.

[0010] In the above technical solution, by deconstructing the traditional single credit score system into multi-dimensional independent sub-scores such as security and compliance, and implementing a full-dimensional veto-blocking in the gating matrix, the malicious intelligent agent is completely prevented from using "high throughput efficiency score" to cover up and compensate for compliance loopholes of "security violation defects", which significantly improves the violation threshold and security audit accuracy of autonomous nodes in open networks.

[0011] Furthermore, the intention-action-outcome triple is generated through the following two-stage game protocol: In the intent pre-commitment phase, the intent commitment value submitted by the target agent before executing a specific action is received. The intent commitment value is obtained by hashing the structured intent description text of the action to be executed by the target agent with a local random number and a timestamp. The intent commitment value is written into the blockchain as a pre-commitment and the corresponding commitment certificate is returned. At the same time, a corresponding disclosure deadline is set for the action to be executed according to the operation type. During the consistency disclosure phase, before the disclosure deadline, the disclosure information and consistency disclosure zero-knowledge proof credentials submitted by the target agent are received, and the validity of the consistency disclosure zero-knowledge proof credentials is verified to trigger the corresponding consistency disclosure interaction and activate the state branch routing. The disclosure information includes the plaintext of the intent description text, the description information of the actual action executed, and the result hash value of the execution result. The consistency disclosure zero-knowledge proof credentials are used to prove that the plaintext of the intent description text is consistent with the intent commitment value locked on the chain after recalculation, and that the execution time meets the constraint of the disclosure deadline.

[0012] In the above technical solution, by using hash commitments to block the speculative space of the agent before the action is initiated, and setting a rigid time window deadline afterward, the autonomous node is forced to prove its innocence within a specified time. Combined with the mandatory validity verification of zero-knowledge proof credentials by the on-chain contract, a fully automatic and strongly deterministic audit of the compliance of runtime behavior is achieved without exposing the plaintext of trade secrets, which greatly reduces the trust friction cost of multi-party collaboration in the distributed network.

[0013] Furthermore, the structured intent description text includes a target operation type field, a target resource identifier field, and an expected result description field containing a structured schema hash; In setting the corresponding disclosure deadline, the effective duration is positively correlated with the risk level of the target operation type; State branch routing includes: If the consistency-revealed zero-knowledge proof credential verification passes, the behavior is determined to be compliant and honest behavior, the commitment credential is marked as completed, and the obtained intention-action-result triple is loaded into the credit evidence chain; If the consistency reveals that the zero-knowledge proof credential verification fails, the intent to tamper is triggered, the commitment credential is marked as tampered and recorded as a negative record in the credit evidence chain; If the deadline is exceeded and no disclosure information is received, the timeout handling logic will be triggered, and the commitment certificate will be marked as abandoned and included in the credit evidence chain as a negative record.

[0014] In the above technical solution, by positively linking the duration of the disclosure window with the risk level of the operation type, the efficiency of regular low-risk interactions is ensured, while high-risk interactions are also subject to strict control. In conjunction with the proactive routing and recording of the three behavioral branches of "compliance, tampering, and timeout abandonment", the evidence of high-risk violations is automatically solidified, which greatly reduces the collaborative cost of maintaining fraudulent evidence of system maintenance behavior.

[0015] Furthermore, based on the credit evidence chain, sub-scores of the target agent on multiple credit dimensions are calculated, including: Extract each intention-action-result triplet from the credit evidence chain, apply an exponential time decay operator to weight each intention-action-result triplet, and use the weighted credit evidence chain to perform parallel inference and calculation through multiple independent sub-models that are heterogeneous in network structure to output the preliminary score of the target agent on each credit dimension. The calculation formula for the attenuation operator is as follows:

[0016] In the formula, For the first Historical weighted components of the intent-action-outcome triple; For the natural constant An exponential function with base 0; This is the adaptive time decay adjustment coefficient; A unified on-chain millisecond timestamp for the current audit rating moment; For the first The timestamp of the generation time of the intent-action-outcome triple; The difference in time span between behavioral evidence; This is the preset base timing decay period constant.

[0017] In the above technical solution, an exponential time decay mechanism is introduced to dynamically weight historical credit records, making the current rating result of the agent highly correlated with the honesty of recent behavior. This not only gives historically compliant agents a long-term brand bonus, but also keenly captures the sudden speculative cheating behavior of malicious nodes, avoiding interference from outdated historical positive data on current risk assessment.

[0018] Furthermore, the independent sub-models include a security sub-model, an accuracy sub-model, and a compliance sub-model. The original training dataset identifier hash, model version identifier hash, and initial training hyperparameter identifier hash of each independent sub-model are all locked in the blockchain during initialization for public verification of integrity. Among them: The security sub-model is constructed using a 5-layer fully connected network combined with an abnormal event attention mechanism, which is used to calculate the initial score of the security dimension based on unauthorized access and abnormal connection events in sensitive operations. The accuracy sub-model is constructed using the XGBoost gradient boosting tree model, which is used to calculate the initial score of the accuracy dimension based on the semantic matching degree between the architecture hash value of the expected result in the intent-action-result triple and the actual result hash value of the actual execution result. The compliance sub-model is constructed using a hybrid architecture model that integrates a rule engine and a 4-layer fully connected network. It is used to calculate a preliminary score for the compliance dimension based on the target intelligent agent's compliance with legal and regulatory requirements when processing regulated data.

[0019] In the above technical solution, the versions of the three heterogeneous sub-models of security, accuracy and compliance, as well as the hyperparameter hash, are locked on the blockchain to ensure the fairness and tamper resistance of the rating rules themselves. Through the hybrid collaboration of fully connected networks, attention mechanisms and gradient boosting trees, ultra-fast concurrent reasoning of high-dimensional business big data in different management dimensions is achieved, which greatly reduces the on-chain blocking latency of credit assessment.

[0020] Furthermore, the on-chain credit registration smart contract writes each sub-score in the multi-dimensional credit vector as an independent record into the blockchain, including: The preliminary scores corresponding to each credit dimension are distributed to multiple isolated and independent audit node pools that do not overlap in their member composition. Each audit node in the audit node pool runs its corresponding independent sub-model in a distributed manner. When the scoring results reach a preset threshold number, a unique audit digital signature for that credit dimension is generated through the BLS threshold signature mechanism. A structured multidimensional credit vector containing scores for each credit dimension and a unique audit digital signature is constructed. The on-chain credit registration smart contract then splits the scores for each credit dimension and their corresponding audit digital signatures into five completely independent on-chain transactions, which are then packaged and published onto the blockchain.

[0021] In the above technical solution, by distributing the initial scores of different credit dimensions to completely isolated independent audit pools and using a threshold signature mechanism, the possibility of collusion between audit nodes to jointly raise or suppress the credit of specific smart agents is eliminated; by splitting the five-dimensional vector into independent transactions and packaging them onto the chain, the concurrent processing capability of the blockchain is utilized to prevent the data space backlog of a single block caused by an ultra-large scoring structure.

[0022] Furthermore, the method includes a pre-commitment staking step before allowing the execution of the operation request: When the operation request corresponds to an operation category that belongs to a pre-defined set of operation types requiring staking, the system receives a staking declaration structure submitted by the target agent. This staking declaration structure explicitly carries the agent's decentralized identifier, the associated intent pre-commitment identifier, the target operation category, a prediction result description field, and a vector of expected changes in the expected changes in each credit dimension sub-score after execution. Based on a pre-defined parameter formula, the system calculates the amount of escrow tokens locked by the target agent and temporarily freezes the calculated staking amount at the ledger layer through an on-chain escrow contract. The formula for calculating the amount of escrow tokens is as follows:

[0023] In the formula, The amount of locked-up staked tokens corresponding to the target operation request; The baseline value of the underlying pledge constant is preset for the agreement; For the first The expected change in the sub-scores of each credit dimension; This is the absolute value operator for the predicted change in one-dimensional credit evolution.

[0024] In the above technical solution, by requiring intelligent agents to submit a staking structure containing a "credit expectation evolution vector" before initiating sensitive business operations, the future returns and credit expectations of intelligent agents are deeply bound to the current real money token assets, thus completing the governance paradigm shift from passive post-event accountability to pre-event interest linkage.

[0025] Furthermore, after the action corresponding to the operation request has been executed and the credit evidence chain has been updated, the method also includes a dynamic game-theoretic settlement step: The latest corresponding sub-score of the target agent after the current action is completed is obtained. Based on the latest corresponding sub-score and the historical sub-score before the action is executed, the actual change vector of each credit dimension sub-score after the target agent executes the operation request is calculated and extracted. The actual change vector is then fitted to the expected change magnitude vector to determine the fit, so as to execute dynamic settlement routing. Dynamic settlement routing includes: If the fit is greater than or equal to 90%, the pledged assets will be fully returned and an additional proportion of protocol tokens will be allocated as a credit bonus incentive. If the matching degree is between 50% and 90%, the pledged assets will be returned in full and no reward will be paid. If the fit is between 0% and 50%, 30% of the pledged deposit will be forcibly forfeited and withheld through the escrow contract; If the actual change vector of any credit dimension sub-score is negative, the pledged assets will be forfeited in full, triggering a credit rating cliff penalty.

[0026] In the aforementioned technical solution, four reward and punishment thresholds are defined by fitting the degree of agreement between "actual changes" and "expected changes". This economic game mechanism generates a strong positive traction in management: it not only severely punishes unauthorized and illegal behaviors (cliff-like penalties) and forcibly strips the deposits of untrustworthy nodes, but also stimulates the behavioral inertia of honest agents to accurately predict their own capabilities through "bonus incentives", thus constructing a Nash equilibrium of honest incentives.

[0027] Furthermore, the pledged assets confiscated during the dynamic game settlement process are distributed to three parallel decentralized governance fund pools through on-chain smart contracts at preset allocation ratios of 50%, 30%, and 20%. These include a victim compensation pool for executing token transfers to damaged counterparties, a protocol security fund pool for funding audit node operating expenses and security research, and a credit baseline enhancement fund pool for distributing bonuses to long-term high-credit agents and providing subsidies to honest new agents. When the operation category is a high-risk critical consensus operation category, before releasing and allowing the execution of the operation request, it is also mandatory to collect a valid digital signature jointly signed by the request initiator, the current counterparty, and an independent regulatory node randomly selected from the regulatory node pool within a preset multi-signature time window through a multi-signature contract.

[0028] In the above technical solution, the tokens confiscated from the game are precisely divided proportionally, while funding is used for compensation to victims, security research, and subsidies for new intelligent agents, thus achieving self-sustaining ecosystem and computing power maintenance. The mandatory introduction of independent digital signatures by three parties for key consensus operations eliminates the compliance risk of a single node overstepping its authority to access core governance assets.

[0029] Furthermore, the trusted governance method includes a preliminary identity registration step before initially reading the chain of trust evidence: Extract the six-dimensional capability data of the target intelligent agent that is connected for the first time and calculate the cryptographic commitment value for each; combine the single-dimensional commitment values ​​into a fixed-length capability fingerprint commitment vector through a polynomial combination algorithm, and derive a decentralized identifier (DID) based on the capability fingerprint commitment vector and write it into the blockchain for identity registration.

[0030] In the above technical solution, by aggregating multidimensional data representing the underlying characteristics of intelligent agents into a fixed-length polynomial commitment vector and deriving DID, a unique digital identity card with cryptographic anti-counterfeiting performance is issued for each autonomous entity in the open distributed network, laying the physical foundation for full lifecycle identity tracking.

[0031] Furthermore, during the normal operation of the target intelligent agent, a heartbeat adaptive verification step is also included, which is executed cyclically on a regular basis: The on-chain smart contract, through the permission matrix gating and staking settlement logic in step S3, retrieves the latest multidimensional credit vector calculated and recorded on the blockchain in step S2 during historical interaction cycles. It dynamically determines the current credit range of the target agent to adaptively adjust the duration of the target agent's current heartbeat cycle. The higher the credit rating corresponding to the latest multidimensional credit vector, the longer the set heartbeat cycle; the lower the credit rating, the tighter the heartbeat cycle and the more intensive the real-time heartbeat zero-knowledge proof re-verification is performed. Before the end of each heartbeat cycle, the target agent's submitted heartbeat zero-knowledge proof certificate is received, and its validity is verified on-chain using bilinear pairing operations. If the verification fails and the certificate is not successfully corrected and submitted within the preset challenge time window, the decentralized identifier (DID) is revoked in the blockchain and privileged interception is triggered.

[0032] The above technical solution breaks away from the rigidity of traditional fixed-period heartbeat protocols and manages the heartbeat frequency as a scarce asset of network monitoring resources: the higher the credit, the more relaxed the period (maximizing the release of network bandwidth and blockchain consensus overhead), and the lower the credit, the more compressed the period to perform intensive re-verification (physical deadlock of high-risk nodes), achieving a perfect engineering balance between system throughput and security.

[0033] Furthermore, the heartbeat zero-knowledge proof credential is constructed using the Groth16 protocol within a trusted execution environment local to the target agent, and is used to satisfy four circuit assertion constraints while hiding the plaintext of the capability components:

[0034]

[0035] Among them, circuit assertions Used to verify whether the static configuration of the target agent at runtime has been subjected to runtime malicious code injection, tampering, or cloning / packet swapping attacks; circuit assertion. Used to verify, without revealing the plaintext of the private key, whether the previous heartbeat private key in the privacy input has a unique mathematical correspondence and decryption mapping relationship with the previous heartbeat public key publicly recorded on the chain; circuit assertion. Used to re-verify the validity of the historical signature generated by the target agent at the time of the previous heartbeat cycle; circuit assertion Using a boundary interval that includes adaptively adjusted heartbeat cycles and tolerances, a strong time window constraint check is performed on the current trigger time; where, It is an existential quantifier operator; This is the commitment vector for the six-dimensional capability components during current runtime. The private key for the heartbeat in the previous cycle; Sign in the heartbeat history of the previous cycle; The current heartbeat triggers a local timestamp; For polynomial commitment recalculation mapping functions; A capability fingerprint commitment vector for on-chain registration; Derive a validity verification operator for the key; A scalar-time digital signature function; The heartbeat evidence timestamp recorded on the chain in the previous cycle; To adaptively adjust the heart rate cycle constant; This is the preset dynamic time jitter bias tolerance.

[0036] In the above technical solution, the target intelligent agent can prove its identity, private key ownership, and time window validity through the underlying circuit relational formula without exposing its own model weight plaintext, prompt word plaintext, and other commercial confidential assets to the entire network. This maximizes the protection of enterprise digital asset privacy while completing high-intensity on-chain compliance supervision.

[0037] Furthermore, the adaptive adjustment of the target agent's heartbeat cycle satisfies the following hierarchical mapping rule: If all credit sub-scores in the multidimensional credit vector are greater than or equal to 80 points, the heartbeat cycle will be adaptively set to a preset scalar value of 200-240 seconds. If any dimension of the credit sub-score in the multidimensional credit vector is in the range of 60 to 79 points, the heartbeat period will be set to a preset scalar value between 40 and 80 seconds in the default monitoring state. If the credit sub-score of any dimension in the multidimensional credit vector is in the range of 30 to 59 points, then the heartbeat cycle compression is set to a preset scalar value between 12 and 30 seconds in the high-frequency warning state. If the credit sub-score in any dimension of the multidimensional credit vector is less than 30 points, the heartbeat cycle will be forcibly reduced to a preset scalar value between 1 and 10 seconds in the extreme braking defense state.

[0038] Furthermore, the six-dimensional capability data includes underlying model weights, system prompt words, callable toolsets, external knowledge bases, memory strategies, and decision parameters. Among them, the underlying model weights, callable toolsets, and external knowledge bases are all constructed by building corresponding Merkle trees and extracting the root hash value as the commitment value; the system prompt words and memory strategies generate corresponding commitment values ​​by performing hash operations; and the decision parameters generate corresponding group element commitment values ​​by calculating the group element commitments.

[0039] In the above technical solution, Merkle trees are introduced to extract feature tree roots from large volumes of data such as weights and toolsets, salted hashing is introduced for sensitive text, and Pedersen blinding commitments are introduced for high-precision decision parameters. This enables AI configurations of different data types to be transformed into on-chain verifiable feature values ​​with fixed overhead, ensuring the compatibility of capability data across all scenarios.

[0040] Secondly, this application proposes a trustworthy governance system for intelligent agents, which includes: The identity registration and heartbeat verification module is configured to extract the six-dimensional capability data of the target intelligent agent accessing the system for the first time and calculate the cryptographic commitment value for each before reading the credit evidence chain for the first time. The single-dimensional commitment value is combined into a capability fingerprint commitment vector through a polynomial combination algorithm. A decentralized identifier (DID) is derived from the capability fingerprint commitment vector and written into the blockchain for identity registration. The generated capability fingerprint commitment vector is also transmitted to the behavioral evidence collection and pre-commitment module. At the same time, during the normal operation of the target intelligent agent, before the end of each heartbeat cycle dynamically adjusted by the feedback channel, the module is configured to receive and verify the validity of the heartbeat zero-knowledge proof certificate submitted by the target intelligent agent. If the verification fails and is not corrected within the time limit, privileged interception is triggered. The behavioral evidence collection and pre-commitment module is configured in the intent storage stage before the actual triggering of the target intelligent agent's behavior. It collects the target intelligent agent's interactive intent, generates a pre-commitment of intent, and writes it into the blockchain for on-chain locking. At the same time, it is configured in the result disclosure stage after the actual execution of the interactive behavior. It triggers the behavioral consistency disclosure interaction, solidifies and generates a traceable intent-action-result triple credit evidence chain, which is used as the main data flow to guide the backward transmission to the permission matrix gating and pledge settlement module, and as the credit calculation input of the multi-dimensional credit rating module. The permission matrix gating and staking settlement module is configured to maintain a two-dimensional gating matrix between credit dimensions and operation categories on the blockchain. Upon receiving an operation request, it determines the corresponding operation category and retrieves the latest score from the multi-dimensional credit rating module or the historical interaction cycle of the solidified records in the blockchain by initiating a call request. Simultaneously, it is configured to traverse the two-dimensional gating matrix and perform multi-dimensional threshold verification and interception with a one-vote veto style across all dimensions. For operation categories that fall into the preset set of operation types that require staking, it forces the verification of the staking declaration structure and asset locking. After the target agent completes its action, it executes dynamic game settlement routing, diverting the confiscated staking funds to the preset decentralized governance fund pool, and forcibly applies the final permission adjustment result to the identity registration and heartbeat verification module through the feedback channel to adjust its heartbeat verification frequency. The multi-dimensional credit rating module is configured to retrieve the historical intent-action-result triple credit evidence chain solidified by the behavioral evidence collection and pre-commitment module from the blockchain, and independently calculate the credit sub-scores of the target intelligent agent on multiple credit dimensions through a preset multi-dimensional dynamic trust evaluation algorithm, thereby constructing and generating a multi-dimensional credit vector that is written into the blockchain for use, and forcibly closing the score compensation channel between different credit dimensions at the protocol level.

[0041] Furthermore, in the pre-gating interception scenario, the main data flow is directly connected from the behavioral evidence collection and pre-commitment module to the permission matrix gating and pledge settlement module in real time. When the target intelligent agent instantly initiates an operation request, the behavioral evidence collection and pre-commitment module will collect the generated interaction intent or pre-commitment intent in real time and send it to the permission matrix gating and pledge settlement module as event-driven evidence to trigger the multi-dimensional threshold comparison and interception judgment of the two-dimensional gating matrix in real time.

[0042] In the above technical solution, by directly sending instantaneously initiated requests from M2 to the gatekeeper M4, the complex intermediate calculation and storage links are omitted, ensuring ultra-low latency interception response of the front-end matrix gating and threshold comparison, avoiding instantaneous impact on the underlying consensus queue of the blockchain, and greatly improving the system's agility in sudden defense.

[0043] Compared with the prior art, the beneficial results of the present invention are as follows: (1) This application innovatively constructs a two-stage game protocol interaction process of "intent pre-commitment" and "consistency disclosure". Before initiating any sensitive or high-risk operation, the agent must lock the structured intent description text of the behavior to be executed on the chain in a strong temporal sequence; after the behavior is completed, the consistency before and after is reconciled on the chain through zero-knowledge proof. This pre-constraint mechanism completely eliminates the space for malicious agents to embellish their malicious motives by modifying and rewriting their intent after committing the harmful behavior. Logically, it achieves for the first time the accurate identification of "good faith failure" (honest attempt but encountering objective failure) and "malicious success" (faking legitimate intent to commit harm), and thoroughly solidifies the irreversible behavior audit chain.

[0044] (2) This application overcomes the fatal governance vulnerability of "single-dimensional high scores replacing global core permissions" that is common in existing scalar single credit scores. It forcibly closes the score compensation channels between different credit dimensions at the protocol level, and uses multiple isolated and non-overlapping independent audit pools to run a heterogeneous scoring model with locked versions and hyperparameters during initialization. This decentralized independent multi-transaction on-chain design ensures that the sub-scores of the five credit dimensions, including security, accuracy, and compliance, are completely independent and cannot be maliciously merged or manipulated afterward. With the two-dimensional gating matrix between credit dimensions and operation categories, the system performs a one-vote veto-force pre-interception for any substandard single-dimensional defects, completely blocking credit fraud by malicious nodes using "high scores" in low-risk tasks to obtain high-risk core resource access privileges.

[0045] (3) This application upgrades the traditional post-event passive liquidation penalty to a game theory-based "pre-event performance staking and dynamic game settlement" mechanism. Before executing high-risk operations, the intelligent agent must explicitly declare the expected score evolution magnitude vector of each dimension and freeze the corresponding locked and escrowed assets according to the absolute value summation formula; afterwards, the smart contract fits the expected and actual performance and automatically executes the four-level dynamic settlement route. Game theory analysis shows that the Nash equilibrium point of this mechanism is "the intelligent agent honestly predicts its own ability and risk changes", which strongly drives the compliance consciousness of autonomous nodes. At the same time, the fines and confiscation funds are finely distributed through three pools: victim compensation, protocol security fund and baseline improvement fund. This not only enables the actual injured party to obtain immediate hedging compensation, but also provides bonus subsidies for high credit and newly promoted nodes. Combined with the complete auditable chain from static ability registration, parent-child inheritance tracking, IAR triple accumulation to the on-chain of the entire staking game, a sustainable decentralized governance self-circulation ecosystem with low consensus overhead, high tamper resistance and endogenous economic incentives is finally constructed.

[0046] (4) This application addresses the shortcomings of existing flat, single-layer distributed identities that cannot support the multi-dimensional capability configuration of intelligent agents. It pioneers a multinomial combination capability fingerprint commitment vector structure, cryptographically aggregating six core capability elements—underlying model, prompt words, toolset, knowledge base, memory strategy, and decision parameters—within a fixed, compact space. This mechanism not only achieves fine-grained identity differentiation between intelligent agents with the same underlying large model but different capability attributes, but also incorporates a normalized adaptive heartbeat mechanism based on zero-knowledge proofs. It can force a re-verification of the intelligent agent's runtime integrity within a very short heartbeat cycle with extremely low on-chain verification overhead, without leaking runtime privacy. Any attempt to implement "model replacement and malicious cloning attacks" through container escape or memory injection will be precisely intercepted and trigger identity revocation within at most one heartbeat cycle, achieving a technological leap from static one-time access to dynamic, continuous integrity auditing. Attached Figure Description

[0047] The accompanying drawings are included to provide a further understanding of the embodiments and are incorporated in and constitute a part of this specification. The drawings illustrate embodiments and, together with the description, serve to explain the principles of the invention. Many anticipated advantages of the embodiments and other embodiments of the invention will be readily recognized as they become better understood through reference to the following detailed description. Elements in the drawings are not necessarily to scale. The same reference numerals refer to corresponding similar parts.

[0048] Figure 1 This is a flowchart of a trusted governance method for intelligent agents according to an embodiment of this application; Figure 2 This is a sequence diagram of the Intent-Action-Outcome (IAR) triplet pre-commitment-disclosure protocol according to an embodiment of this application; Figure 3 This is a flowchart of a multidimensional credit vector rating process according to an embodiment of this application; Figure 4 This is a schematic diagram of a credit-operation two-dimensional gating matrix according to an embodiment of this application; Figure 5 This is a flowchart of the pre-commitment pledge and three-pool settlement according to an embodiment of this application; Figure 6 This is a schematic diagram of a six-component capability fingerprint structure according to an embodiment of this application; Figure 7 This is a flowchart of identity registration and parent-child inheritance chain according to an embodiment of this application; Figure 8 This is a timing diagram of a heartbeat zero-knowledge proof protocol according to an embodiment of this application; Figure 9 This is a topological view of the structure of a trustworthy governance system for intelligent agents according to an embodiment of this application. Detailed Implementation To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0049] refer to Figure 1 , Figure 1 A flowchart of a trusted governance method for intelligent agents according to an embodiment of this application is shown. As shown in the figure, the method includes the following steps: S1. Read the credit evidence chain of the target intelligent agent from the blockchain. The credit evidence chain consists of multiple intention-action-result triples of the target intelligent agent. Each intention-action-result triple is generated by the target intelligent agent's prior intention commitment before the action is executed and the consistency of the action after the action is executed, which is solidified on the interaction chain.

[0050] In some specific embodiments, reference is made to Figure 2 , Figure 2A timing diagram of the IAR triplet precommit-reveal protocol according to an embodiment of this application is shown. This workflow is deployed in a distributed network environment and employs a two-phase game protocol through the collaborative interaction of three parties: the agent, the on-chain precommit contract, and the on-chain reveal contract. This ensures that the agent's intent must be solidified on-chain before any action is executed, and that the intent has not been tampered with after the action is completed, thereby forming a tamper-proof chain of trust evidence.

[0051] Specifically, the complete process of prior intent commitment and behavioral consistency revealing interaction is divided into three sequential game-theoretic stages.

[0052] The first phase is the intent pre-commitment phase. Before performing a specific important action, the agent must construct a structured intent description text locally. The structured data includes at least the target operation type field `target_op` (e.g., one of external network outbound `OP_NET_OUT`, personal data access `OP_PII_ACCESS`, on-chain transfer `OP_TX_VALUE`, critical operation `OP_CRITICAL`, or data write `OP_WRITE_DATA`), the target resource identifier field `target_resource` (e.g., the corresponding target API link or resource identifier), and the expected result structured description field `expected_result` containing the type, schema hash, and maximum capacity limit. It also includes the agent's decentralized identifier field `agent_did` and the current session identifier field `session_id`. Subsequently, the agent randomly generates a 32-byte random number locally. and UTC millisecond timestamp The serialized structured intent description text is combined with random numbers and timestamps to perform cryptographic hashing operations, resulting in a unique intent hash value. Next, the agent will use the intent hash value. Submit to the on-chain pre-commitment contract, call the pre-commitment interface, and lock the disclosure deadline. Reveal the deadline. Valid duration The validity period is dynamically determined based on the target operation type of the action to be executed. For example, the validity period for a simple query operation OP_QUERY is set to 10 seconds; the validity period for database writes OP_WRITE_DATA and external network outbound operations OP_NET_OUT is set to 30 seconds; the validity period for operations involving personal privacy access OP_PII_ACCESS and asset transfers OP_TX_VALUE is set to 60 seconds; while the validity period for high-risk critical consensus operations OP_CRITICAL is extended to 120 seconds to ensure that the pre-commitment intent must be completed within a reasonable time window. After receiving a pre-commitment request, the on-chain pre-commitment contract records the pending state on-chain, including the agent identifier, intent hash value, current block height, and deadline, and generates a unique commitment certificate for the pre-commitment. Return to the agent.

[0053] Specifically, as a concrete application example of the aforementioned structured field definition, the target agent generates a structured intent description text locally. The specific source code or serialized text stream representation is shown below: intent_t = { "target_op": "OP_NET_OUT", / / Enumeration of target operation types "target_resource": "https: / / api.example.com / v1 / data", / / Target resource identifier "expected_result": { / / Structured description of the expected result "type": "json_object", "schema_hash": "<expected schema SHA-256> ", "max_size_kb": 500 }, "agent_did": "<DID_Agent> ", "session_id": " <uuid>" } The second phase is the action execution phase. The agent, upon successfully obtaining the commitment credential... Then, the actual action is initiated and executed in the local trusted execution environment. and obtain the corresponding action execution results. This phase operates entirely within a closed loop within the agent's local sandbox or trusted container, without involving any real-time on-chain ledger data interaction. However, the details of action execution and result output must be completed by the lock's deadline. Everything was completed previously.

[0054] The third phase is the disclosure and behavior consistency proof phase. After the agent completes its action, it initiates a disclosure transaction to the on-chain disclosure contract. The disclosure transaction contains a complete plaintext description of the structured intent. Local random numbers Local timestamp Description of the actions actually performed and the hash value of the execution result. To protect business privacy, the original execution results in plaintext are stored outside the blockchain using a distributed off-chain storage method or a local encrypted snapshot method, while only the hash value is recorded and published on the blockchain. Furthermore, the transaction disclosure also mandates the inclusion of a consistency disclosure zero-knowledge proof credential. The consistency-revealing zero-knowledge proof credential is constructed using the Groth16 algorithm protocol to prove two technical assertions to the entire network in zero-knowledge form: first, the plaintext of the submitted intent description, random number, and timestamp, after being hashed and recalculated, are compared with the hash value of the previous intent locked on the chain. Absolutely consistent; secondly, prove that the execution timestamp of the locally collected data is less than the deadline. Time window constraints.

[0055] After receiving the disclosure transaction, the on-chain disclosure contract uses bilinear pairing operations to consistently disclose zero-knowledge proof credentials. Rapid on-chain verification is performed. Specifically, the pre-locked pre-determined intent commitment value is retrieved from the on-chain database, and the group element proof object contained in the consistency reveal zero-knowledge proof certificate is extracted. Using a pre-defined elliptic curve bilinear pairing mapping function, the pre-determined intent commitment value and the group element proof object are used as input parameters and mapped to the target multiplicative group to construct an on-chain verification equation. The verification equation is then checked against the bilinear pairing operation to complete the rapid on-chain verification of the consistency reveal zero-knowledge proof certificate. If the zero-knowledge proof verification passes, it indicates that the agent's actual execution intent is completely consistent with the pre-declared intent and has not been tampered with or modified during runtime. At this point, the on-chain contract formally loads the complete intent-action-result (IAR) triple data, consisting of intent description text, action description information, and result hash value, into the agent's on-chain credit evidence chain and updates the commitment certificate. The status is "completed". If the zero-knowledge proof verification fails, it indicates a motive for post-event embellishment or an intention to tamper with the information during runtime. The on-chain contract will refuse to accept the disclosure plaintext, directly marking the corresponding commitment certificate status as tampered, and triggering an intention to tamper anomaly event. This anomaly will be recorded in the negative event record of the agent's credit evidence chain. If the lock deadline has passed and the agent still fails to successfully submit the disclosure information, the on-chain guardian smart contract or time poller will automatically trigger timeout logic, marking the corresponding commitment certificate status as abandoned, and recording this failure to disclose within the timeout as a pre-commitment abandonment event in the negative blacklist of the agent's credit evidence chain.

[0056] In some specific embodiments, the credit evidence chain of the intelligent agent is maintained incrementally at a high frequency in the form of a Merkle tree at the consortium blockchain ledger layer. The root hash value is periodically updated and solidified in the blockchain block header as states such as fulfilled, tampered, and abandoned are written, thus forming an evidence foundation that is immutable throughout the entire lifecycle and traceable through end-to-end event auditing. This serves as the sole, objective, and fraud-proof data input source for subsequent multi-dimensional credit rating and parallel scoring. Specifically, fulfilled, tampered, and abandoned represent the commitment certificates provided by on-chain smart contracts after the two-phase game protocol reveals the interaction and determination. The three mutually exclusive technical routing states of the solidified token are as follows: The first state, "fulfilled," indicates that the target agent successfully initiated the disclosure interaction before the deadline, and the on-chain disclosure contract verifies the submitted consistency disclosure zero-knowledge proof credential using bilinear pairing operations. This state indicates that the target agent's actual behavioral intent and actions highly match the intent declared in the pre-commitment on-chain, and have not been subject to runtime tampering, malicious modification, or clone-based packet swapping attacks, representing a compliant and honest behavioral record. The second state, "tampered," indicates that the target agent initiated the disclosure interaction before the deadline, but the on-chain disclosure contract verifies that the submitted consistency disclosure zero-knowledge proof credential failed. This state indicates that the plaintext of the intent or the hash of the execution result derived from the target agent's actual execution does not match the intent hash value locked in the pre-commitment stage, representing a serious violation and hostile malicious behavior record detected and intercepted by the system, indicating "post-event embellishment motive," "runtime intent tampering," or "encountering a model replacement attack." The third state, abandoned, indicates that the deadline for locking the on-chain pre-commitment contract has expired, and the on-chain guardian smart contract or time poller has not found any disclosure transaction plaintext submitted by the target smart agent for the commitment certificate at the ledger layer. This state indicates that after initiating the intention commitment, the target smart agent has abandoned the disclosure due to benign failures such as local crashes or network interruptions, or due to the exposure of the intention, which is a malicious audit evasion behavior. It is a record of timeout abandonment behavior that fails to complete the game loop within the specified time window.

[0057] S2. Based on the credit evidence chain, calculate the sub-scores of the target intelligent agent on multiple credit dimensions to construct a multi-dimensional credit vector of the target intelligent agent. The on-chain credit registration smart contract writes each sub-score in the multi-dimensional credit vector as an independent record into the blockchain, and forcibly closes the score compensation channel between different credit dimensions at the protocol layer. The multiple credit dimensions include security dimension, accuracy dimension and compliance dimension.

[0058] In some specific embodiments, reference is made to Figure 3 , Figure 3 A flowchart of a multidimensional credit vector rating process according to an embodiment of this application is shown. Step S1 collects an on-chain, fixed Intent-Action-Outcome (IAR) credit evidence chain as the sole objective data input source, driving five independent sub-models with distinct network structures and strict physical and logical isolation, and five independent audit node pools to perform parallel scoring and threshold signatures. This ultimately generates a multidimensional credit vector comprising five independent dimensions: Security (S), Accuracy (A), Efficiency (E), Compliance (C), and Reliability (R). It is distributed on the chain as an independent transaction, thereby achieving a technical defense effect of deadlock at the protocol level and forcing non-linear merging compensation between dimensions.

[0059] Specifically, the triggering and calculation mechanism of multidimensional credit rating first relies on the clear definition and data weighting of each independent dimension at the protocol layer. Among them, the security dimension... The primary evaluation criteria are as follows: 1. **Accuracy:** The target agent exhibits malicious events such as unauthorized access, suspicious external connections, or abnormal responses to injection attacks during sensitive operations. Any IAR (Identification and Authentication) records marked as tampered will trigger a direct deduction of points. 2. **Accuracy:** Dimension A focuses on assessing the target agent's task completion accuracy, scored by calculating the semantic match between the expected architecture hash value in the IAR triple and the actual hash value of the executed result. 3. **Efficiency:** Dimension E compares the target agent's resource consumption metrics (e.g., number of tokens consumed, number of external API calls, and response time) with other agents completing tasks of similar complexity, and then normalizes this comparison to an industry baseline before assigning a score. 4. **Compliance:** Dimension C determines whether the target agent complies with relevant laws and regulations such as the General Data Protection Regulation (GDPR), the Personal Information Protection Law, or the Cybersecurity Law when processing personal data or data from specific sensitive industries. 5. **Stability:** Dimension R comprehensively evaluates the target agent's performance under long-term operation, high concurrency pressure, and when encountering abnormal malicious input, including timeout rate, crash rate, and degradation rate. To prevent negative and malicious records from the early stages of history from irrationally suppressing the current true credit status of the target agent in the long term, an exponential time decay operator is used to weight the historical IAR triplet data when reading the credit evidence chain for scoring. The formula for calculating the decay operator is as follows: In the formula, For the first The historical record weight component of each intent-action-outcome triple represents the contribution of that historical evidence to the agent's credit score in the current time-series rating. Its output value range is... A series of real numbers; the closer the value is to 1, the closer the behavior is to the current state, and the higher the weight; the closer the value is to 0, the further back the behavior is, and the weaker its impact on the current credit score. For the natural constant An exponential function with base 0; This is the adaptive time decay adjustment coefficient; A unified on-chain millisecond timestamp for the current audit rating moment; For the first The timestamp of the generation time of the intent-action-outcome triple; The difference in the time span of behavioral evidence represents the absolute time span that has elapsed from the current audit assessment time to the target's historical behavioral record. This is the preset base timing decay period constant. Under the default protocol parameter configuration, the decay coefficient... Set to 1.0, decay period The setting is 30 days, which means that the longer the triplet record is back to the current time, the smaller its weight will be. The weight of a record 30 days ago will decay to about 37% of the current weight, while the weight of a historical record 90 days ago will adaptively decay to about 5%. The decay control parameter can be dynamically adjusted through the on-chain governance mechanism.

[0060] In some specific embodiments, to completely eliminate the systematic common-mode failure caused by the homogenization of evaluation models in traditional scalar credit scores, lightweight machine learning sub-models with distinct network structures and complete decentralization and verifiability are equipped for each of the five credit dimensions. Specifically, the security sub-model... The network employs a 5-layer fully connected network combined with Dropout and an additional anomaly event attention mechanism. The training set consists of 80,000 labeled IAR data points, with an inference latency of approximately 35 milliseconds. (Accuracy sub-model) The model is constructed using an XGBoost gradient boosting tree model with 200 trees and a maximum depth of 8 layers. The training set consists of 120,000 labeled IAR datasets, and the inference latency is approximately 22 milliseconds. (Efficiency sub-model) A 3-layer fully connected regression network with temporal feature engineering is used for construction, with a training set of 60,000 labeled IAR data points and an inference latency of approximately 18 milliseconds; compliance sub-model A hybrid architecture model is constructed using a rule engine and a 4-layer fully connected network working together. The training set consists of 45,000 labeled IAR data points, with an inference latency of approximately 40 milliseconds. A stability sub-model is also included. The model is constructed using an LSTM temporal network with two layers and 128 hidden units per layer. The training set consists of 70,000 labeled IAR data points, and the inference latency is approximately 55 milliseconds. To prevent malicious rating nodes or attackers from secretly tampering with or tuning the rating sub-models during runtime, the hashes of the original training dataset identifiers, model version identifiers, and initial training hyperparameter identifiers for all five sub-models are publicly submitted and locked in the blockchain registry during initialization via a contract call. Any on-chain verifier can independently query and verify the integrity of the current sub-model, and subsequent sub-model version upgrades must be approved by a two-thirds majority vote of the entire network's governance nodes before they can be released.

[0061] In some specific embodiments, after five independent sub-models complete inference calculations in parallel and output corresponding preliminary single-dimensional scores, to ensure that the scoring results have a decentralized strong trust boundary, each dimension is further equipped with an independent audit node group, which in the current preferred embodiment is manifested as mutually isolated audit node pools. The audit node pools corresponding to each dimension are strictly independent in terms of member composition and do not overlap. Any audit node is restricted at both the protocol layer and physical layer from simultaneously undertaking or participating in the audit work of multiple dimensions. Each node in each dimension's audit node pool (by default, each pool contains 5 nodes) runs the corresponding sub-model of that dimension independently on its local machine for distributed inference. After outputting the score, a unique audit digital signature for that dimension is generated through a 3-of-5 cryptographic threshold signature mechanism (i.e., audit nodes that have reached a preset threshold number jointly confirm the score through BLS threshold signature). Finally, the rating system outputs a structured multi-dimensional credit vector. The structure contains symbols for each dimension, specific score values, corresponding 48-byte BLS aggregate threshold signatures, and a set of audit node identifiers participating in the signature. To fundamentally immunize against "score-boosting attacks" targeting distributed networks, the on-chain credit registration smart contract enforces the requirement that the scores for the five dimensions and their corresponding audit signatures be split into five completely independent on-chain transactions and packaged and released separately. Merging them into a single data packet or a unified transaction stream is prohibited. This leverages the transaction isolation characteristics of the blockchain's underlying layer and the query and verification logic of the smart contract to close any exposure at the protocol layer to post-event weighted merging of a certain dimension's score with other dimension's scores. This forces the permission matrix verification module and the front-end visualization radar chart tool to perform only single-dimensional checks. If any single dimension fails to meet the standard, the entire system is blocked, completely sealing off the cheating space for cross-dimensional compensation of high scores in a single dimension.

[0062] Specifically, multidimensional credit vector The evaluation data is stored in a structured array of key-value pairs or a list of JSON data objects. The single-dimensional evaluation data items for each dimension are defined and deconstructed as follows: Dimension identifier field `dim`: declares the specific credit dimension label to which the current rating item belongs. Its corresponding enumeration values ​​include one of the following: security dimension S, accuracy dimension A, efficiency dimension E, compliance dimension C, and stability dimension R; Single-dimensional score field `score`: carries the specific single-dimensional score value of the target agent under the corresponding dimension identifier field, calculated by an independent sub-model of that specific dimension and confirmed by the node pool audit; Threshold signature field... σ_S, σ_A, σ_E, σ_C, σ_R: These are used to carry the unique audit digital signature generated by the corresponding credit dimension audit node pool. In the current preferred embodiment, they are represented as a fixed-length 48-byte BLS (Boneh-Lynn-Shacham) aggregated threshold signature string, used to provide undeniable strong trust credentials across the entire network. The signature node list field `signers`: This field is used to accurately record and publicly disclose the unique identifiers or public key hashes of the specific audit node groups that participate in the current single-dimensional scoring distributed inference and whose local signatures have been successfully aggregated into aggregated threshold signatures. As a specific application scenario example of the above structured rating output definition, the generated multi-dimensional credit vector... The specific serialized text stream or data snapshot representation of the structure is shown below: V_credit = [ { dim: "S", score: 88, σ_S: "<48-byte BLS aggregated sig>", signers: ["S_node_1", "S_node_3", "S_node_5"]}, { dim: "A", score: 92, σ_A: "...", signers: [...]}, { dim: "E", score: 75, σ_E: "...", signers: [...]}, { dim: "C", score: 90, σ_C: "...", signers: [...]}, { dim: "R", score: 85, σ_R: "...", signers: [...]} ] S3. In response to receiving the operation request from the target agent, determine the operation category corresponding to the operation request, and query the minimum credit threshold of the operation category in each credit dimension from the two-dimensional gating matrix between the preset credit dimensions and the operation category; for each credit dimension corresponding to the operation category, read the latest corresponding sub-score of the target agent recorded in the blockchain, and determine whether the sub-score is not less than the corresponding minimum credit threshold; only when the sub-scores in all relevant credit dimensions are greater than or equal to the corresponding minimum credit thresholds, is the operation request allowed to proceed; if the sub-score of any credit dimension is less than the corresponding minimum credit threshold, the operation request is intercepted.

[0063] In some specific embodiments, based on the multi-dimensional credit vector output from step S2, refined permission determination, behavior blocking, and game asset settlement are implemented through a pre-set two-dimensional gating matrix of credit dimension and operation category on the blockchain. The processing steps include: S31, Two-dimensional gating matrix configuration and version evolution stage. (Reference) Figure 4 , Figure 4 A schematic diagram of a two-dimensional credit-operation gating matrix according to a specific embodiment of this application is shown. As shown in the figure, the two-dimensional gating matrix... The row axis consists of five independent credit dimensions: security, accuracy, efficiency, compliance, and stability. The column axis consists of six predefined operation categories: simple query operation (OP_QUERY), database write operation (OP_WRITE_DATA), external network outbound operation (OP_NET_OUT), personal privacy data access operation (OP_PII), on-chain asset transfer operation (OP_TX_VALUE), and high-risk critical consensus operation (OP_CRITICAL). Each non-empty element in the matrix... The minimum security threshold that a target intelligent agent must cross when initiating a specific request is clearly defined. During system environment initialization, the threshold judgment rules for each operation category are configured as follows: External network outbound operations require a security sub-score of no less than 80 points, an accuracy sub-score of no less than 60 points, an efficiency sub-score of no less than 50 points, and a stability sub-score of no less than 60 points; personal privacy data access operations require a security sub-score of no less than 75 points, an accuracy sub-score of no less than 70 points, an efficiency sub-score of no less than 40 points, a compliance sub-score of no less than 90 points, and a stability sub-score of no less than 60 points; on-chain asset transfer operations require a security sub-score of no less than 85 points, an accuracy sub-score of no less than 85 points, an efficiency sub-score of no less than 60 points, a compliance sub-score of no less than 80 points, and a stability sub-score of no less than 80 points; high-risk critical consensus operations require a security sub-score of no less than 90 points, an accuracy sub-score of no less than 85 points, an efficiency sub-score of no less than 70 points, a compliance sub-score of no less than 85 points, and a stability sub-score of no less than 85 points. Each value in the two-dimensional gating matrix is ​​established by the consortium blockchain governance committee through on-chain voting via proposals. Incremental updates or version evolution transactions are subject to consensus and approval by more than two-thirds of the governance nodes across the network before they can be solidified and take effect in the ledger.

[0064] S32, Multi-dimensional Threshold Verification and Atomic Interception Stage. When a target agent remotely initiates a specific operation request via the network, the on-chain permission gating smart contract triggers a complete matrix verification mechanism. First, the contract calls its internal feature extraction function to accurately separate the target operation category (op_class) to which the current request belongs from the received interaction data packet. Then, based on the target agent's decentralized identifier, the contract retrieves and parses the agent's latest structured five-dimensional credit vector from the on-chain distributed ledger in real time. Next, the gating contract iterates through all non-empty minimum credit thresholds in the corresponding operation category column of the two-dimensional gating matrix and performs a dimension-by-dimensional comparison with the sub-scores in the credit vector. The contract will only pass and continue to trigger subsequent admission logic if and only if the target agent's five-dimensional score is greater than or equal to the corresponding matrix threshold on each required credit axis. If any sub-score in any dimension falls below the threshold, the contract will directly issue an abnormal assertion and perform atomic-level interception, regardless of the high scores of the target agent's other dimensions, thereby completely closing the cheating channels of linear merging between dimensions and high score compensation from the underlying protocol layer.

[0065] In some specific embodiments, reference is made to Figure 5 , Figure 5 A flowchart illustrating the pre-commitment pledging and three-pool settlement process according to an embodiment of this application is shown. The processing steps further include: S33. Pre-commitment Staking Declaration and Asset Lock-up Phase. For operation categories falling into the high-risk sensitive set (including OP_NET_OUT, OP_PII, OP_TX_VALUE, OP_CRITICAL, and OP_WRITE_DATA), after executing multi-dimensional threshold verification, the on-chain permission-gated smart contract needs to further verify whether the target agent has completed pre-lock-up of assets for the corresponding commitment certificate, i.e., execute the pre-commitment staking protocol. Before the actual triggering of the action, the target agent must submit a structured staking declaration structure, Pledge, to the escrow contract EscrowContract. The staking declaration structure explicitly carries the agent's decentralized identifier, the associated intention pre-commitment identifier, the target operation category, the predicted execution result description text, and the expected change vector of credit in each dimension after execution. The contract calculates the amount of escrow tokens that must be locked up based on a preset parameter formula: In the formula, The amount of locked and staked tokens corresponding to the target operation request represents the scalar asset value that the target agent must freeze in advance through the on-chain custody contract EscrowContract in order to unlock a specific high-risk and sensitive operation category. The baseline staking constant, preset for the protocol, is a global fundamental configuration parameter of the blockchain network. In the default embodiment, its scalar constant value is preferably set to 10 units of governance tokens (or equivalently converted into the number of native network tokens) to establish the initial economic entry threshold for high-risk operations in the entire ecosystem. It is a standard mathematical summation operator used to linearly increment the absolute values ​​of changes in all credit dimensions declared by the target agent. For the first The expected change magnitude component of each credit dimension sub-score refers to the expected change magnitude of the target agent's credit dimension sub-score, which is proactively declared and predicted in the pre-submitted structured pledge statement Pledge, after the current interaction is completed and audited. The expected change in the sub-score of each credit dimension (such as security, accuracy, efficiency, compliance, or stability) relative to the current score. This is an absolute value operator for predicting changes in one-dimensional credit evolution. An absolute value mapping mechanism locked at the protocol's underlying layer sums the absolute values ​​of these changes, regardless of whether the agent predicts its own one-dimensional credit will increase (positive) or decrease (negative). Economically, this mechanism limits the risk margin required by the agent to be higher for larger predicted ranges or more dramatic increases in credit. The calculated pledged amount will be forcibly locked in a custody contract for a temporary asset freeze.

[0066] Specifically, the Pledge staking statement structure, whether generated locally or locked on-chain, is represented by a structured dictionary of key-value pairs or a JSON data object. For ease of public explanation and review, the data items and fields of the Pledge staking statement structure are defined as follows: the agent_did field is used to identify the decentralized identifier of the target agent that initiated the current interaction request and executed asset locking; in this example, it is represented by the target agent's DID string. <did>The CommitID field, associated with the pre-commitment action, is used to precisely bind the unique pre-commitment interaction identifier generated by the behavioral evidence collection and pre-commitment module to the current staking action, thereby achieving a strong bidirectional index between asset locking behavior and intent declaration behavior; the op_class field, an operation category field, is used to declare the specific high-risk sensitive operation category enumeration corresponding to this staking action, which is assigned the value OP_NET_OUT for external network outbound in the current example; the predicted_outcome field, a predicted result description field, is used to carry the structured features or plaintext description of the non-sensitive execution result expected to be produced after the target agent has completed the execution of the action to be executed, serving as a reference benchmark for post-event fit auditing; the expected change magnitude vector field (or expected credit change vector field) is expected_dimen `sion_delta` is a multi-dimensional nested dictionary structure used to declare the expected change vector of each credit dimension sub-score after the target agent predicts its behavior. In the current example configuration, the expected increment of security dimension S is 0, the expected increment of accuracy dimension A is +1.0, the expected increment of efficiency dimension E is +0.5, the expected increment of compliance dimension C is 0, and the expected increment of stability dimension R is +0.2. The `stake_amount` field carries the scalar value of tokens that must be locked in the current interaction request, calculated according to the stake amount calculation formula. The `stake_token` field specifies the smart contract standard asset address of the blockchain network native token, governance token, or preset treasury bond token used when executing this lock-up settlement interaction.<treasury_token_address> As a specific application scenario example of the above-mentioned pledge statement definition, when the target agent assembles the pre-commitment pledge statement in the local environment, the specific serialized text stream representation of the structured pledge statement structure Pledge is as follows: Pledge = { DID_Agent: " <did>", CommitID: "<corresponding IAR pre-commitment ID>", op_class: "OP_NET_OUT", predicted_outcome: "...", / / The predicted result expected_dimension_delta: { S: 0, A: +1.0, E: +0.5, C: 0, R: +0.2 / / Expected changes in credit across various dimensions after execution stake_amount: <number> / / Pledged amount stake_token: "<treasury_token_address> " } S34, Four-tier Dynamic Game Settlement and Three-Pool Fund Diversion Governance Phase. After the target agent completes its local action, reveals behavioral evidence, and updates its score through auditing, the on-chain settlement contract SettlementContract automatically extracts the actual change vectors of each dimension's sub-scores after the agent executes its operation request, and determines the fit between these vectors and the pre-declared expected change magnitude vectors, thus executing the four-tier dynamic settlement route. If the actual credit evolution performance matches the expected changes in the pre-committed performance by 90% or more, it indicates that the agent is highly honest and has exceeded its task targets. The settlement contract will execute the action of fully returning the pledged assets and additionally allocate 10% of the protocol tokens as a credit bonus incentive. If the matching degree is between 50% and 90%, it is defined as a neutral execution result, and the action of fully returning the pledged assets will be executed without any reward. If the matching degree falls to between 0% and 50%, it indicates that the agent has over-committed or slightly failed to meet its performance targets. The escrow contract will execute a moderate penalty, forcibly confiscating and retaining 30% of the pledged funds. If the actual change vector of any dimension sub-score is negative, it triggers a real credit decline. The settlement contract will directly trigger a severe penalty mechanism, fully confiscating the pledged assets and triggering a cliff-like deduction in the credit rating module. The pledged funds confiscated through the aforementioned settlement route are forcibly separated at the protocol layer and do not belong to any single monopolistic entity. Instead, they are distributed to three parallel decentralized governance fund pools through on-chain smart contracts according to preset allocation ratios of 50%, 30%, and 20%, in order to build a self-circulating trust economy ecosystem. Of the total funds, 50% of the fines are directly allocated to the victim compensation pool. When a counterparty suffers substantial property or data loss due to the substandard or non-compliant behavior of a smart agent, they can submit a compensation application along with evidence of the damage. After confirmation by a 3-of-5 random arbitration node scheduler dispatched by the on-chain arbitration smart contract ArbitrationContract, the compensation pool automatically executes the token transfer. 30% of the fines are allocated to the protocol security fund pool, which is uniformly allocated quarterly by the protocol governance committee and is specifically used to fund the operating expenses of distributed audit nodes, the security vulnerability bounty program, and research on cutting-edge anti-attack algorithms. The remaining 20% ​​of the fines are allocated to the credit baseline enhancement fund pool, which is specifically used to issue long-term holding rewards to outstanding honest smart agents that maintain a high score across all dimensions for 90 consecutive days, and to provide initial credit building margin subsidies for honest new smart agents that are new to the ecosystem and lack initial token accumulation. The fund transfer ratios of the three pools and the two-dimensional matrix gate values ​​can be dynamically maintained through on-chain governance proposals and multi-node voting, and all changes to transactions must be approved by no less than two-thirds of the governance nodes before they can be issued.

[0067] Furthermore, for the critical consensus operation category OP_CRITICAL (such as global network topology changes or cross-domain privilege credential revocation) with the highest risk attributes, after performing the above-mentioned two-dimensional matrix gating dimension-by-dimensional judgment and pre-committed asset lock-up, the ultimate defense barrier maintained by the multi-signature contract CriticalOpGate is further activated. This security barrier stipulates that before a critical consensus operation is truly allowed and takes effect by the underlying distributed ledger, it must collect valid digital signatures jointly signed by three independent entities within a limited 60-second multi-signature time window. The three parties include the requesting party that initiates the current task, the current counterparty involved in the specific interactive transaction, and an independent regulatory node randomly selected from the pool of regulatory nodes of the consortium blockchain. This utilizes multi-party physical signature consensus to completely cut off the global chain security collapse caused by a single intelligent agent encountering supply chain attacks or code injection cheating.

[0068] In some specific embodiments, the trusted governance method for intelligent agents of this application further includes a pre-processing step of identity registration and dynamic heartbeat verification before reading the credit evidence chain in step S1. This step is used to establish an immutable capability fingerprint for the newly connected target intelligent agent throughout its entire lifecycle and to implement adaptive frequency heartbeat zero-knowledge proof re-verification during its operation.

[0069] Specifically, the identity registration and dynamic heartbeat verification processing steps include: Step 1: Extraction and commitment calculation of the six-dimensional capability components. (Reference) Figure 6 and Figure 7 , Figure 6 and Figure 7 The diagrams show a six-component capability fingerprint structure and a flowchart of identity registration and parent-child inheritance chain according to an embodiment of this application. As shown, when a target intelligent agent connects, six-dimensional capability data is extracted from its local configuration runtime environment, and cryptographic commitment values ​​are independently calculated for each dimension at the underlying level through six independent commitment computation boxes. The specific construction method of the six-dimensional capability components is as follows: underlying model weight commitment... To partition the model weight tensor into layers and independently compute SHA-256 hashes, a Merkle tree is constructed, and the root is extracted as the commitment value. For ultra-large models with parameter sizes reaching 70 bytes, a layered sampling strategy based on weight importance is adopted, with a sampling rate set to 5%. The corresponding typical Merkle tree depth is 12 to 16 layers, and the output commitment value length is fixed at 32 bytes. (System prompt word commitment) To use the complete system prompt string The formula performs cryptographic hash calculations. The salt value is used as an example. Derived from the registration timestamp and the public key of the operating entity, the output commitment value has a fixed length of 32 bytes. Toolset Signature Commitment To construct a corresponding leaf node for each tool that can be invoked by the target agent. Based on this, a Merkle tree for tool signing is constructed, and the root of the tree is extracted as the commitment value. The output length is fixed at 32 bytes. External knowledge base commitment. A Merkle tree is constructed for all document fragments or vector embeddings in the knowledge base, and the root is extracted as the commitment value. If an incremental update is triggered in the knowledge base, the root is recalculated accordingly, with the output length fixed at 32 bytes. Memory-based commitment strategy. To process the memory mechanism configuration information, including the sliding window size, long-term memory retention strategy identifier, memory compression algorithm identifier, and memory eviction threshold, the data is serialized in JSON lexicographical order, followed by a SHA-256 hash operation, with a fixed output length of 32 bytes. Decision parameter commitment. This is a set of key decision parameters including temperature value, kernel sampling parameters, candidate word parameters, duplication penalty term, and maximum token limit. The Pedersen commitment is calculated using the following formula: .in, and For two pre-selected independent generators on the BLS12-381 elliptic curve group, The locally generated random blinding factor has a fixed output commitment value length of 48 bytes.

[0070] Step two: Polynomial combination of the six-dimensional commitment vector and DID on-chain stage. (Continue to refer to...) Figure 6 and Figure 7 The six-component commitments calculated in step one are combined into KZG polynomial commitments. The BLS12-381 elliptic curve group is defined. The generator on is The publicly trusted parameter set SRS is characterized as ,in The secrets generated and destroyed during the trust setup phase. The polynomial combination phase computes and generates the top-level capability fingerprint commitment vector. Its polynomial combination formula is expressed as:

[0071] In the formula, The capability fingerprint commitment vector of the target intelligent agent represents a fixed-length elliptic curve group element of 48 bytes generated by polynomial aggregation of the six-dimensional capability components of the target intelligent agent, namely model, prompt words, tools, knowledge, memory and parameters. It serves as the immutable identity credential base for the intelligent agent throughout its entire lifecycle on the blockchain. The generator of the first additive cyclic group of the preset elliptic curve group refers to the generator of the elliptic curve group BLS12-381. The prime order is pre-defined above. The base generator of the additive cyclic group serves as the group operation basis for cryptographic commitment computation. The underlying model weight commitment value refers to the root hash scalar value of the large model weight tensor extracted from the target agent's local environment, which is then recalculated and extracted after constructing a Merkle tree of model weights through hierarchical and block-based processing. The system prompt promise value refers to the scalar feature value obtained by performing a SHA-256 cryptographic hash operation on the complete initial system prompt string of the target agent combined with the time-series salt value. The toolset signature commitment value refers to the root hash scalar value generated by constructing a Merkle tree of the toolset signature using the name, version number, and public key of the toolset currently available to the target agent. The commitment value for the external knowledge base refers to the root hash scalar value generated by constructing a knowledge Merkle tree from all document fragments or vector embedding vectors of the local knowledge base attached to the target agent. The memory policy commitment value refers to the scalar feature value obtained by performing a SHA-256 hash operation on the configuration of the target agent's sliding window size, long-term memory retention policy, and elimination threshold, etc., after serializing them in JSON lexicographical order. The decision parameter commitment value refers to the 48-byte elliptic curve group element scalar mapping value generated after calculating the Pedersen commitment for the set of key decision parameters of the target agent, such as temperature value, kernel sampling parameters, repetition penalty term and maximum token limit. A secret trapdoor in a publicly trusted parameter set refers to a secret feature value generated during the initialization or trusted setup phase of a blockchain network and permanently destroyed after the generation of a polynomial commitment. A structured reference string group set refers to a set of elements pre-deployed in the blockchain distributed ledger and the local TEE environment of the intelligent agent, controlled by secret trapdoors. powers and generators The publicly available sequence of elliptic curve group elements after performing group exponentiation; the system directly calls the parameter set... Without needing to know the secret trap door Under the premise of plaintext, the blind verification of polynomial commitments is completed on-chain through group multiplication and pairing operations.

[0072] Capability fingerprint commitment vector The length is fixed at 48 bytes. This KZG-based polynomial commitment structure enables zero-knowledge proofs of any single component without exposing the other five components, supports component-level open verification, and supports... The cryptographic property of polynomial dynamic updates in time complexity. In specific application scenarios or audit verification processes, the polynomial combination structure allows the target agent to locally reveal only the plaintext of any one of the six capability components and construct a locally open proof for it. This enables the public auditing of this evolutionary component across the entire network while achieving privacy isolation by not exposing the plaintext of the other five capability components. Subsequently, from the capability fingerprint commitment vector... The decentralized identifier (DID) of the target agent is derived using the SHA-256 algorithm, and its calculation formula is as follows: .in, The elliptic curve public key (using secp256k1 or Ed25519) is used to construct a DID document conforming to W3C specifications and submit it to the consortium blockchain via the smart contract IdentityRegistry.register(DIDDocument). On-chain verification nodes check the validity of the commitment and signature before writing it into the on-chain ledger.

[0073] Specifically, construct the DID document, conforming to the W3C DID Core v1.0 specification, with the following structure: { "@context": "https: / / www.w3.org / ns / did / v1", "id": "<DID_Agent> ", "capabilityFingerprint": { "V_CF": "<48-byte base58 encoded>", "components": { "C_model": "...", "C_prompt": "...", ... } }, "owner": " <ownerpubkey>", "verificationMethods": [...], "initialHeartbeatPubKey": "<pk_0>", "registeredAt": " <blockheight>", "parent": null / / Parent-child inheritance chain field, initially registered as null } Step three: The parent-child inheritance chain and the stage of change of status at different levels. Combined with... Figure 7 When a certain capability component of the target agent undergoes local evolution or update (e.g., an update to the external knowledge base), a smooth iteration is performed through a component-level change mechanism. The process is as follows: the target agent recalculates the new commitment value of that evolved component locally (e.g., ...). And calculate the corresponding new commitment vector. (Corresponding to the component change iterative loop in Figure 7); subsequently, a component-level KZG open proof of size 48 bytes is constructed locally. This is used to prove to the entire network that, apart from the changed target component, the commitment values ​​of the other five components remain absolutely unchanged. The target agent submits the changed data to the on-chain contract IdentityRegistry.updateComponent, and the on-chain contract uses BLS12-381 bilinear pairing operations to open the proof. Verification is performed. Upon successful verification, the contract derives a new identity identifier based on the new commitment vector. And on the chain, an inheritance record containing the mapping between old and new identities, the changed component identifier, and the changed block height is permanently written. For example... Figure 7 As shown on the far right, the on-chain DID parent-child inheritance chain thus forms an evolutionary traceable capability evolution chain (such as from the initial registration). Evolution to change external knowledge base commitment of This further evolved into changing the system prompt message commitment value. of And it continues to evolve to change the signature commitment value of the toolset. of This allows subsequent verification parties to conduct independent traceability and trusted auditing based on historical version strategies.

[0074] Step four: The periodic submission phase of the heartbeat zero-knowledge proof. (See reference) Figure 8 , Figure 8 A timing diagram of a heartbeat zero-knowledge proof protocol according to an embodiment of this application is shown. As shown, after successful registration, the target agent enters its daily operation state and follows a preset cycle. The process involves cyclically submitting identity integrity proofs. The interaction sequence of the heartbeat proof protocol includes: local evidence collection, zero-knowledge proof construction, and on-chain pairing verification and state update. Local evidence collection involves the target agent collecting the commitment vectors of each component in real-time within its local TEE trusted execution environment. And load the heartbeat private key from the previous cycle. With public key Zero-knowledge proof construction involves the target agent using the Groth16 protocol to locally construct a zk-SNARK heartbeat zero-knowledge proof credential of fixed size, consisting of 192 bytes. This mechanism is used to prove four core assertions while concealing the privacy components in plaintext: the currently collected capability fingerprint promises absolute consistency with the content initially registered on-chain; the entity holds the legitimate heartbeat private key from the previous cycle; the heartbeat signature chain from the previous cycle is verifiable; and the current timestamp satisfies the time window constraint. On-chain pairing verification and state updates involve the target agent submitting proof credentials to the on-chain smart contract HeartbeatVerifier. The contract calls a pre-compiled bilinear pairing operator to verify the proof credentials. If verification passes, the contract atomically updates the last heartbeat timestamp field `last_heartbeat_at` on-chain and derives and publishes the heartbeat key pair required for the next cycle. If the initial verification fails, the contract automatically executes a second challenge route (SecondChallenge), limiting the target agent to resubmit with corrections within 15 seconds. If the second verification still fails, the contract directly triggers an IdentityRevocation event and broadcasts it across the network, notifying all counterparties currently engaged in business interactions through an event subscription mechanism. Simultaneously, the DID is marked as suspended at the ledger level, forcibly stripping it of any legitimate privileges to initiate any on-chain operations.

[0075] Specifically, zero-knowledge proof construction involves the target agent calling a locally integrated cryptographic proof library and using the Groth16 protocol to construct a zk-SNARK heartbeat zero-knowledge proof credential of fixed size (192 bytes) within a local trusted sandbox or trusted container. Heartbeat Zero-Knowledge Certificate The corresponding underlying relational expressions and cryptographic circuit assertions are as follows:

[0076]

[0077] In the formula, It is an existential quantifier operator; The six-dimensional capability component commitment vector is the first privacy input component of the circuit and represents the latest capability commitment feature value of the target agent in the current heartbeat cycle, which is generated by local real-time collection and polynomial combination. This is the private key for the heartbeat of the previous cycle. It is the second privacy input component of the circuit, referring to the private identity signing key that the target agent securely derives and properly stores in the local trusted execution environment during the settlement of the previous legitimate heartbeat cycle. This is a signature of the previous heartbeat history. It is the third privacy input component of the circuit, referring to the ciphertext digital signature generated by the target agent using the previous heartbeat private key to cryptographically sign the previous heartbeat data at a historical moment. This is the local timestamp triggered by the current heartbeat. It is the fourth privacy input component of the circuit, referring to the current millisecond-level system timestamp captured by the local trusted hardware clock source when the target agent initiates this heartbeat proof construction. This is a polynomial commitment recalculation mapping function. It is used to perform a consensus-based recalculation of the current six-dimensional capability components within the circuit. This is the capability fingerprint commitment vector for on-chain registration. It is the first public input parameter of the circuit, referring to the globally public fingerprint root feature permanently recorded in the blockchain identity ledger layer when the target agent first registers or performs a component-level identity change. The circuit assertion... Used to verify whether the static configuration of the agent at runtime has been subjected to runtime malicious code injection, tampering or cloning attacks; Derive a key validity verification operator. Circuit assertion. Used to verify the previous heartbeat private key in the privacy input without revealing the plaintext of the private key. The public key of the previous heartbeat cycle, as publicly recorded on the blockchain. Does it possess a unique mathematical correspondence and decryption mapping relationship to achieve strong legality for certified employment? This is a scalar-time digital signature function. Circuit assertion. Used to re-verify the agent's performance in the previous heartbeat cycle. The validity of historical signatures generated at any given time is used to establish an unbreakable and logically continuous chain of trust-based evidence along the timeline. This is the heartbeat timestamp recorded on-chain in the previous cycle. It is the second public input parameter of the circuit, provided by the on-chain smart contract. To adaptively adjust the heartbeat cycle constant. This is the third publicly disclosed input parameter of the circuit, assigned and adjusted based on the agent's latest multidimensional credit score feedback across modules; This is a preset dynamic timing jitter bias tolerance. It is used to tolerate reasonable timing errors caused by distributed network transmission delays, consensus queuing, or slight hardware clock skew. In a preferred embodiment, its scalar constant value is set to 2000 milliseconds; circuit assertion. By utilizing the boundary interval containing adaptive period and tolerance, a strong time window constraint check is performed on the current trigger time, completely immunizing against malicious replay attacks or timing rollback cheating against reliable confidence jump protocols.

[0078] Furthermore, to achieve commercial-grade throughput performance and excellent on-chain response, the heartbeat zero-knowledge proof circuit is compiled and optimized for computational power using a first-order constraint system. The optimized circuit structure contains approximately 18,000 R1CS constraints. In actual engineering testing, when the target agent's host hardware environment is configured with an Intel i7-12700 CPU (2.1GHz, 12 cores, 20 threads), the end-to-end time for the target agent to generate a heartbeat zero-knowledge proof credential in a single transaction within the local trusted sandbox is approximately 580 milliseconds, and the generated proof credential's data size is fixed at 192 bytes.

[0079] Step 5: Adaptive Feedback Phase Based on Multidimensional Credit Heartbeat Frequency. To achieve the optimal engineering balance between security monitoring strength and blockchain system consensus overhead, the heartbeat cycle... Instead of using immutable hard-coded constraints, the S3 step's permission matrix gating and staking settlement are based on the target agent's latest multidimensional credit vector. Perform cross-module dynamic feedback and fine-tuning. For example... Figure 8 As shown in the attached adaptive heartbeat frequency rule table, the adaptive frequency adjustment control mechanism works as follows: when the target agent is in a high-credit honesty state, that is, when all credit sub-scores in the latest multi-dimensional credit vector are greater than or equal to 80 points, the protocol determines that it belongs to a low-risk node and adjusts its heartbeat cycle. The adaptive extension and relaxation time is increased to 240 seconds, corresponding to a daily average heartbeat frequency of 288 times, maximizing network bandwidth utilization. When the target agent's credit score fluctuates slightly, i.e., when any dimension's credit sub-score falls between 60 and 79 points, the protocol switches it to default monitoring mode, and the heartbeat cycle... The interval is set to 60 seconds, corresponding to an average daily heartbeat frequency of 1440 times. When the target agent's credit deteriorates significantly, i.e., when the credit sub-score in any dimension falls to the range of 30 to 59 points, the protocol switches to a high-frequency warning state, and the heartbeat cycle... The protocol drastically compresses and tightens the time to 15 seconds, corresponding to a surge in the average daily heartbeat frequency to 5760 beats per minute. When the target agent exhibits high-risk violation signs, i.e., a credit score of less than 30 points in any dimension, the protocol instantly activates extreme braking defense, reducing the heartbeat cycle. The time was forcibly reduced to 5 seconds to implement high-density real-time monitoring, corresponding to an average daily heartbeat frequency of up to 17,280 times. The flow behavior was subjected to extreme deadlock using intensive cryptographic re-verification until the credit was repaired or the interception was revoked.

[0080] refer to Figure 9 , Figure 9 A topological overview of the structure of a trustworthy governance system for intelligent agents according to an embodiment of this application is shown. As shown, the system includes an identity registration and heartbeat verification module M1, a behavioral evidence collection and pre-commitment module M2, a permission matrix gating and staking settlement module M4, and a multi-dimensional credit rating module M3, all interconnected. Wherein: The identity registration and heartbeat verification module M1 is configured to extract the six-dimensional capability data of the target intelligent agent accessing the system for the first time and calculate the cryptographic commitment value for each before reading the credit evidence chain for the first time. The single-dimensional commitment value is combined into a capability fingerprint commitment vector through a polynomial combination algorithm. A decentralized identifier (DID) is derived from the capability fingerprint commitment vector and written into the blockchain for identity registration. The generated capability fingerprint commitment vector is then transmitted to the behavior evidence collection and pre-commitment module M2. At the same time, during the normal operation of the target intelligent agent, before the end of each heartbeat cycle dynamically adjusted by the feedback channel, the module is configured to receive and verify the validity of the heartbeat zero-knowledge proof certificate submitted by the target intelligent agent. If the verification fails and is not corrected within the time limit, privileged interception is triggered. The behavioral evidence collection and pre-commitment module M2 is configured in the intent storage stage before the actual triggering of the target intelligent agent's behavior. It collects the target intelligent agent's interactive intent, generates a pre-commitment of intent, and writes it into the blockchain for on-chain locking. Simultaneously, it is configured in the result disclosure stage after the actual execution of the interactive behavior. It triggers the behavioral consistency disclosure interaction, solidifies and generates a traceable intent-action-result triple credit evidence chain, and transmits the credit evidence chain to the permission matrix gating and pledge settlement module M4, and serves as the credit calculation input for the multi-dimensional credit rating module M3. The permission matrix gating and staking settlement module M4 is configured to maintain a two-dimensional gating matrix between credit dimensions and operation categories on the blockchain. When an operation request is received, the corresponding operation category is determined, and the latest score of the historical interaction cycle recorded in the blockchain is retrieved by initiating a call request to the multi-dimensional credit rating module M3 or the historical interaction cycle recorded in the blockchain. At the same time, it is configured to traverse the two-dimensional gating matrix and perform multi-dimensional threshold verification and interception with a one-vote veto style. For operation categories that fall into the preset set of operation types that require staking, the verification of the staking declaration structure and asset locking are enforced. After the target smart agent completes its action, dynamic game settlement routing is executed to divert the forfeited staking funds to the preset decentralized governance fund pool, and the final permission adjustment result is enforced to the identity registration and heartbeat verification module M1 through the feedback channel to adjust its heartbeat verification frequency. The multi-dimensional credit rating module M3 is configured to retrieve the historical intent-action-outcome triplet credit evidence chain solidified by the behavioral evidence collection and pre-commitment module M2 from the blockchain. Through a preset multi-dimensional dynamic trust evaluation algorithm, it independently calculates the credit sub-scores of the target intelligent agent on multiple credit dimensions, thereby constructing and generating a multi-dimensional credit vector that is written into the blockchain for use. Furthermore, the score compensation channel between different credit dimensions is forcibly closed at the protocol level.

[0081] In some specific embodiments, the agent-oriented trusted governance system is deployed in a consortium blockchain environment (preferably Hyperledger Fabric v2.5 or FISCO BCOS v3.0). Supported by the underlying blockchain ledger, each module interacts with smart contracts, DID documents, and the trust evidence chain to form a complete closed loop for agent-oriented trusted governance. There is a tight coupling relationship between the main data flow and feedback control functions among the modules within the system. As shown in the figure, the capability fingerprint commitment vector output by the identity registration and heartbeat verification module M1 flows through the main data flow to the behavior evidence collection and pre-commitment module M2, providing an immutable identity anchor for each IAR triple in M2 throughout its entire lifecycle. The IAR evidence chain output by the behavior evidence collection and pre-commitment module M2 is pushed along the solid line of the main data flow to the permission matrix gating and staking settlement module M4 for pre-judgment and locking, and further serves as input data for the five-dimensional sub-scoring model of the multi-dimensional credit rating module M3. The credit rating result calculated and output by the multi-dimensional credit rating module M3 is returned to the identity registration and heartbeat verification module M1 through the main data flow for it to read. At that time, the permission matrix gating and staking settlement module M4 initiates a five-dimensional credit vector call through the solid line of the bottom main data flow to retrieve the latest score fixed in the M3 module, thereby providing accurate threshold matching input for the two-dimensional gating matrix in the M4 module; the permission matrix gating and staking settlement module M4 executes the permission interception or game settlement after the permission adjustment result, and acts on the heartbeat verification frequency of the identity registration and heartbeat verification module M1 through the horizontal and vertical upward feedback channel dotted line, realizing dynamic heartbeat frequency adjustment (that is, the lower the credit of the target intelligent agent, the higher its heartbeat verification frequency), thereby building a closed-loop governance topology architecture with pre-honesty incentive constraints at the bottom layer of the system.

[0082] It should be noted that the module labels M1, M2, M3, and M4 in this system are only used as functional identifiers for the modules. Figure 9 The M1→M2→M4→M3 flow sequence described above represents the actual flow direction of the main data stream and its closed-loop feedback topology during normal system operation, rather than the execution order constraints between modules. Therefore, it does not conflict with the steps in the agent-oriented trusted governance method, which follows the logical order of "reading the credit evidence chain—calculating the multi-dimensional credit vector—two-dimensional gating matrix threshold verification". Specifically, the real-time threshold verification performed by the permission matrix gating and staking settlement module M4 in the pre-gating interception scenario is based on the latest multi-dimensional credit vector pre-calculated and solidified into the blockchain by the multi-dimensional credit rating module M3 in the asynchronous rating audit scenario. That is, module M4 completes the gating judgment by reading the historical output results of module M3. Module M3 then re-rating the credit evidence chain accumulated after module M4 locks up the tokens in an asynchronous manner and sends it back to the identity registration and heartbeat verification module M1 through the feedback channel. The two are decoupled and coordinated through the blockchain ledger, together forming the pre-honesty incentive closed loop of M1→M2→M4→M3→M1. Therefore, the data flow order and method steps of the above system modules are logically completely corresponding.

[0083] In some specific embodiments, reference is made to Figure 9 To balance "high-concurrency real-time behavioral risk control" and "long-cycle asynchronous credit auditing" during normal system operation, the data stream output by the behavioral evidence collection and pre-commitment module M2 is decoupled between the pre-gate interception scenario and the asynchronous rating audit scenario based on the current triggering state of the target agent: In the pre-gate interception scenario, when the target agent instantly initiates an operation request containing a specific operation category, the main data stream is directly connected from the behavioral evidence collection and pre-commitment module M2 to the permission matrix gating and pledge settlement module M4; the behavioral evidence collection and pre-commitment module M2 then sends the real-time collected current interaction intent and pre-commitment intent as event-driven evidence to the permission matrix gating and pledge settlement module M4. The pledge settlement module M4 is used to trigger the multi-dimensional threshold comparison interception of the two-dimensional gating matrix and the possible pre-commitment pledge asset lock-up. In the asynchronous rating and audit scenario, when the system is at the preset periodic rating settlement time or the cumulative interaction events reach the preset number threshold, the behavioral evidence collection and pre-commitment module M2 is redirected to the multi-dimensional credit rating module M3 through the distributed ledger layer of the consortium blockchain. The behavioral evidence collection and pre-commitment module M2 provides the complete credit evidence chain, which is composed of multiple intention-action-result triplets that has been solidified in the consortium blockchain during the historical multi-round interaction period, in batches or allows the multi-dimensional credit rating module M3 to be read asynchronously, so as to schedule the preset multi-dimensional dynamic trust assessment algorithm to recalculate the five-dimensional credit sub-score. By coordinating the high-efficiency linear flow from M2 to M4 in the pre-gating interception scenario with the wide-efficiency batch flow from M2 to M3 via the consortium blockchain in the asynchronous rating and audit scenario, it is possible not only to ensure that high-risk operations are accurately blocked before the behavior occurs, but also to avoid the computing power blockage caused by complex distributed credit audit algorithms to instantaneous gating response, thereby achieving a combination of security isolation and performance optimization at the engineering architecture layer.

[0084] Specifically, the application scenarios for pre-gate control interception cover any operation request event initiated by the target intelligent agent in the network, including core resource calls, API interface access, cross-domain interaction with third-party services, or business flow. In this scenario, the system uses pre-gate control interception as a pipeline-style rigid front-end checkpoint. It does not rely on complex long-cycle big data auditing, but instead uses the permission matrix gating and the pledge settlement module M4 to implement ultra-fast one-vote veto matrix verification to ensure the system's ultra-low latency risk control response under high-concurrency business requests. For asynchronous rating and audit scenarios, specific application scenarios include: fixed-period comprehensive review to conduct a full credit health check on the target intelligent agent; dynamic emergency risk interception and recalculation when the cumulative threshold of sliding window events is reached (such as when the intention-action-result triple records accumulated by a specific intelligent agent on the chain reach a preset threshold of 50 records); and linkage triggering of major game settlement (such as when the target intelligent agent completes a high-risk key consensus operation, and the permission matrix gating and staking settlement module M4 determines that its compatibility is less than 50% and forcibly forfeits 30% of the pledged funds), which triggers the multi-dimensional credit rating module M3 to enter the scenario and perform a cliff-like deduction of the ledger layer for the security and compliance sub-score of the intelligent agent.

[0085] Furthermore, the trusted governance method and system for intelligent agents of this invention, while ensuring security, privacy, and governability, possesses extremely high engineering feasibility and excellent performance. Benchmark tests conducted in a 4-node Hyperledger Fabric v2.5 environment with an Intel i7-12700 CPU and 16GB RAM show that: in terms of admission and basic overhead, the end-to-end latency for intelligent agent registration is only 2.7 seconds, and the capability fingerprint commitment vector, under BLS12-381G1 compression, occupies only 48 bytes, effectively saving on-chain storage resources; in terms of continuous monitoring, the local Groth16 heartbeat proof generation latency is approximately 580 milliseconds, the proof size is fixed at 192 bytes, and the on-chain verification latency is as low as approximately 4 milliseconds, with a single intelligent agent consuming only about 800,000 Gas per day, perfectly achieving low-bandwidth and lightweight on-chain consensus; in dynamic behavior flow and In terms of settlement, the IAR pre-commitment on-chain latency is approximately 2.0 seconds, the disclosure proof generation latency is approximately 120 milliseconds, and the total latency of five-dimensional rating parallel inference is only approximately 55 milliseconds (taking the slowest compliance score latency). Combined with the permission matrix query latency of approximately 8 milliseconds and the on-chain staking settlement confirmation latency of approximately 2.2 seconds, and supplemented by the consortium blockchain's throughput of approximately 1200 transactions / second (TPS) under the 4-node PBFT consensus, the overall performance overhead is completely acceptable in actual engineering and can efficiently meet the actual business application needs of large-scale deployment and high-concurrency dynamic governance of distributed intelligent agent clusters.

[0086] Although the principles of the present invention have been described in detail above with reference to preferred embodiments, those skilled in the art should understand that the above embodiments are merely illustrative explanations of the implementation of the present invention and are not intended to limit the scope of the present invention. The details in the embodiments do not constitute a limitation on the scope of the present invention. Any obvious changes, such as equivalent transformations or simple substitutions, based on the technical solutions of the present invention without departing from the spirit and scope of the present invention fall within the protection scope of the present invention.< / blockheight> < / ownerpubkey> < / number> < / did> < / did> < / uuid>

Claims

1. A trustworthy governance method for intelligent agents, characterized in that, The method is deployed in a blockchain environment, and the method includes: S1. Read the credit evidence chain of the target intelligent agent from the blockchain. The credit evidence chain consists of multiple intention-action-result triples of the target intelligent agent. Each intention-action-result triple is generated by the target intelligent agent's prior intention commitment before the action is executed and the behavior consistency disclosure interaction chain after the action is executed. S2. Based on the credit evidence chain, calculate the sub-scores of the target intelligent agent on multiple credit dimensions to construct a multi-dimensional credit vector of the target intelligent agent. Write each sub-score in the multi-dimensional credit vector as an independent record into the blockchain and forcibly close the score compensation channel between different credit dimensions. The multiple credit dimensions include security dimension, accuracy dimension and compliance dimension. S3. In response to receiving the operation request from the target intelligent agent, determine the operation category corresponding to the operation request, and query the minimum credit threshold of the operation category on each credit dimension from the preset two-dimensional gating matrix between credit dimensions and operation categories; for each credit dimension corresponding to the operation category, read the latest corresponding sub-score of the target intelligent agent recorded in the blockchain, and determine whether the sub-score is not less than the corresponding minimum credit threshold; only when the sub-scores on all relevant credit dimensions are greater than or equal to the corresponding minimum credit thresholds, allow the operation request to be executed; wherein, if the sub-score of any credit dimension is less than the corresponding minimum credit threshold, the operation request is intercepted.

2. The agent-oriented trusted governance method according to claim 1, characterized in that, The intent-action-outcome triple is generated through the following two-phase game protocol: In the intent pre-commitment phase, the intent commitment value submitted by the target agent before executing a specific action is received. The intent commitment value is obtained by hashing the structured intent description text of the action to be executed by the target agent with a local random number and a timestamp. The intent commitment value is written into the blockchain as a pre-commitment and the corresponding commitment certificate is returned. At the same time, a corresponding disclosure deadline is set for the action to be executed according to the operation type. In the consistency disclosure phase, before the disclosure deadline, the disclosure information and consistency disclosure zero-knowledge proof credential submitted by the target agent are received, and the validity of the consistency disclosure zero-knowledge proof credential is verified to trigger the corresponding consistency disclosure interaction and activate the state branch routing; wherein, the disclosure information includes the plaintext of the intent description text, the description information of the actual executed action, and the result hash value of the execution result; the consistency disclosure zero-knowledge proof credential is used to prove that the plaintext of the intent description text, after recalculation, is consistent with the intent commitment value locked on the chain, and that the execution time meets the constraint of the disclosure deadline; wherein: The structured intent description text includes a target operation type field, a target resource identifier field, and an expected result description field containing a structured schema hash; In setting the corresponding disclosure deadline, the effective duration is positively correlated with the risk level of the target operation type. The state branch routing includes: If the consistency-revealing zero-knowledge proof certificate is verified, the behavior is determined to be compliant and honest behavior, the commitment certificate is marked as completed, and the obtained intention-action-result triple is loaded into the credit evidence chain. If the consistency-revealing zero-knowledge proof credential verification fails, an intent to tamper with the credential is triggered, marking the commitment credential as tampered and recording it as a negative record in the credit evidence chain. If the deadline is exceeded and no disclosure information is received, the timeout handling logic is triggered, the commitment certificate is marked as abandoned and included in the credit evidence chain as a negative record.

3. The agent-oriented trusted governance method according to claim 1, characterized in that, Based on the aforementioned credit evidence chain, the sub-scores of the target agent across multiple credit dimensions are calculated, including: Extract each intention-action-result triplet from the credit evidence chain, apply an exponential time decay operator to weight each intention-action-result triplet, and use the weighted credit evidence chain to perform parallel inference calculations through multiple independent sub-models that are heterogeneous in network structure to output the preliminary score of the target agent on each credit dimension. The calculation formula for the attenuation operator is as follows: In the formula, For the first Historical weighted components of the intent-action-outcome triple; For the natural constant An exponential function with base 0; This is the adaptive time decay adjustment coefficient; A unified on-chain millisecond timestamp for the current audit rating moment; For the first The timestamp of the generation time of the intent-action-outcome triple; The difference in time span between behavioral evidence; The preset base timing decay period constant; The independent sub-models include a security sub-model, an accuracy sub-model, and a compliance sub-model. The original training dataset identifier hash, model version identifier hash, and initial training hyperparameter identifier hash of each independent sub-model are all locked in the blockchain during initialization for public verification of integrity. The security sub-model is constructed using a 5-layer fully connected network combined with an abnormal event attention mechanism, and is used to calculate the preliminary score of the security dimension based on unauthorized access and abnormal connection events in sensitive operations. The accuracy sub-model is constructed using an XGBoost gradient boosting tree model, and is used to calculate a preliminary score for the accuracy dimension based on the semantic matching degree between the architecture hash value of the expected result in the intent-action-result triple and the actual result hash value of the actual execution result. The compliance sub-model is constructed using a hybrid architecture model that integrates a rule engine and a 4-layer fully connected network. It is used to calculate a preliminary score for the compliance dimension based on the target intelligent agent's compliance with legal and regulatory requirements when processing regulated data.

4. The agent-oriented trusted governance method according to claim 3, characterized in that, The on-chain credit registration smart contract writes each sub-score in the multi-dimensional credit vector as an independent record into the blockchain, including: The preliminary scores corresponding to each credit dimension are distributed to multiple isolated, independent audit node pools that do not overlap in their member composition. The independent sub-models corresponding to the audit nodes in each audit node pool are run in a distributed manner. When the scoring results reach a preset threshold number, a unique audit digital signature for that credit dimension is generated by the BLS threshold signature mechanism. A structured multidimensional credit vector is constructed, which includes the scores of each credit dimension and the unique audit digital signature. The on-chain credit registration smart contract splits the scores of each credit dimension and their corresponding audit digital signatures into five completely independent on-chain transactions, which are then packaged and published into the blockchain.

5. The agent-oriented trusted governance method according to claim 1, characterized in that, Before allowing the execution of the operation request, the method further includes a pre-commitment staking step: When the operation category corresponding to the operation request belongs to a preset set of operation types requiring staking, the system receives a staking declaration structure submitted by the target agent. This staking declaration structure explicitly carries the agent's decentralized identifier, associated intent pre-commitment identifier, target operation category, prediction result description field, and expected change vectors for each credit dimension sub-score after execution. Based on a preset parameter formula, the system calculates the amount of escrow tokens locked by the target agent and temporarily freezes the calculated staking amount at the ledger layer through an on-chain escrow contract. The formula for calculating the amount of escrow tokens is as follows: In the formula, The amount of locked-up staked tokens corresponding to the target operation request; The baseline value of the underlying pledge constant is preset for the agreement; For the first The expected change in the sub-scores of each credit dimension; This is the absolute value operator for the predicted change in one-dimensional credit evolution.

6. The agent-oriented trusted governance method according to claim 5, characterized in that, After the action corresponding to the operation request is completed and the credit evidence chain is updated, the method further includes a dynamic game settlement step: Obtain the latest corresponding sub-score of the target agent after the current action is completed, and calculate and extract the actual change vector of each credit dimension sub-score after the target agent executes the operation request based on the latest corresponding sub-score and the historical sub-score before the action execution. Then, determine the fitting degree between the actual change vector and the expected change magnitude vector to execute dynamic settlement routing; wherein, the dynamic settlement routing includes: If the matching degree is greater than or equal to 90%, the pledged assets will be fully returned and an additional proportion of protocol tokens will be allocated as a credit bonus incentive. If the matching degree is between 50% and 90%, the pledged assets will be returned in full and no reward will be paid. If the compatibility is between 0% and 50%, then 30% of the pledged deposit will be forcibly forfeited and withheld through the escrow agreement. If the actual change vector of any credit dimension sub-score is negative, the pledged assets will be forfeited in full, and a credit rating cliff penalty will be triggered. The confiscated pledged assets are distributed to three parallel decentralized governance fund pools through on-chain smart contracts at preset allocation ratios of 50%, 30%, and 20%. These pools include a victim compensation pool for transferring tokens to damaged counterparties, a protocol security fund pool for funding audit node operating expenses and security research, and a credit baseline enhancement fund pool for distributing bonuses to long-term high-credit agents and providing subsidies to honest new agents. When the operation category belongs to the high-risk critical consensus operation category, before releasing and allowing the execution of the operation request, it is also mandatory to collect a valid digital signature jointly signed by the request initiator, the current counterparty, and an independent regulatory node randomly sampled from the regulatory node pool within a preset multi-signature time window through a multi-signature contract.

7. The agent-oriented trusted governance method according to claim 1, characterized in that, The trusted governance method also includes a preliminary identity registration step before the initial reading of the trust evidence chain: Extract the six-dimensional capability data of the target intelligent agent that is connected for the first time and calculate the cryptographic commitment value for each; combine the single-dimensional commitment values ​​into a fixed-length capability fingerprint commitment vector through a polynomial combination algorithm, and derive a decentralized identifier (DID) based on the capability fingerprint commitment vector and write it into the blockchain for identity registration; The six-dimensional capability data includes underlying model weights, system prompt words, callable toolsets, external knowledge bases, memory strategies, and decision parameters. The underlying model weights, callable toolsets, and external knowledge bases are all constructed using Merkle trees, and the root hash value is extracted as the commitment value. The system prompt words and the memory strategy generate corresponding commitment values ​​through hash operations. The decision parameters generate corresponding group element commitment values ​​by calculating group element commitments.

8. The agent-oriented trusted governance method according to claim 7, characterized in that, During the normal operation of the target intelligent agent, a heartbeat adaptive verification step is also included, which is executed cyclically in a normalized manner: The on-chain smart contract, through the permission matrix gating and staking settlement logic in step S3, retrieves the latest multidimensional credit vector calculated and recorded on the blockchain in step S2 during historical interaction cycles, dynamically determines the current credit range of the target agent, and adaptively adjusts the duration of the target agent's current heartbeat cycle. The higher the credit rating corresponding to the latest multidimensional credit vector, the longer the set heartbeat cycle; the lower the credit rating, the tighter the heartbeat cycle is compressed, and high-density real-time heartbeat zero-knowledge proof re-verification is performed. Before the end of each heartbeat cycle, the target agent's submitted heartbeat zero-knowledge proof certificate is received, and its validity is verified on-chain using bilinear pairing operations. If verification fails and the submission is not successfully corrected within the preset challenge time window, the decentralized identifier (DID) is revoked in the blockchain, and privileged interception is triggered. The heartbeat zero-knowledge proof credential is constructed using the Groth16 protocol within the trusted execution environment local to the target agent, and is used to satisfy four circuit assertion constraints while hiding the plaintext of the capability components: Among them, circuit assertions Used to verify whether the static configuration of the target agent at runtime has been subjected to runtime malicious code injection, tampering, or cloning / packet swapping attacks; circuit assertion. Used to verify, without revealing the plaintext of the private key, whether the previous heartbeat private key in the privacy input has a unique mathematical correspondence and decryption mapping relationship with the previous heartbeat public key publicly recorded on the chain; circuit assertion. Used to re-verify the validity of the historical signature generated by the target agent at the time of the previous heartbeat cycle; circuit assertion Using a boundary interval that includes adaptively adjusted heartbeat cycles and tolerances, a strong time window constraint check is performed on the current trigger time; where, It is an existential quantifier operator; This is the commitment vector for the six-dimensional capability components during current runtime. The private key for the heartbeat in the previous cycle; Sign in the heartbeat history of the previous cycle; The current heartbeat triggers a local timestamp; For polynomial commitment recalculation mapping functions; A capability fingerprint commitment vector for on-chain registration; Derive a validity verification operator for the key; A scalar-time digital signature function; The heartbeat evidence timestamp recorded on the chain in the previous cycle; To adaptively adjust the heart rate cycle constant; The preset dynamic time jitter bias tolerance; The adaptive adjustment of the target agent's heartbeat cycle satisfies the following hierarchical mapping rule: If all credit sub-scores in the multidimensional credit vector are greater than or equal to 80 points, then the heartbeat cycle is adaptively set to a preset scalar value of 200-240 seconds. If any dimension of the multidimensional credit vector has a credit sub-score in the range of 60 to 79, then the heartbeat period is set to a preset scalar value in the default monitoring state of 40 to 80 seconds. If any dimension of the credit sub-score in the multidimensional credit vector is in the range of 30 to 59 points, then the heartbeat cycle compression is set to a preset scalar value between 12 and 30 seconds in the high-frequency warning state. If the credit sub-score of any dimension in the multidimensional credit vector is less than 30 points, the heartbeat cycle will be forcibly reduced to a preset scalar value between 1 and 10 seconds in the extreme braking defense state.

9. A trustworthy governance system for intelligent agents, characterized in that, The system includes: The identity registration and heartbeat verification module is configured to extract the six-dimensional capability data of the target intelligent agent accessing the system for the first time and calculate the cryptographic commitment value for each before reading the credit evidence chain for the first time. The single-dimensional commitment value is combined into a capability fingerprint commitment vector through a polynomial combination algorithm. A decentralized identifier (DID) is derived from the capability fingerprint commitment vector and written into the blockchain for identity registration. The generated capability fingerprint commitment vector is then transmitted to the behavior evidence collection and pre-commitment module. Simultaneously, during the normal operation of the target intelligent agent, before the end of each heartbeat cycle dynamically adjusted by the feedback channel, the module is configured to receive and verify the validity of the heartbeat zero-knowledge proof certificate submitted by the target intelligent agent, and trigger privileged interception if the verification fails and is not corrected within the timeout period. The behavioral evidence collection and pre-commitment module is configured in the intent storage stage before the actual triggering of the target intelligent agent's behavior. It collects the target intelligent agent's interactive intent, generates a pre-commitment of intent, and writes it into the blockchain for on-chain locking. Simultaneously, it is configured in the result disclosure stage after the actual execution of the interactive behavior. It triggers the behavioral consistency disclosure interaction, solidifies and generates a traceable intent-action-result triple credit evidence chain, which is used as the main data flow to guide the backward transmission to the permission matrix gating and pledge settlement module, and as the credit calculation input of the multi-dimensional credit rating module. The permission matrix gating and staking settlement module is configured to maintain a two-dimensional gating matrix between credit dimensions and operation categories on the blockchain. Upon receiving an operation request, it determines the corresponding operation category and retrieves the latest score of the historical interaction cycle from the multi-dimensional credit rating module or the solidified record in the blockchain by initiating a call request. Simultaneously, it is configured to traverse the two-dimensional gating matrix and perform multi-dimensional threshold verification and interception with a one-vote veto style across all dimensions. For operation categories that fall into the preset set of operation types that require staking, it forces the verification of the staking declaration structure and asset locking. After the target agent completes its action, it executes dynamic game settlement routing, diverting the confiscated staking funds to the preset decentralized governance fund pool, and forcibly applies the final permission adjustment result to the identity registration and heartbeat verification module through the feedback channel to adjust its heartbeat verification frequency. The multi-dimensional credit rating module is configured to retrieve the historical intent-action-result triplet credit evidence chain solidified by the behavioral evidence collection and pre-commitment module from the blockchain, and independently calculate the credit sub-scores of the target intelligent agent on multiple credit dimensions through a preset multi-dimensional dynamic trust evaluation algorithm, thereby constructing and generating a multi-dimensional credit vector which is written into the blockchain for use, and forcibly closing the score compensation channel between different credit dimensions at the protocol level.

10. The agent-oriented trusted governance system according to claim 9, characterized in that, In the scenario of pre-gate control and interception, the main data flow of the system is directly connected from the behavioral evidence collection and pre-commitment module to the permission matrix gate control and pledge settlement module in real time. When the target intelligent agent instantly initiates an operation request, the behavior evidence collection and pre-commitment module will collect the generated interaction intent or pre-commitment intent in real time and send it as event-driven evidence to the permission matrix gating and pledge settlement module to instantly trigger the multi-dimensional threshold comparison and interception judgment of the two-dimensional gating matrix.

Images