Appraisal certification system, program, and appraisal certification method

Abstract

This solution addresses the problem that while artificial intelligence (AI) agents (acting entities) may perform actions that exceed the scope of their authority or execution conditions, the impact on external resources is determined solely by the exercise of the acting entity's execution authority. [Solution] The appraisal and certification system 1 obtains action request information 100 for the target action to an external resource 40 from the acting entity 30, suspends the determination of execution or impact until confirmation that the confirmed permission information 130 is valid is established, obtains determination results 110 of the scope of authority or conformity to execution conditions from multiple verification artificial intelligence (AI) models 60, at least one of which operates in a different entity or management area from the acting entity 30, and only if the result of the integration satisfies the authorization conditions 121, generates or sets confirmed permission information 130 to a valid state based on a confirmed control authority 138 different from the execution authority, releases the suspension upon confirmation and determines the impact, and records authorization trail information 140, including information corresponding to the determination result and the authorization determination result 123, in a recording base 50 that can detect tampering.

Description

Technical Field

[0001] The present invention relates to an appraisal and certification system, a program, and an appraisal and certification method for appraising and certifying target actions by an artificial intelligence (AI) agent.

Background Art

[0002] Techniques are known for appraising and certifying the authenticity or history of assets, products or data by recording on a blockchain or other recording infrastructure and collating verification information (see, for example, Patent Documents 1 to 5). Further, techniques are known for weighting and integrating verification results from a plurality of different verification entities or a plurality of verification models to evaluate validity (see, for example, Patent Documents 6 and 7).

[0003] On the other hand, the use of so-called artificial intelligence (AI) agents, which are centered on large language models and other artificial intelligence models, and autonomously plan, select, or execute actions on external tools, APIs (Application Programming Interfaces), databases, other information systems, and other external resources, is expanding. Artificial intelligence (AI) agents can perform various actions such as connecting to external resources, authentication, using keys, generating electronic signatures, transmitting data, changing settings, executing transactions, delegating to other artificial intelligence (AI) agents, and the like, without going through sequential confirmation by humans.

[0004] In the field of access control, authorization frameworks related to granting and exercising authority are known (see, for example, Non-Patent Document 1). In addition, the zero-trust concept that does not place implicit trust and verifies each access is known (see, for example, Non-Patent Document 2). Furthermore, standards related to Decentralized Identifiers (DIDs) and verifiable credentials that enable machine-verifiable assertions of subject identifiers and attributes are known (see, for example, Non-Patent Documents 3 and 4).

Prior Art Documents

Patent Documents

[0005] [Patent Document 1] Patent No. 6894033 [Patent Document 2] Patent No. 7371301 [Patent Document 3] Patent No. 7367270 [Patent Document 4] Patent No. 7833630 [Patent Document 5] Patent No. 7477937 [Patent Document 6] Patent No. 7755716 [Patent Document 7] Patent No. 7872466 [Non-patent literature]

[0006] [Non-Patent Document 1] D. Hardt, ed., "The OAuth 2.0 Authorization Framework", RFC 6749, IETF, 2012. [Non-Patent Document 2] "Zero Trust Architecture", NIST Special Publication 800-207, National Institute of Standards and Technology, 2020 [Non-Patent Document 3] "Decentralized Identifiers (DIDs) v1.0", W3C Recommendation, 2022 [Non-Patent Document 4] "Verifiable Credentials Data Model", W3C Recommendation [Overview of the project] [Problems that the invention aims to solve]

[0007] Artificial intelligence (AI) agents may perform actions that deviate from the scope of their granted authority or the execution conditions attached to that authority due to the characteristics of probabilistically operating AI models, misdirection caused by information taken in from external sources (including so-called prompt injection), discrepancies between the training and operational environments, and failures or infringements of the execution environment. Furthermore, once the impact of such actions is determined in external resources, it may become difficult or impossible to retrieve transmitted data, restore altered states, or cancel executed transactions.

[0008] Conventional authentication and certification techniques (Patent Documents 1 to 7) determine the authenticity or provenance of assets, products, or data, and do not control the actions that artificial intelligence (AI) agents intend to perform on external resources before they are performed or before the effects are determined. Furthermore, conventional authorization frameworks (Non-Patent Documents 1 and 2) are based on the premise that the effect occurs when the entity holding the authority exercises that authority. Therefore, if the artificial intelligence (AI) agent itself holds the authority to perform the action, it is not possible to prevent the determination of the effects through the sole exercise of that authority by the entity performing the action.

[0009] Furthermore, even with prior verification by a single verifier, it is difficult to address errors in the verifier itself, common defects or infringements between the actor and the verifier, changes in the state between the time of verification and the time of execution, reuse of authorization, dilution of control through a chain of delegation to other artificial intelligence (AI) agents, and changes in preconditions after authorization. In addition, in case the appropriateness of an action is contested afterward, it is necessary to be able to prove, in a manner that is difficult to tamper with, the basis on which the action was determined.

[0010] Therefore, the present invention aims to provide an appraisal system, program, and appraisal method in which the impact on external resources is not determined solely by the exercise of execution rights by the artificial intelligence (AI) agent acting as the subject, but is determined only based on the determination of a verification artificial intelligence (AI) model and definitive control rights separated from the subject, and the process of determination and confirmation is recorded in a way that makes tampering detectable. [Means for solving the problem]

[0011] To solve the above problems, an appraisal and certification system according to one aspect of the present invention is an appraisal and certification system for appraising and certifying a target act performed by an artificial intelligence (AI) agent, comprising a processor and a storage unit, and comprising an action request information acquisition unit, a hold control unit, a judgment result acquisition unit, an integrated judgment unit, a confirmation control unit, an execution control unit and a record control unit realized by the processor, wherein the action request information acquisition unit acquires action request information for a target act that an artificial intelligence (AI) agent, which is the acting entity, intends to perform on an external resource, including action identification information that identifies the target act, acting entity identification information that identifies the acting entity, and action content information that indicates the content of the target act, the hold control unit holds at least one of the execution of the target act or the determination of the impact of the target act on the external resource until confirmation that the confirmation permission information is valid is established, and the judgment result acquisition unit acquires a judgment result from each of a plurality of verification artificial intelligence (AI) models indicating whether the target act conforms to at least one of the scope of authority granted to the acting entity or the execution conditions attached to said authority, and at One is operated by an entity different from the entity that operates the acting entity, or operates in a management area different from the management area in which the acting entity operates, and the integrated determination unit integrates the acquired multiple determination results based on predetermined integration rules and generates an approval determination result indicating whether the result of the integration satisfies the pre-set approval conditions, and the confirmation control unit, when the approval determination result indicates approval, generates or sets to an effective state the confirmation permission information associated with the action request information and permits the confirmation of the impact on the external resources based on a confirmation control authority different from the execution authority of the target action granted to the acting entity, and does not generate or set to an effective state the confirmation permission information by the exercise of the execution authority by the acting entity alone, and does not generate or set to an effective state the confirmation permission information when the approval determination result indicates disapproval, and when confirmation is made that the confirmation permission information is valid and corresponds to the action request information, the execution control unit releases the hold and allows the target action to be executed or the impact on the external resources to be confirmed, and when confirmation is not made or the confirmation permission information does not exist,Without allowing the aforementioned target act to be performed or determining its effects, the record control unit generates authorization trail information including the act identification information, information corresponding to a plurality of the aforementioned determination results, and the authorization determination result, and records it in a record infrastructure capable of detecting tampering in association with the act identification information.

[0012] In another aspect of the present invention, the process of generating or setting the confirmed permission information to an active state is performed in a confirmed control area where access to the confirmed control authority is separated from both the management area where the acting entity operates and the management area where the multiple verification artificial intelligence (AI) models operate, and the acting entity is not granted the authority to generate, change to an active state, or restore the confirmed permission information from an active state. Another aspect of the present invention is a program for causing one or more computers to function as the above-mentioned parts, and yet another aspect of the present invention is an appraisal and certification method in which one or more computers perform the steps corresponding to the processing by the above-mentioned parts. [Effects of the Invention]

[0013] According to the present invention, the impact on external resources cannot be determined solely by the exercise of execution authority by the artificial intelligence (AI) agent acting as the principal investigator. Therefore, even if the principal investigator malfunctions, misleads, or infringes, it is possible to prevent the impact of actions that deviate from the scope of authority or execution conditions from being determined. In a configuration that integrates the judgment results of multiple verification artificial intelligence (AI) models, resilience to errors in a single model is obtained. In a configuration where at least one verification artificial intelligence (AI) model is operated by a different entity or in a different management domain than the principal investigator, failures or infringements on the principal investigator's side are prevented from spreading to the verification. Since the judgment, authorization, and confirmation processes are recorded as authorization trail information in a record infrastructure that can detect tampering, subsequent audits and verifications become possible.

[0014] Furthermore, according to each embodiment described later, effects such as guarantee of suspension by safety-side handling that treats the target action as unapproved when a determination result cannot be obtained, prevention of inconsistency caused by a state change between the time of reusing confirmed permission information and granting approval and the time of confirmation, control and collective revocation for delegation chains, continuous re-evaluation in response to changes in prerequisites after approval, and compensation for already confirmed impacts can be obtained. Brief Description of the Drawings

[0015] [Figure 1] This is an explanatory diagram showing an example of the overall configuration of an appraisal and certification system according to the present embodiment. [Figure 2] This is an explanatory diagram showing an example of the data structure of action request information. [Figure 3] This is an explanatory diagram showing an example of verification artificial intelligence registration information and determination results. [Figure 4] This is an explanatory diagram showing an example of integrated rules and approval conditions. [Figure 5] This is an explanatory diagram showing an example of the data structure of confirmed permission information. [Figure 6] This is an explanatory diagram showing an example of recording approval trail information. [Figure 7] This is a flowchart showing an example of the overall process of appraisal and certification. [Figure 8] This is a diagram showing an example of a sequence related to recording, generation of confirmed permission information, and impact confirmation. [Figure 9] This is an explanatory diagram showing an example of a pending mode. [Figure 10] This is a flowchart showing an example of a determination process related to a sequence of target actions. [Figure 11] This is a flowchart showing an example of a process for conditional delegation information and revocation of a delegation lineage. [Figure 12] This is a flowchart showing an example of processes for continuous re-evaluation, generation of state change information, and compensation. [Figure 13] This is an explanatory diagram showing an example of allowable impact range information, impact candidate information, and confirmed result information. [Figure 14]This is an explanatory diagram showing an example hardware configuration. [Figure 15] This is an explanatory diagram showing an example configuration of the management area and the deterministic control area. [Figure 16] This is an explanatory diagram showing an example of a modified configuration. [Modes for carrying out the invention]

[0016] Embodiments of the present invention will be described below with reference to the drawings. The present invention is not limited to the following embodiments, and various modifications are possible without departing from the spirit thereof. The following configurations, information, conditions, and processes can be combined with each other to the extent that they are not technically contradictory, and some of them may be omitted or replaced with other configurations having equivalent functions.

[0017] (First Embodiment: Overall Configuration) Figure 1 is an explanatory diagram showing an example of the overall configuration of the appraisal and certification system 1 according to this embodiment. The appraisal and certification system 1 is a system in which, with respect to a target action that an artificial intelligence (AI) agent, which is the acting entity 30, intends to perform on an external resource 40, one or more verification artificial intelligence (AI) models 60 determine whether the target action conforms to the scope of authority or execution conditions granted to the acting entity 30, and only if the result of the integration of the determination satisfies the authorization condition 121, it permits the execution of the target action or the determination of its impact based on a definitive control authority 138 different from the execution authority 139 of the acting entity 30, and records the process in a record base 50 that can detect tampering.

[0018] The authentication and certification system 1 comprises a processor 10 and a storage unit 20. The authentication and certification system 1 may consist of one or more computers and may be implemented as a server, cloud environment, on-premise environment, edge device, virtual machine, container, or a combination thereof. The processor 10 may include one or more CPUs (central processing units), GPUs, or other computing devices. The storage unit 20 may include memory, storage, databases, or external storage devices.

[0019] The processor 10 implements the action request information acquisition unit 81, the hold control unit 82, the determination result acquisition unit 83, the integrated determination unit 84, the confirmation control unit 85, the execution control unit 86, and the record control unit 87. Each unit may be implemented in a single device, or it may be implemented in a distributed environment operated by different devices, different management areas 190, or different entities. Some of the functions of each unit may be implemented through cooperation with external services or devices.

[0020] (Overview of each section) The action request information acquisition unit 81 acquires action request information 100 related to the target action. The hold control unit 82 holds the execution or determination of the impact of the target action. The judgment result acquisition unit 83 acquires the judgment result 110 from the verification artificial intelligence (AI) model 60. The integrated judgment unit 84 integrates the judgment results 110 to generate an authorization judgment result 123. The confirmation control unit 85 generates or sets confirmed permission information 130 to an active state based on the confirmation control authority 138. The execution control unit 86 releases the hold and determines the impact, conditional on the confirmation being successful. The recording control unit 87 records the authorization evidence information 140 on the recording base 50.

[0021] The acting entity 30 is an artificial intelligence (AI) agent that intends to perform the target action. The acting entity 30 may operate in the management area 190a of the entity that operates the acting entity 30. The external resource 40 is a resource that is the target or destination of the target action. A preceding device 41 (including a gateway, proxy, API infrastructure, or execution infrastructure) may be provided in front of the external resource 40, and the execution control unit 86 may be implemented inside the external resource 40 or in the preceding device 41.

[0022] The verification artificial intelligence (AI) models 60 (60a to 60n) are one or more artificial intelligence models that determine the suitability of the target action. At least one of the multiple verification artificial intelligence (AI) models 60 is operated by an entity different from the entity that operates the acting entity 30, or operates in a different management area 190 (e.g., management areas 190b to 190n) than the management area 190a in which the acting entity 30 operates.

[0023] The record infrastructure 50 is a record infrastructure capable of detecting tampering and may include one or more of a server 51, a distributed server 52, or a blockchain or other distributed ledger 53. The confirmation control unit 85 may be implemented in a confirmation control area 191, where access rights to the confirmation control authority 138 are separated from any of the management areas 190. Each component of the authentication certification system 1, the acting entity 30, the external resources 40, the verification artificial intelligence (AI) model 60, and the record infrastructure 50 may be connected to each other so as to be able to communicate with each other via a communication network 70.

[0024] (Boundaries of trust) In this embodiment, the management area 190a where the acting entity 30 operates, the management area 190 where each verification artificial intelligence (AI) model 60 operates, and the deterministic control area 191 are separated from each other by a boundary of trust. Communication between each area may be conducted via encrypted and mutually authenticated paths. The paths of authority, keys, and access are separated so that a failure, breach, or infringement in any one area does not directly disable the functionality of the other areas.

[0025] (Operating body) The appraisal and certification system 1 may be operated by the entity that operates the acting entity 30, or by an independent third-party organization, a consortium composed of multiple entities, or the operator of the data linkage platform. The neutrality of control is enhanced by having at least a portion of the definitive control domain 191 and the verification artificial intelligence (AI) model 60 operated by an independent entity.

[0026] (Second Embodiment Terminology) In this specification, “artificial intelligence (AI) agent” means an entity that uses one or more artificial intelligence models to autonomously plan, select, or execute actions toward external resources 40 based on objectives or instructions. An artificial intelligence (AI) agent may be implemented as software, or as a control entity mounted on a robot, vehicle, manufacturing equipment, or other device. An artificial intelligence (AI) agent may include mechanisms for calling external tools, mechanisms for generating plans, or mechanisms for coordinating with other artificial intelligence (AI) agents 32.

[0027] In this specification, "artificial intelligence model" may include machine learning models, deep learning models, decision tree models, random forest models, gradient boosting models, support vector machine models, Bayesian models, neural network models, graph neural network models, time series models, anomaly detection models, classification models, regression models, ranking models, rule-based models, combinations of machine learning models and machine learning models, or large-scale language models.

[0028] In this specification, “external resources” means resources that exist outside of the artificial intelligence (AI) agent acting as the subject 30 and that are the subject or destination of the target action. External resources 40 may include servers, APIs, databases, storage, files, network equipment, key management systems, record infrastructure, smart contracts, other information systems, other artificial intelligence (AI) agents 32, or external devices (including robots, manufacturing equipment, vehicles, medical devices, and other equipment).

[0029] In this specification, “Target Action” means an action that the agent 30 intends to perform on the external resource 40. The Target Action may include at least one of the following: connecting to the external resource 40, authentication, using a key, generating an electronic signature, issuing verifiable credentials (VC), transmitting data, retrieving data, changing the state of the external resource 40 (including changing settings, writing, updating, deleting, executing a transaction, making a payment, sending money, or committing), delegating authority or processing to another artificial intelligence (AI) agent 32, initiating another artificial intelligence (AI) agent 32, or controlling an external device.

[0030] In this specification, “impact” means any change in the state, transfer or availability of data, change in rights and obligations, or other effects that may occur in the external resource 40 as a result of the action in question. “Confirmation of impact” means that the impact becomes effective in the external resource 40, becomes difficult to cancel or roll back, or becomes accessible or available to third parties. Confirmation of impact may include committing in the database, making transmitted data decryptable, approving the transaction, outputting an electronic signature, or validating a delegation.

[0031] In this specification, “delegation” means that an agent 30 allows another artificial intelligence (AI) agent 32 to exercise or execute all or part of the authority or processing it holds. A chain of delegations can be formed when the other artificial intelligence (AI) agent 32 that has received the delegation further delegates (re-delegates).

[0032] In this specification, "suspended" means that the execution or determination of the subject act is suspended until confirmation that the confirmation permission information 130 is valid is established, and "release of suspension" means that the suspended execution or determination is resumed, conditional on the establishment of said confirmation.

[0033] In this specification, “managed area” means the scope of computing resources managed under a common administrative authority, authentication infrastructure, key management, network boundary, or operational entity. Different managed areas 190 may be separated in such a way that failures, breaches, or unauthorized actions are less likely to affect each other.

[0034] In this specification, “execution authority” means the authority granted to the agent 30 to perform the target action, which may be granted in the form of an authorization token, API key, credentials, or permission setting. “Determinative control authority” means an authority different from the execution authority 139, which is the authority to generate or set to an active state determinative authorization information 130 that permits the determination of the impact on the external resource 40. The agent 30 is not granted determinative control authority 138.

[0035] In this specification, "tamper-detectable record infrastructure" means a record infrastructure that has the property of being able to detect alterations to recorded information, and may include blockchain or other distributed ledgers 53, append-type records using hash chains, records with electronic signatures or timestamps, or a hybrid configuration that uses a record infrastructure utilizing distributed ledger technology in combination with a record infrastructure that does not utilize distributed ledger technology. The record may be a record of the main body of the information, or it may be a combination of a record of a digest of the information and separate storage of the main body.

[0036] In this specification, "digest" means a hash value or other value calculated from information that can be used to verify the identity of said information. In this specification, "fail close" means a safe approach that prevents the target action from being performed or the impact from being determined when the information necessary for determination or verification cannot be obtained or when there is other uncertainty.

[0037] (Notation used in this specification) In this specification and drawings, multiple elements of the same type are distinguished by a common number followed by a lowercase English letter subscript (e.g., verification artificial intelligence (AI) models 60a-60n, management areas 190a-190n), and when distinction is not required, the subscript is omitted and they are referred to collectively. The configurations shown in each figure are examples, and the number and arrangement of elements are not limited to those shown.

[0038] (Third Embodiment: Action Request Information 100) Figure 2 is an explanatory diagram showing an example of the data structure of the action request information 100. The action request information acquisition unit 81 acquires the action request information 100 for the target action that the artificial intelligence (AI) agent, which is the acting entity 30, intends to perform on the external resource 40. The action request information 100 includes action identification information 101, acting entity identification information 102, and action content information 103. The action request information 100 may further include external resource identification information 104, action type information 105, external resource type information 106, risk information 107, the time of the request, or information indicating the instructions or context that form the basis of the action.

[0039] (Action Identification Information 101) The action identification information 101 is information that identifies the target action. The action identification information 101 may be a unique identifier, a random number, a sequential number, or a digest of the action request information 100. The action identification information 101 is used to consistently refer to the same target action throughout the subsequent judgment, confirmation permission, and recording processes.

[0040] (Identification of the actor 102) The agent identification information 102 is information that identifies the agent 30. The agent identification information 102 may be an agent identifier, a distributed identifier (DID), a public key, a certificate, or an identifier for the execution environment.

[0041] (Act content information 103) The action content information 103 is information indicating the content of the target action. The action content information 103 may include the type of action, the target, arguments, the data to be transmitted or its digest, the content of the changes, or the content of the delegation. If the action content information 103 contains sensitive information, a digest of that information may be used instead of the main body of the information.

[0042] (External resource identification information 104) External resource identification information 104 is information that identifies the external resource 40. External resource identification information 104 may be an identifier, address, URL, name, or a digest thereof of the external resource 40.

[0043] (Activity type information 105 and external resource type information 106) The action type information 105 indicates the type of action, for example, whether it is transmission, modification, delegation, activation, or control. The external resource type information 106 indicates the type of external resource 40, for example, whether it is a database, API, recording infrastructure, device, or other artificial intelligence (AI) agent.

[0044] (Risk Information 107) Risk information 107 is information indicating the risk related to the target activity, and may be calculated or assigned based on the importance, irreversibility, amount, scope of impact, sensitivity, or history of past anomalies of the target activity. Risk information 107 may be calculated according to predetermined rules, or it may be calculated by an artificial intelligence model for risk calculation.

[0045] (Series 108) A sequence 108 is a collection of multiple target actions performed by the same actor 30 that are temporally consecutive or related in terms of purpose, target, or context. The action request information acquisition unit 81 may acquire a sequence 108 of multiple target actions. Details of the determination regarding the sequence 108 will be described later with reference to the ninth embodiment and Figure 10.

[0046] The action request information acquisition unit 81 may acquire the action request information 100 as a request from the acting entity 30, or it may acquire the action request information 100 by capturing a request for the target action in the path between the acting entity 30 and the external resource 40 (including the preceding device 41). In the latter configuration, the target action can be subject to control even if the acting entity 30 does not presuppose the existence of the appraisal certification system 1.

[0047] (Specific example of action request information 100) For example, if the target act is a money transfer, the act identification information 101 is a unique identifier issued for each request, the act subject identification information 102 is the DID of the act subject 30, the act content information 103 includes the recipient, amount, and currency, the external resource identification information 104 is the identifier of the settlement system that processes the money transfer, the act type information 105 is "money transfer", and the risk information 107 is the degree of risk based on the history of the amount and recipient.

[0048] (Information indicating the time and context of the request) The time of the request, which may be included in the action request information 100, is used for the certification, validity period management, and auditing of the series 108. The information indicating the instructions or context on which the action is based may be a summary or digest of instructions given to the acting entity 30, referenced information, or plan, and may be used as input for the determination by the verification artificial intelligence (AI) model 60.

[0049] (Completeness of action request information 100) The action request information 100 may be signed or summarized by the act subject 30 or the action request information acquisition unit 81. This makes it possible to detect any alteration of the action request information 100 between the judgment, confirmation permission, and recording processes.

[0050] (Association of action request information 100 belonging to series 108) Multiple action request information 100 belonging to series 108 may be associated by a common series identifier or by mutual reference. This association enables the calculation of cumulative values ​​in the ninth embodiment and the querying of series units in the authorization trail information 140.

[0051] (Fourth Embodiment: Hold Control) Figure 9 is an explanatory diagram showing an example of a hold mechanism. The hold control unit 82 holds either the execution of the target act or the determination of the impact of the target act on external resources 40 until confirmation that the confirmed permission information 130 is valid is established. The hold mechanism may be selected according to the nature of the target act.

[0052] In other words, if the target action is an action for which the determination of its effects can be postponed after execution, the hold control unit 82 may provisionally execute the target action and postpone the determination of its effects. If the target action is an action for which the determination of its effects cannot be postponed after execution, the hold control unit 82 may postpone the execution of the target action itself until confirmation that the confirmation permission information 130 is valid is obtained. Provisional execution may include executing a transaction in the database before commit, writing to the staging area, registering in a pending approval state, or transmitting data in an encrypted state.

[0053] (Cancellation of provisional execution) If the final authorization information 130 does not become valid after the provisional execution, the hold control unit 82 or the execution control unit 86 may cancel the provisional execution. Cancellation of the provisional execution may include rolling back the transaction, discarding the staging area, or deleting the registration awaiting approval.

[0054] (Managing pending status) The hold control unit 82 may manage the hold status (hold, awaiting confirmation, confirmed, disapproved, cancelled, etc.) of the hold target action in association with the action identification information 101. The hold status may be stored in the storage unit 20, or it may be recorded in the recording base 50 in association with the approval trail information 140.

[0055] (Release hold) The release of a hold is contingent only upon confirmation by the execution control unit 86. The hold control unit 82 and the acting entity 30 are not granted the authority to release a hold without confirmation. This centralizes the process for releasing holds and prevents circumvention of control.

[0056] (Holding by leaving the key) If the target action includes a process of transmitting data to an external resource 40, the hold control unit 82 may transmit the data to the external resource 40 in an encrypted format (encrypted data 181) that cannot be used without decryption information 180, and may hold the decryption information 180 in a state that cannot be used by the artificial intelligence (AI) agent which is the subject of the action 30. The decryption information 180 may be a decryption key, a key fragment, a key for encrypting the key, or a token that permits decryption, and may be held in the deterministic control area 191, a key management system (KMS), a hardware security module (HSM), or a trusted execution environment.

[0057] The execution control unit 86 makes the decryption information 180 available to the external resource 40 (including sending, disclosing, releasing the key, or performing decryption of the decryption information 180) if it confirms that the confirmed permission information 130 is valid. If the confirmation is not successful, it does not make the decryption information 180 available to the external resource 40. This allows the data transmission itself to proceed, while the availability of its contents is delayed until the impact is confirmed.

[0058] (Generation of conditional delegation information 150) If the target action involves the delegation of authority or processing from the acting entity 30 to another artificial intelligence (AI) agent 32, the hold control unit 82 may generate conditional delegation information 150, which includes delegate identification information 151, delegate identification information 152, delegation scope information 153, and information 154 identifying the parent delegation, and which remains in an invalid state until confirmation that the confirmed permission information 130 is valid is established. Details of the conditional delegation information 150 will be described later with reference to the eighth embodiment and Figure 11. The generation of the conditional delegation information 150 corresponds to step S301 in Figure 11.

[0059] (Fifth embodiment: Acquisition of verification artificial intelligence (AI) model 60 and judgment result 110) Figure 3 is an explanatory diagram showing an example of verification artificial intelligence registration information 117 and judgment result 110. The judgment result acquisition unit 83 acquires a judgment result 110 from each of one or more verification artificial intelligence (AI) models 60 indicating whether the target act conforms to at least one of the scope of authority granted to the acting entity 30 or the execution conditions attached to said authority. This specification mainly describes a configuration using multiple verification artificial intelligence (AI) models 60, but a configuration using only a single verification artificial intelligence (AI) model 60 may also be used depending on the required level of assurance, the risk of the target act, or operational constraints.

[0060] The validation artificial intelligence (AI) model 60 may include one or more of the AI ​​models exemplified in the second embodiment and may be generated, updated, evaluated, or selected using authority definitions, execution conditions, policies, past target actions and their results, or anomaly cases. The validation artificial intelligence (AI) model 60 may use information indicating the plan or rationale for the execution of the target action, or the results of a trial execution, in its determination.

[0061] (Scope of authority and execution conditions) The scope of authority may include the types, targets, limits on the amount or value of permitted actions, or the scope of permitted external resources 40. The execution conditions may include time limits, frequency limits, limits on the amount or value, prerequisite conditions, whether approval is required, or prohibited actions. The scope of authority and execution conditions may be stored in the storage unit 20 or retrieved from an external authority management system.

[0062] (Request for judgment and input information) The judgment result acquisition unit 83 may provide all or part of the action request information 100 to the verification artificial intelligence (AI) model 60 when a judgment request is made. For sensitive information not necessary for the judgment, a digest may be provided instead of the full information, and disclosure may be limited to the extent necessary for the judgment. The judgment request may include the deadline for the judgment, the scope and execution conditions of the applicable authority, or the perspective of the judgment.

[0063] When multiple verification artificial intelligence (AI) models 60 are used, at least one of them is operated by an entity different from the entity operating the acting entity 30, or operates in a different management area 190 than the management area 190a in which the acting entity 30 operates. This prevents any obstacles, infringements, or conflicts of interest on the part of the acting entity 30 from affecting the verification. All verification artificial intelligence (AI) models 60 may be operated by entities different from the acting entity 30, or operate in different management areas 190, and may be distributed across different entities or different management areas 190.

[0064] (Judgment result 110) The determination result 110 is information indicating whether the target action conforms to the scope of authority or the conditions for execution. The determination result 110 may include a binary determination indicating conformity or non-conformity, a score, conditional conformity, information indicating certainty, or information indicating the reason for the determination.

[0065] (Verification model identification information 111) The verification model identification information 111 is information that identifies the verification artificial intelligence (AI) model 60 that produced the judgment result 110. The verification model identification information 111 may include the model identifier and version information.

[0066] (Decision-making entity identification information 112) The determination entity identification information 112 is information that identifies the entity operating the verification artificial intelligence (AI) model 60 or the environment in which the determination was performed. The determination entity identification information 112 makes it possible to determine the similarities and differences of the entities in the quorum condition 122.

[0067] (Judgment time information 113) The determination time information 113 is information indicating the time when the determination was made. The determination time information 113 may be used to manage the validity period of the determination and to track changes between the time of authorization and the time of confirmation.

[0068] (Judgment result signature information 114) The judgment result signature information 114 is an electronic signature attached to the judgment result 110 using the signature key of the verification artificial intelligence (AI) model 60 or the entity that operates it. The judgment result signature information 114 makes the origin and completeness of the judgment result 110 machine verifiable.

[0069] (Decision authority information 115 and public key information 116) The judgment authority information 115 is information indicating the authority of the verification artificial intelligence (AI) model 60 to make judgments, and may include the type of action on which the judgment can be made, the type of external resource 40, its validity period, or whether it has expired. The public key information 116 is information of the public key used to verify the judgment result signature information 114.

[0070] (Verification AI registration information 117) The memory unit 20 may store verification artificial intelligence (AI) registration information 117 for each of the multiple verification artificial intelligence (AI) models 60, which associates verification model identification information 111, judgment authority information 115, and public key information 116. Registration, updating, and revocation of the verification artificial intelligence registration information 117 are performed by an authorized administrator or through administrative procedures, and the history thereof may be recorded in the recording base 50.

[0071] The judgment result acquisition unit 83 acquires the verification model identification information 111, the judgment subject identification information 112, the judgment time information 113, and the judgment result signature information 114 in association with the judgment result 110. Based on the verification artificial intelligence registration information 117, it verifies the validity of the judgment authority information 115 and the judgment result signature information 114. A judgment result 110 in which the judgment authority information 115 is valid and the verification of the judgment result signature information 114 is successful may be used for integration by the integration judgment unit 84. This prevents judgment results 110 made by models that do not have the authority to make judgments, or judgment results 110 that have been impersonated or tampered with from being mixed into the integration.

[0072] (Renewal and expiration of registration for Validated Artificial Intelligence (AI) Model 60) If the verified artificial intelligence (AI) model 60 is updated, the version information in the verified model identification information 111 will be updated, and the judgment authority information 115 may be reset as necessary. If there is a change in the operating entity, the model is discontinued, or fraud is detected, the judgment authority information 115 will expire, and the judgment results 110 after the expiration will not be used for integration.

[0073] (Determination of combinations and numbers of 60 validation artificial intelligence (AI) models) The judgment result acquisition unit 83 may determine the combination of verification artificial intelligence (AI) models 60 to acquire the judgment result 110 and the number of required judgment results 110 based on at least one of the action type information 105, external resource type information 106, or risk information 107. For example, it may be determined that a higher-risk action requires the use of a larger number of verification artificial intelligence (AI) models 60 belonging to a wider variety of entities or management domains 190. The integrated judgment unit 84 may apply the integrated rules 120 and authorization conditions 121 corresponding to the determined combination and number.

[0074] (Specific example of judgment result 110) For example, if the target act is a money transfer, the Verified Artificial Intelligence (AI) model 60a determines compliance or non-compliance from the perspective of amount and frequency, the Verified Artificial Intelligence (AI) model 60b from the perspective of the risk level of the recipient, and the Verified Artificial Intelligence (AI) model 60c from the perspective of consistency with the context of the instructions of the acting entity 30. Each determination result 110 may be accompanied by information indicating the degree of certainty and information indicating the reason for the determination.

[0075] (Independence of the decision) The judgments made by multiple verification artificial intelligence (AI) models 60 may be performed independently of each other. The judgment result acquisition unit 83 may suppress tracking and guidance between judgments by not providing the judgment result 110 of one verification artificial intelligence (AI) model 60 to other models.

[0076] (Validity period of judgment result 110) The judgment result 110 may have an expiration period set, and a judgment result 110 that has elapsed a predetermined time from the judgment time information 113 will be excluded from the input for integration, and a re-determination may be requested if necessary.

[0077] (Sixth Embodiment: Integrated Determination) Figure 4 is an explanatory diagram showing an example of the integration rule 120 and the authorization condition 121. The integration determination unit 84 integrates the acquired multiple determination results 110 based on the predetermined integration rule 120 and generates an authorization determination result 123 indicating whether the integration result satisfies the pre-set authorization condition 121. In a configuration using a single verification artificial intelligence (AI) model 60, the integration determination unit 84 may generate the authorization determination result 123 based on the determination results 110 of the single model.

[0078] (Integration input) The input for integration consists of multiple judgment results 110 that have passed verification, and verification model identification information 111, judgment subject identification information 112, judgment time information 113, and information indicating the degree of certainty associated with each judgment result 110 may be used together. Judgment results 110 that have exceeded a predetermined validity period may be excluded from the input for integration.

[0079] (Integration Rule 120) The integration rule 120 is a rule for integrating multiple decision results 110. The integration rule 120 may include logical AND, logical OR, majority vote, weighted majority vote, weighted average, score synthesis, adoption of minimum or maximum value, adoption of median, average after exclusion of outliers, or graded evaluation. The integration rule 120 may be definitively defined to give the same integrated result for the same multiple decision results 110.

[0080] (Authorization Condition 121) Approval condition 121 is a condition that determines whether or not to approve the target action based on the results of the integration. Approval condition 121 may include that the value indicated by the results of the integration is above a threshold, that the required number of conformity determinations have been obtained, or that conformity determinations by a specific verification artificial intelligence (AI) model 60 have been included. Approval condition 121 may be switched depending on the action type information 105, the external resource type information 106, or the risk information 107.

[0081] (Quorum condition 122) The authorization condition 121 may include a quorum condition 122 that requires that a predetermined determination result 110 (e.g., a determination result 110 indicating compliance) be obtained by at least two verification artificial intelligence (AI) models 60 operated by different entities or operating in different management domains 190. The final control unit 85 does not generate or set the final authorization information 130 to a valid state if the quorum condition 122 is not met.

[0082] (Specific example of quorum condition 122) For example, quorum condition 122 may be defined as "a determination result 110 indicating conformity is obtained from at least two of the verification artificial intelligence (AI) models 60 operated by three different entities." The number required by quorum condition 122 may be three or more, and quorum condition 122 may also include conditions relating to combinations of entities or management domains 190 (for example, including at least one entity that does not have a capital relationship with the entity operating the acting entity 30).

[0083] (Authorization decision result 123) The approval decision result 123 is information indicating whether the merger result satisfies the approval condition 121, i.e., whether it is approved or not. The approval decision result 123 may be associated with information identifying the merger rule 120 or approval condition 121 applied as a result of the merger, or the time of the decision.

[0084] (Specific examples of integration and authorization) For example, if the three judgment results 110 are conforming (confidence 0.9), conforming (confidence 0.8), and non-conforming (confidence 0.6), respectively, and the integration rule 120 is a weighted majority vote based on confidence, and the authorization condition 121 is "the result of the weighted majority vote is conforming, and the quorum condition 122 (two or more conforming by different entities) is satisfied," then the integration judgment unit 84 calculates conforming as the result of the integration, confirms that the quorum condition 122 is satisfied, and generates an authorization judgment result 123 indicating authorization.

[0085] (Management of Integration Rule 120 and Authorization Condition 121) The establishment and modification of the Integration Rule 120 and the Authorization Condition 121 can only be carried out by an authorized administrator or through administrative procedures, and the history thereof may be recorded in the record base 50. The acting entity 30 is not authorized to modify the Integration Rule 120 and the Authorization Condition 121.

[0086] (Information on acceptable impact range 124) Figure 13 is an explanatory diagram showing an example of permissible impact range information 124, candidate impact information 170, and confirmed result information 171. The integrated determination unit 84 may generate permissible impact range information 124 indicating the range within which the determination of the impact on the external resource 40 is permissible, based on a plurality of determination results 110 or the result of integration. The permissible impact range information 124 may include upper limits on the amount, quantity, or number of items, limitations on the target, limitations on the period, or other associated conditions. The confirmation control unit 85 may associate the confirmed permission information 130 with the permissible impact range information 124. The execution control unit 86 determines the impact only within the range indicated by the permissible impact range information 124 and does not determine the impact beyond that range. This makes partial approval possible, rather than having to choose between approving all or not approving all of the target action.

[0087] (7th Embodiment: Deterministic Control) Figure 5 is an explanatory diagram showing an example of the data structure of the confirmed permission information 130. When the approval determination result 123 indicates approval, the confirmed control unit 85 generates or sets to an active state the confirmed permission information 130, which is associated with the action request information 100 and permits the confirmation of the impact on the external resource 40, based on a confirmed control authority 138 that is different from the execution authority 139 of the target action granted to the acting entity 30. The confirmed permission information 130 is not generated or set to an active state by the acting entity 30's exercise of the execution authority 139 alone. When the approval determination result 123 indicates disapproval, the confirmed control unit 85 does not generate or set to an active state the confirmed permission information 130.

[0088] (Relationship between execution authority 139 and deterministic control authority 138) The execution authority 139 is the authority to initiate the execution of the target act and is granted to the acting entity 30. The determination control authority 138 is the authority to permit the determination of the effects and is not granted to the acting entity 30, but is exercised in the determination control area 191. Because the two authorities are separated, even if the execution authority 139 is stolen or abused, it will not lead to the determination of the effects.

[0089] (Final permission information 130) The definitive authorization information 130 is information that authorizes the determination of the impact on the external resource 40. The definitive authorization information 130 may be newly generated, or it may be set to an active state from an existing invalid state. The definitive authorization information 130 may be in the form of a token, signed data, a state on the recording infrastructure 50, or a flag.

[0090] (Retention and reference of confirmed permit information 130) The confirmed authorization information 130 may be stored in the storage area of ​​the confirmed control area 191, the recording infrastructure 50, or a storage area accessible from the execution control unit 86. In any configuration, the generation of the confirmed authorization information 130, the change to an active state, and the return from an inactive state are performed only as processes based on the confirmed control authority 138.

[0091] (Definite control region 191) Figure 15 is an explanatory diagram showing an example configuration of the management area 190 and the deterministic control area 191. The deterministic control unit 85 may be implemented in the deterministic control area 191, where access rights to the deterministic control authority 138 are separated from either the management area 190a where the artificial intelligence (AI) agent, which is the acting entity 30, operates, or the management area 190 where the multiple verification artificial intelligence (AI) models 60 operate. Similarly, the process of generating or setting the deterministic permission information 130 to a valid state may be executed in the deterministic control area 191. The deterministic control area 191 may be implemented by a dedicated server, a partition of an isolated network, a hardware security module, a key management system, a trusted execution environment, or an environment operated by an independent entity.

[0092] The artificial intelligence (AI) agent, which is the acting entity 30, is not granted the authority to generate confirmed permission information 130, change it to a valid state, or revert it from an invalid state. As a result, the confirmed permission information 130 is not generated or set to a valid state by the control in the management area 190a on which the acting entity 30 operates. The confirmed permission information 130 may be kept accessible from the external resource 40 or the execution control unit 86 (execution control processing) located upstream of the external resource 40, in a state in which the acting entity 30 is not permitted to change it. "Upstream" means located on the path from the acting entity 30 to the external resource 40, on the side of the acting entity 30 that is closer to the acting entity 30 than to the external resource 40.

[0093] (Generation of confirmed permit information 130 based on the recording of authorization trail information 140) The configuration may be such that the action request information 100 includes external resource identification information 104, the recording control unit 87 records the authorization evidence information 140 on the recording base 50, and after obtaining recording completion information 141 indicating the completion of the recording, the confirmation control unit 85 generates confirmed permission information 130. This order (permission after recording) structurally guarantees that there is always evidence corresponding to the confirmed impact.

[0094] (Control signature generation key 135, control signature target information 136, and control signature 137) The confirmation control unit 85 may, when the authorization determination result 123 indicates authorization and the recording completion information 141 is obtained, use the control signature generation key 135 associated with the confirmation control authority 138 to generate confirmation authorization information 130 by attaching a control signature 137 to control signature target information 136, which includes act identification information 101, act subject identification information 102, a digest of act content information 103, external resource identification information 104, a digest of authorization trail information 140, and the recording completion information 141 or a digest of the recording completion information 141. The control signature generation key 135 may be held in the confirmation control area 191, hardware security module, key management system, or trusted execution environment in a state where it is not output externally.

[0095] (Providing information for verifying control signature 137) Information regarding the verification key for verifying the control signature 137 is provided for reference by the execution control unit 86. The execution control unit 86 determines the effect only if the verification of the control signature 137 and the correspondence between the control signature target information 136 and the action request information 100 are successful. This allows the execution control unit 86 to machine-verify the authenticity of the determined permission information 130 without requiring a query to the determined control area 191.

[0096] (Treated as denial) If a predetermined time has elapsed to obtain the judgment result 110 (e.g., no response within the predetermined time, an error response), if the number of judgment results 110 required for the authorization condition 121 cannot be obtained, or if the validity of the obtained judgment result 110 cannot be confirmed, the confirmation control unit 85 will treat the target act as unauthorized and will not generate or set the confirmation authorization information 130 to a valid state. In uncertain situations such as communication failures, shutdown of the verification artificial intelligence (AI) model 60, or failure to verify the judgment result signature information 114, a safe state where the impact is not determined, i.e., a fail-close, is guaranteed.

[0097] (Expiration date information 131) Expiration date information 131 indicates the validity period during which the determination of the impact is permitted. Determination of the impact based on the determination permission information 130 after the expiration of the validity period is not permitted. After the expiration of the validity period, the target action may be re-requested and re-evaluated as necessary.

[0098] (Re-request after the expiration of the validity period) If the subject act 30 requests the same subject act again for an act that could not be finalized due to the expiration of its validity period, new act identification information 101 will be assigned, and the judgment, approval, and final permission processes will be executed again. Expired final permission information 130 will not be reused.

[0099] (One-time identification information 132) The one-time identification information 132 is identification information whose number of uses is limited to one, and may be a nonce or a one-time token. The status of the one-time identification information 132 (unused or used) may be managed in a storage area or recording substrate 50 that can be referenced and updated by the execution control unit 86.

[0100] (Status version 133 and commit identification information 134) The state version 133 may be a number, sequence number, or state digest indicating the version of the state of the external resource 40. The information identifying the commit 134 is information identifying the commit of the processing in the external resource 40, and may be a transaction identifier. By including this information in the final authorization information 130, it becomes possible to confirm that the state or processing of the external resource 40 assumed at the time of authorization is maintained at the time of final authorization.

[0101] The execution control unit 86 confirms that the validity period indicated by the expiration information 131 is within the validity period, that the one-time identification information 132 is unused, and that the state version 133 or information identifying the commit 134 corresponds to the current state or process of the external resource 40. It then updates the one-time identification information 132 from unused to used and determines the impact on the external resource 40, all within the same transaction or an exclusively controlled process. The exclusively controlled process may include acquiring a lock, atomic updating by compare and swap, or equivalent processes. This prevents the reuse (replay) of the confirmed authorization information 130 and inconsistencies caused by changes in the state of the external resource 40 between the time of authorization and the time of confirmation.

[0102] (Invalidation of confirmed permission information 130) The confirmation control unit 85 can change the valid confirmation permission information 130 to an invalid state based on the confirmation control authority 138. Invalidation may be performed based on the results of re-evaluation in the tenth embodiment, instructions from the administrator, detection of an abnormality in the acting entity 30, or the expiration of the validity period. Confirmation of the impact based on the invalidated confirmation permission information 130 is not permitted.

[0103] (Specific example of confirmed permit information 130) For example, the confirmed authorization information 130 for the transaction subject to the remittance includes, as control signature target information 136, a digest of the transaction identification information 101, the entity identification information 102, the transaction content information 103 (recipient, amount), the external resource identification information 104 (settlement system identifier), the authorization trail information 140, the record completion information 141, the expiration date information 131 (e.g., 5 minutes from issuance), the one-time identification information 132, and the information 134 identifying the commit, and is configured as a signed token to which a control signature 137 is attached.

[0104] (Eighth Embodiment: Execution Control) The execution control unit 86 releases the hold and executes the target action or determines the impact on the external resource 40 if it confirms that the confirmed permission information 130 is valid and corresponds to the action request information 100. If the confirmation is not successful or if the confirmed permission information 130 does not exist, the execution control unit 86 does not execute the target action or determine the impact.

[0105] (Specific examples of confirming the response) The confirmation of the correspondence may be performed by checking for a match in the action identification information 101, checking for a match in the digest of the action content information 103, or verifying the control signature 137 (e.g., matching the hash value). For example, the execution control unit 86 checks for a match between the digest calculated from the action content information 103 of the pending target action and the digest included in the control signature target information 136, and if they do not match, it does not determine the impact.

[0106] (Execution control in the preceding device 41) The execution control unit 86 may be located inside the external resource 40, in the preceding device 41 of the external resource 40, or in the definitive control area 191. In the configuration where it is located in the preceding device 41, the existing external resource 40 can be made subject to control without modifying the external resource 40.

[0107] (Handling of cases where verification is not successful) If confirmation is not successful, the execution control unit 86 may not allow the target action to be performed or its effects to be determined, and may record that fact in the recording base 50 in association with the authorization trail information 140. The execution control unit 86 may notify the acting entity 30 or the administrator of the reason for the failure of confirmation (e.g., expiration of validity period, use of one-time identification information 132, mismatch of status version 133).

[0108] (Behavior in the event of a failure in the execution control unit 86) If the execution control unit 86 fails and is unable to perform the verification, the impact will not be determined. In other words, the failure of the execution control unit 86 will manifest not as an incorrect determination, but as a suspension of determination. After recovery, verification and determination will resume only if valid determination permission information 130 exists.

[0109] (Potential impact information 170 and confirmed result information 171) The execution control unit 86 acquires candidate impact information 170 indicating the impact to be determined in the external resource 40, and may determine the impact if the candidate impact information 170 corresponds to the content of the target action indicated in the action request information 100 or the determination permission information 130. The candidate impact information 170 may include the result of the provisional execution, the difference scheduled to be written, or a digest of the transmitted data. After determining the impact, the execution control unit 86 acquires determination result information 171 indicating the determined impact in the external resource 40, and may record the determination result information 171 in the recording base 50 in association with the authorization trail information 140. This makes it possible to verify afterward whether the authorized content matches the actually determined content.

[0110] (Conditional delegation information 150) Conditional delegation information 150 represents the content of the delegation, but remains invalid until confirmation that the valid authorization information 130 is valid is obtained. While conditional delegation information 150 is invalid, the other artificial intelligence (AI) agent 32 to whom the delegation is granted cannot exercise the delegated authority.

[0111] (Delegator identification information 151, delegate identification information 152, delegation scope information 153, and information identifying the parent delegation 154) Delegator identification information 151 is information that identifies the artificial intelligence (AI) agent (including the acting entity 30) that is delegating. Delegator identification information 152 is information that identifies the other artificial intelligence (AI) agent 32 that is receiving the delegation. Delegation scope information 153 may include the type, subject, limit, duration, or whether sub-delegation is permitted for the delegated authority or process. Parent delegation identification information 154 is information that identifies the higher-level delegation that is a prerequisite for the delegation, thereby enabling the tracking of the delegation chain 156 from delegation to sub-delegation.

[0112] (Application of the constraints of delegation scope information 153) If another artificial intelligence (AI) agent 32 delegated to the system attempts to perform a target action based on the delegated authority, that target action may also be subject to the suspension, judgment, and confirmation control by the appraisal and certification system 1. In this case, the judgment result acquisition unit 83 and the integrated judgment unit 84 may use the delegated scope information 153 as part of the scope of authority or execution conditions for the judgment.

[0113] (Activation of conditional delegation information 150 and identification and revocation information 155 of the delegation system 156) The execution control unit 86 may change the conditional delegation information 150 to a valid state if it has confirmed that the confirmed authorization information 130 is valid. If the confirmed authorization information 130 later becomes invalid or if the information on which the authorization determination result 123 is based is changed, the execution control unit 86 may trace the conditional delegation information 150 to identify the delegation system 156 originating from the conditional delegation information 150 and generate invalidation information 155 that invalidates the delegation included in the identified system 156. As shown in Figure 11, the execution control unit 86 determines whether the confirmation is successful (step S302), and if successful, changes the conditional delegation information 150 to a valid state (step S303). Subsequently, if invalidation of the confirmed authorization information 130 or a change in the information on which it was based is detected (step S304), the delegation system 156 is identified (step S305), and invalidation information 155 is generated (step S306). The revocation information 155 may be recorded in the record base 50 or notified to the relevant entities or systems, including the delegates. This prevents dilution of control through the chain of delegations and ensures that the revocation of an upstream delegation has a ripple effect on downstream delegations.

[0114] (Ninth Embodiment: Series of Target Actions 108 and Cumulative Impact) Figure 10 is a flowchart showing an example of the processing for determining a series of target actions 108. Even if individual target actions are small and fall within the scope of authority, their accumulation may cause the overall process to deviate from the scope of authority or the conditions for execution. This embodiment prevents such circumvention of control through fragmentation.

[0115] (Acquisition of Series 108) The action request information acquisition unit 81 acquires a series 108 of multiple target actions performed by the artificial intelligence (AI) agent, which is the acting entity 30 (step S201). The determination of the series 108 may be based on the same acting entity identification information 102, temporal proximity, identity or relevance of the target external resource 40, or commonality of the instructions or context that form the basis of the actions.

[0116] (Condition 109 for cumulative effects) The conditions 109 for cumulative impact on external resources 40 may include at least one of the following: a condition relating to the number of times the target action is performed on the same external resource 40 within a predetermined period; a condition relating to the cumulative amount of changes to the external resource 40 or the amount of data transmitted to the external resource 40 due to the target actions included in the series 108; or a condition relating to the depth of the chain of authorizations or delegations of processing to other artificial intelligence (AI) agents 32 included in the series 108. In addition, the conditions 109 for cumulative impact may include conditions relating to the cumulative amount, conditions relating to the expansion of the scope of the target, conditions relating to the frequency of execution, or conditions relating to combinations of target actions that span multiple interrelated external resources 40. The thresholds and periods for each condition may be set according to the action type information 105, the external resource type information 106, or the risk information 107.

[0117] The determination result acquisition unit 83 determines whether the sequence 108 of multiple target actions satisfies the condition 109 for cumulative effects (step S202). Even if each of the multiple target actions is within the scope of the authority granted to the act subject 30, if the sequence 108 satisfies the condition 109 for cumulative effects (step S202: YES), the determination result acquisition unit 83 acquires a determination result 110 from the verification artificial intelligence (AI) model 60 as a single target action (step S203). If it does not satisfy the condition (step S202: NO), the determination result 110 may be acquired for each individual target action (step S204). Subsequent integration, confirmation control, and execution control are performed in the same manner as in the sixth to eighth embodiments (step S205).

[0118] (Determination as a single, unified act) If the sequence 108 is determined to be a single target act, the act content information 103 may include a set of contents for each target act included in the sequence 108, as well as cumulative values ​​for the number of executions, the amount of change, the amount of data, and the depth of the delegation chain. The confirmed permission information 130 may be generated singlely for the entire sequence 108, or it may be generated for each target act included in the sequence 108.

[0119] (Partial confirmation for Series 108) When permissible impact range information 124 is generated for sequence 108, the execution control unit 86 determines the impact only for the portion of the target actions included in sequence 108 that falls within that range, and does not determine the impact for the remaining target actions that exceed the range. The breakdown of the partial determination is recorded as determination result information 171.

[0120] (Specific example from Series 108) For example, if the scope of authority is "transfers of 100,000 yen or less per transaction," and the agent 30 attempts to make 20 transfers of 90,000 yen each in a short period of time, each individual transaction falls within the scope of authority. However, the total transfer of 1,800,000 yen is determined as a single transaction, satisfying the conditions regarding the cumulative amount within a specified period (series 108). This prevents circumvention of the upper limit.

[0121] (Tenth embodiment: Continuous re-evaluation, status change information 160, and compensation) Figure 12 is a flowchart illustrating an example of continuous re-evaluation, generation of status change information 160, and compensation processing. Even if a subject action was determined to be compliant at the time of approval, compliance may be lost due to subsequent changes in the underlying assumptions. This embodiment ensures that control continues even after approval.

[0122] (Detecting changes) After generating the confirmed authorization information 130 or determining its impact, the integrated determination unit 84 detects a change in at least one of the following: the execution environment of the artificial intelligence (AI) agent that is the agent 30, the authority granted to the agent 30, the execution conditions, or the information on which the determination result 110 is based (step S401). The execution environment includes the software, model, configuration, or infrastructure on which the agent 30 operates. Detection of the change may be performed by receiving a notification, periodic inquiries, or monitoring the record base 50 or other information sources. Examples of changes include updating the model of the agent 30, reducing its authority, revising the execution conditions, detecting an indication of breach of the agent 30, or revoking the determination authority information 115 of the verified artificial intelligence (AI) model 60.

[0123] The integrated determination unit 84 identifies the target act or confirmed permission information 130 that will be affected by the change based on the authorization trail information 140 (step S402), and performs a re-evaluation of the identified target act using multiple verification artificial intelligence (AI) models 60 (step S403). If the result of the re-evaluation satisfies the authorization condition 121 (step S404: YES), the existing confirmed permission information 130 may be maintained (step S405).

[0124] (Frequency and scope of re-evaluation) Re-evaluation may be triggered by the detection of a change, performed at predetermined intervals, or preferentially for target actions with high risk information 107. Depending on the nature of the change, the scope of re-evaluation may be limited to target actions related to a specific actor 30, target actions related to a specific external resource 40, or target actions based on the judgment result 110 of a specific verification artificial intelligence (AI) model 60.

[0125] (Status change information 160) If the result of the re-evaluation does not satisfy the authorization condition 121 (step S404: NO), the finalization control unit 85 changes the finalization authorization information 130 that has not yet been used to determine the impact to an invalid state (step S406), and generates status change information 160 indicating non-conformity or expiration for the target activity whose impact has already been determined (step S407). The status change information 160 may be recorded in the recording infrastructure 50 or notified to the relevant entity or system. As a result, even for target activities that were compliant at the time of authorization, control continues in accordance with the change in assumptions.

[0126] (Compensation action 161 and compensation control information 162) When state change information 160 is generated, the execution control unit 86 may identify a compensatory action 161 that cancels, reduces, disables, or replaces an already determined effect with another effect, and output compensatory control information 162 to execute or cause the external resource 40 to execute the compensatory action 161 (step S408). The compensatory action 161 may include a reverse transaction, a refund, deletion of data, recording of invalidity, revocation of authority, or issuance of corrective information. The execution control unit 86 records the result of the compensatory action 161 in the recording base 50 in association with the authorization trail information 140 (step S409).

[0127] Compensation act 161 itself may be subject to the appraiser certification system 1's pending, judgment, and final control. In this case, action request information 100 is also generated for compensation act 161, and it is executed after judgment by the verification artificial intelligence (AI) model 60.

[0128] (11th Embodiment: Authorization Evidence Information 140 and Records) Figure 6 is an explanatory diagram showing an example of the recording of authorization trail information 140. The recording control unit 87 generates authorization trail information 140 which includes action identification information 101, information corresponding to a plurality of judgment results 110, and authorization judgment result 123, and records it in the recording base 50 which is capable of detecting tampering in association with the action identification information 101. The information corresponding to the judgment result 110 may include the judgment result 110 itself, a digest of the judgment result 110, verification model identification information 111, judgment subject identification information 112, judgment time information 113, or judgment result signature information 114.

[0129] (Composition of authorization trail information 140) The authorization trail information 140 may further include the action request information 100 or a digest thereof, the result of the integration, information identifying the applied integration rule 120 or authorization condition 121, information identifying the confirmed authorization information 130, or time information. In a configuration using a single verification artificial intelligence (AI) model 60, the authorization trail information 140 may include information corresponding to the judgment result 110 of the single model. Even in the case of denial, the judgment result 110 that led to denial and the judgment result 123 that led to authorization are recorded as authorization trail information 140, enabling subsequent audits.

[0130] (Recording complete information 141) The recording control unit 87 may acquire recording completion information 141 indicating the completion of recording to the recording infrastructure 50. The recording completion information 141 may include the transaction identifier, block identifier, confirmation of receipt, or a digest thereof in the recording infrastructure 50. The execution control unit 86 or the auditing entity can use the recording completion information 141 to verify the existence and content of the record in the recording infrastructure 50.

[0131] (Hybrid record) Recording may be performed using a hybrid configuration that combines recording to server 51 or distributed server 52 with recording to distributed ledger 53. In this case, the main body of the authorization trail information 140 may be recorded on server 51 or distributed server 52, and a digest thereof may be recorded on distributed ledger 53. Tampering can be detected by comparing the main body record with the digest record.

[0132] (Inquiry and audit of records) The auditing entity can use the action identification information 101 as a key to query the authorization trail information 140, confirmed result information 171, status change information 160, and revocation information 155 recorded in the record base 50, and verify in chronological order what kind of impact was determined based on what kind of judgment. The source and completeness of the trail can be machine-verified by verifying the recorded judgment result signature information 114 and control signature 137.

[0133] (Time information) The authorization trail information 140 and the record completion information 141 may be accompanied by time information. The time information may be accompanied by a timestamp or corroborated by the order of recording in the recording base 50, thereby enabling the chronological relationship of judgment, authorization, recording, and confirmation to be proven retrospectively.

[0134] (12th Embodiment: Overall Processing Procedure) Figure 7 is a flowchart showing an example of the overall processing of an expert certification. First, the action request information acquisition unit 81 acquires action request information 100 for the target action that the act subject 30 intends to perform on the external resource 40 (step S101).

[0135] Next, the hold control unit 82 holds either the execution of the target act or the determination of the impact of the target act on the external resources 40 until confirmation that the confirmed permission information 130 is valid is obtained (step S102). The form of the hold may be selected from holding the execution or holding the determination after provisional execution, depending on the nature of the target act, as in the fourth embodiment.

[0136] Next, the judgment result acquisition unit 83 determines the combination and required number of verification artificial intelligence (AI) models 60 that acquire the judgment result 110 based on the action type information 105, the external resource type information 106, or the risk information 107 (step S103).

[0137] Next, the judgment result acquisition unit 83 requests a judgment for each determined verification artificial intelligence (AI) model 60 (step S104). The judgment request may include all or part of the action request information 100, the scope of applicable authority, and the execution conditions.

[0138] Next, the judgment result acquisition unit 83 acquires judgment results 110 from each verification artificial intelligence (AI) model 60 and verifies the validity of the judgment authority information 115 and the judgment result signature information 114 based on the verification artificial intelligence registration information 117 (step S105).

[0139] Next, the judgment result acquisition unit 83 determines whether the required number of valid judgment results 110 for the authorization condition 121 have been obtained within a predetermined time (step S106). If they are not obtained (step S106: NO), the confirmation control unit 85 treats the target act as unapproved and does not generate or set the confirmed permission information 130 to a valid state (step S112), and the process proceeds to step S110.

[0140] If the required number of valid determination results 110 are obtained (step S106: YES), the integration determination unit 84 integrates the multiple determination results 110 based on the integration rule 120 and generates an authorization determination result 123 indicating whether the result of the integration satisfies the authorization conditions 121 (which may include quorum conditions 122) (step S107).

[0141] Next, the process branches depending on whether the authorization determination result 123 indicates authorization or not (step S108). If the authorization determination result 123 indicates non-authorization (step S108: NO), the confirmation control unit 85 does not generate or set the confirmation authorization information 130 to a valid state (step S112), and the process proceeds to step S110.

[0142] If the authorization determination result 123 indicates authorization (step S108: YES), the record control unit 87 generates authorization trail information 140 and records it in the record base 50 and obtains record completion information 141. The finalization control unit 85 generates finalization authorization information 130 with a control signature 137 attached, or sets it to a valid state, based on the finalization control authority 138 (step S109).

[0143] Next, the recording control unit 87 reflects the results of the processing, including whether or not the approval determination result 123 and the confirmed permission information 130 have been generated, in the approval evidence information 140 and records it on the recording base 50 (step S110).

[0144] Next, the execution control unit 86 verifies that the confirmed authorization information 130 is valid and corresponds to the action request information 100 (including verification of the control signature 137, the expiration date information 131, the one-time identification information 132, and the status version 133 or the information 134 identifying the commit). If this verification is successful, the hold is released, and the target action is executed or the impact on the external resource 40 is determined (step S111). If this verification is not successful or the confirmed authorization information 130 does not exist, the execution control unit 86 does not execute the target action or determine the impact. This completes the series of processes.

[0145] In the event of denial (procedure via step S112), the suspended target action will not be executed. If a provisional execution has been performed, it will be canceled as described in the fourth embodiment. If conditional delegation information 150 has been generated, it will remain in an invalid state or be discarded.

[0146] (Procedure variation) The procedure in Figure 7 may be performed by incorporating the determination regarding series 108 in Figure 10 (steps S201 to S205) before step S103, and the re-evaluation in Figure 12 (steps S401 to S409) may be performed continuously after step S111. Furthermore, if the target action includes delegation, the release of the hold in step S111 is performed as step S303 in Figure 11 (activation of conditional delegation information 150).

[0147] (13th Embodiment Sequence) Figure 8 shows an example of a sequence related to the generation of records, confirmed permission information 130, and the determination of their impact. This sequence illustrates the interaction between the acting entity 30, the expert certification system 1 (decision result acquisition unit 83, integrated decision unit 84, and record control unit 87), the confirmed control area 191 (confirmation control unit 85), the record base 50, the execution control unit 86, and the external resources 40.

[0148] First, the request for the target action by the acting entity 30 is acquired as action request information 100, and the action is set to be held. Next, the judgment results 110 from multiple verification artificial intelligence (AI) models 60 are acquired and integrated, and if the approval judgment result 123 indicates approval, the record control unit 87 records the approval evidence information 140 on the record base 50, and the record completion information 141 is returned to the confirmation control unit 85.

[0149] Next, the confirmation control unit 85 generates confirmation authorization information 130 with a control signature 137 attached using the control signature generation key 135, assuming that it has obtained the recording completion information 141. The execution control unit 86 verifies the control signature 137, confirms the correspondence between the control signature target information 136 and the action request information 100, and performs the confirmation of unused one-time identification information 132 and its update to used, as well as the determination of the impact on the external resource 40, within the same transaction or an exclusively controlled process.

[0150] Finally, the confirmed result information 171 is obtained and recorded in the recording base 50 in association with the authorization trail information 140. This sequence structurally ensures that there is no confirmation without recording, and that the confirmed authorization information 130 is consumed only once.

[0151] (14th Embodiment Hardware Configuration) Figure 14 is an explanatory diagram showing an example of hardware configuration. The computer constituting the appraisal and certification system 1 may include a processor 10, main memory, auxiliary storage, a communication interface, and input / output devices as needed. The storage unit 20 may be implemented by main memory, auxiliary storage, or an external storage device.

[0152] Each of the above parts and each of the processes described later may be realized by the processor 10 reading and executing a program stored in the memory unit 20, and some of these may be realized by dedicated circuits. The communication interface is used for communication with the acting entity 30, external resources 40, verification artificial intelligence (AI) model 60 and recording base 50 via the communication network 70.

[0153] (15th Embodiment: Application Example) As an example of application, we will explain the application to financial transactions. An artificial intelligence (AI) agent, which is the acting entity 30, is entrusted with managing funds and attempts to execute a transfer (target action) to a settlement system (external resource 40). The action request information acquisition unit 81 acquires action request information 100 including the recipient, amount, and currency, and the hold control unit 82 postpones determining the impact by registering it in a state of awaiting approval in the settlement system (provisional execution).

[0154] The judgment result acquisition unit 83 selects three verification artificial intelligence (AI) models 60a, 60b, and 60c based on risk information 107 (amount and risk level of the recipient), and obtains judgment results 110 from each to verify the signature and authority. The integrated judgment unit 84 generates an authorization judgment result 123 indicating authorization based on weighted majority voting (integrated rule 120) and quorum conditions 122 (2 or more conformances by different entities).

[0155] The recording control unit 87 records the authorization trail information 140 in the recording base 50 and obtains recording completion information 141. The finalization control unit 85 generates finalization authorization information 130 (including expiration date information 131, one-time identification information 132, and information 134 identifying the commit) with a control signature 137 in the finalization control area 191. The execution control unit 86 performs verification in the upstream device 41 of the settlement system, executes the consumption of the one-time identification information 132 and the approval of the transfer in the same transaction, and records the finalization result information 171. If the acting entity 30 attempts to transfer funds beyond the scope of its authority due to misdirection, the authorization judgment result 123 will be denied, the registration awaiting approval will be canceled, and the impact will not be finalized.

[0156] As an example of application, we will explain the application to a data linkage platform. The acting entity 30 attempts to transmit sensitive data (target action) to the linked system (external resource 40). The hold control unit 82 transmits the data as encrypted data 181 and stores the decryption information 180 in the definitive control area 191. The verification artificial intelligence (AI) model 60 determines the compatibility with the destination, data category, and usage conditions, and only if the authorization determination result 123 indicates authorization, the execution control unit 86 makes the decryption information 180 available for use at the linked system. If authorization is denied, the encrypted data 181 that reached the linked system is not decrypted, and the impact (availability of the data) is not determined.

[0157] As application example 3, we describe the application to a chain of delegations involving multiple artificial intelligence (AI) agents. Acting entity 30 attempts to delegate procurement negotiations (target act) to another artificial intelligence (AI) agent 32. The hold control unit 82 generates conditional delegation information 150, including delegation scope information 153 (the subject and upper limit of the negotiation), in an invalid state, and after authorization, the execution control unit 86 changes it to an active state. If the delegate further sub-delegates, the delegation chain 156 is connected by information 154 identifying the parent delegation, and the depth of the delegation chain is controlled by a condition 109 where the effects accumulate. If signs of infringement by acting entity 30 are detected later, the confirmed authorization information 130 is invalidated by re-evaluation, and the revocation information 155 propagates throughout the entire chain 156.

[0158] As application example 4, we will explain the application to the control of an external device. The acting entity 30 attempts to perform a setting change (target action) on the manufacturing device (external resource 40). Since the setting change is difficult to cancel after execution, the hold control unit 82 holds the execution itself. The verification artificial intelligence (AI) model 60 determines whether the content of the change conforms to safety standards and execution conditions, and after approval, the execution control unit 86 provided in the upstream device 41 performs a confirmation before the command for the change reaches the manufacturing device.

[0159] In Application Example 4, if confirmation is not successful, the change instruction will not reach the manufacturing equipment, and the status of the manufacturing equipment will not change. The denial and the reason for it will be recorded as approval trail information 140, which can be verified by the administrator afterward.

[0160] (Methods and programs for authentication and certification) The program according to this embodiment is a program for causing one or more computers to function as an action request information acquisition unit 81, a hold control unit 82, a determination result acquisition unit 83, an integrated determination unit 84, a confirmation control unit 85, an execution control unit 86, and a record control unit 87. The program may be provided by being recorded on a computer-readable recording medium, or it may be provided via a communication network 70.

[0161] The appraisal and certification method according to this embodiment is a method executed by one or more computers and includes each step corresponding to the processing by each of the above parts, namely, a step of acquiring action request information 100, a step of holding it in reserve, a step of acquiring a judgment result 110, a step of generating an approval judgment result 123, a step of generating confirmed permission information 130 or setting it to an effective state, a step of releasing the reserve and having the target action executed or determining the impact, and a step of generating and recording approval trail information 140.

[0162] The authentication and certification method may further include steps corresponding to the processes described in each embodiment above, for example, a step of transmitting in encrypted form and holding decryption information 180, a step of making the decryption information 180 available, a step of executing in the definitive control area 191, a step of not granting authority to the acting entity 30, a step of holding it in a referable state, a step of including external resource identification information 104, a step of obtaining record completion information 141, a step of generating definitive permission information 130 with a control signature 137, a step of determining the impact, a step of including expiration information 131, one-time identification information 132, and state version 133 or information 134 identifying a commit, a step of verification, a step of executing in the same transaction or an exclusively controlled process, a step of obtaining the series 108, a step of obtaining the determination result 110 as a single target act, a step of accepting delegation, a step of generating conditional delegation information 150, a step of changing to a valid state, a step of identifying the delegation lineage 156, and a step of generating revocation information 155. All control mechanisms on the system side, as detailed in each embodiment, are understood and fully disclosed and supported as each step of the appraisal and certification method and each process of the program.

[0163] Furthermore, the order of each of the above steps and processes may be rearranged to the extent that it does not contradict the technical order, and some steps and processes may be executed in parallel or repeatedly.

[0164] (Security considerations) In the event of an infringement or misdirection by the actor 30, the separation of the execution authority 139 and the determinative control authority 138 ensures that the impact cannot be determined solely by the exercise of authority by the actor 30. Even if the actor 30 attempts to forge the determinative authorization information 130, the control signature generation key 135 is not output outside the determinative control area 191, and therefore a valid control signature 137 cannot be generated.

[0165] In the event of any errors or breaches in the verification artificial intelligence (AI) model 60, the integration of multiple judgment results 110 and quorum conditions 122 ensure that the judgment of a single model alone will not lead to authorization. Verification of the judgment result signature information 114 and the judgment authority information 115 prevents the inclusion of impersonation and unauthorized judgments.

[0166] To prevent theft or reuse of the confirmed authorization information 130, replay and time-delay exploitation are prevented by verification of the expiration information 131, the one-time identification information 132, and the status version 133 or commit identification information 134, along with atomic consumption. To prevent interception of communications, encrypted and mutually authenticated routes are used.

[0167] In the event of tampering with or concealment of evidence, the existence of evidence corresponding to the confirmed impact is guaranteed by recording the tampering on a record-keeping infrastructure 50, and by a sequence in which the completion of the recording is a prerequisite for the generation of confirmed authorization information 130. In the event of failure or uncertainty, a safe shutdown is guaranteed by handling it as a fail-close by handling it as unauthorized.

[0168] To counter dilution of control through chains of delegation, upstream controls are ensured to extend downstream through conditional delegation information 150, tracking of delegation lines 156, control of the depth of delegation chains by conditions 109 where effects accumulate, and cascading of revocation information 155.

[0169] To prevent misconduct by insiders, an operational system may be adopted in which the authority to operate the fixed control area 191, the authority to manage the integrated rules 120 and authorization conditions 121, and the authority to manage the verified artificial intelligence registration information 117 are assigned to different individuals. Since the history of the exercise of each authority is recorded in the record base 50, it is possible to conduct post-incident audits of the actions of insiders.

[0170] (modified version) Figure 16 is an explanatory diagram showing an example of a configuration related to a modified example. As the first modified example, the number of verification artificial intelligence (AI) models 60 can be one or more. Even in a configuration using a single verification artificial intelligence (AI) model 60, the separation of definitive control authority 138, definitive control area 191, handling of disauthorization as a fail-close, expiration date information 131, one-time identification information 132, recording of authorization trail information 140, and other configurations can be applied.

[0171] As a second variation, in addition to, or instead of, the judgments of multiple verification artificial intelligence (AI) models 60, the results of approval by a human approver 200 may be included in the input for integration. The approval condition 121 may include a condition that requires the approval of the approver 200 for a given target action.

[0172] As a third variation, even if the act in question is an act of issuance managed by another system that oversees the issuance of credentials and other proof information, the act of issuance itself can be treated as an act subject to the appraisal and certification system 1 and subject to retention, judgment, and confirmation control. In other words, the management of issuance and the confirmation control of the act may be applied in a layered manner.

[0173] As a fourth variation, decentralized identifiers (DIDs) may be used as the agent identification information 102, the decision-making agent identification information 112, the delegate identification information 151, or the delegate identification information 152. Furthermore, smart contracts may be used to record or manage the authorization trail information 140, the definitive permission information 130, the conditional delegation information 150, the revocation information 155, or the status change information 160, and the definitive permission information 130 or the authorization trail information 140 may be associated with non-fungible tokens (NFTs) or other tokens.

[0174] As a fifth variation, the appraisal and certification system 1 can also be applied to target activities involving digital goods (including the transfer, issuance, or modification of digital goods). In this case, the external resource 40 may be a record infrastructure or platform for managing digital goods.

[0175] As a sixth modification, the control signature 137, the determination result signature information 114, and other signature and cryptographic schemes may include not only public-key cryptography or symmetric-key cryptography, but also cryptographic schemes that are resistant to the risk of decryption by quantum computers (post-quantum cryptography). If the conditions for the transition of the signature scheme are met, the confirmed authorization information 130 or authorization trail information 140 may be re-signed or re-recorded using the transition scheme.

[0176] As a seventh variation, the integration rule 120 may use weights set for each verified artificial intelligence (AI) model 60, and these weights may be updated based on the actual results of the judgment. Furthermore, if the judgment result 110 includes information indicating the degree of certainty, the integration judgment unit 84 may perform weighting according to the degree of certainty.

[0177] As an eighth variation, if the acting entity 30 intends to perform multiple interrelated target acts with respect to multiple external resources 40, the expert certification system 1 may treat the multiple target acts as a single target act or series 108 and control them with a single approval decision result 123 and a single confirmed permission information 130, or control them with confirmed permission information 130 for each external resource 40.

[0178] As a ninth variation, the determination of the impact by the execution control unit 86 may be carried out in two stages. That is, in the first stage, the execution control unit 86 prepares (provisionally determines) the impact on the external resource 40, and in the second stage, after confirming the acquisition and recording of the determined result information 171, it may treat the impact as ultimately valid.

[0179] As a tenth variation, the information indicating the reason for the judgment included in the judgment result 110 may be generated as natural language by a large-scale language model, and its digest or full text may be recorded in association with the authorization trail information 140. This allows the basis for the judgment to be referenced in a human-understandable form during subsequent audits.

[0180] As an eleventh modification, multiple definitive control areas 191 may be provided for redundancy. Even in this case, the generation of definitive permission information 130 for a single target action or setting it to an active state is performed in only one of the definitive control areas 191 by mutual exclusion.

[0181] As a twelfth variation, the hold control unit 82 may set a predetermined deadline for the hold, and if confirmation that the final permission information 130 is valid is not obtained by the deadline, it may automatically treat the held target act as unapproved and cancel the provisional execution.

[0182] As a 13th variation, the expert certification system 1 may provide the auditing entity with means for retrieving authorization trail information 140 and related records. The authority to retrieve is limited to viewing the records and does not include the authority to modify the records.

[0183] As a 14th variation, the authentication certification system 1 may be applied to data linkage in a data linkage platform or data space, supply chain management, financial transactions, manufacturing or logistics control, medical information systems, administrative information systems, or critical infrastructure control. For example, in a data linkage platform, the target actions would be the transmission of data or changes in the setting of usage conditions by an artificial intelligence (AI) agent, which is the acting entity 30, and the external resource 40 may be a linked system or record infrastructure.

[0184] As a 15th variation, the judgment result acquisition unit 83 may not reuse past judgment results 110 for the same target action. Each time a target action is requested, a new judgment result 110 is acquired, so that changes in the premise are reflected in the judgment.

[0185] As a 16th modification, the recording control unit 87 may record the authorization trail information 140 in parallel on multiple independent recording substrates 50. By comparing the multiple records, the resilience of a single recording substrate 50 to failure or breach is enhanced.

[0186] As a 17th variation, the expert certification system 1 may provide the subject 30 with a summary of the reasons for the disapproval of the disapproved subject act. The subject 30 may then re-request the subject act, modified based on these reasons, as new act request information 100.

[0187] As the 18th variation, the action content information 103 may include not only text but also images, audio, structured data, or other forms of information or digests thereof, and the verification artificial intelligence (AI) model 60 may use multiple forms of information together for determination. [Industrial applicability]

[0188] The present invention can be used in all information systems that utilize artificial intelligence (AI) agents, particularly in fields where actions involving irreversible effects on external resources are entrusted to AI agents, such as data linkage, supply chains, finance, manufacturing, logistics, healthcare, government administration, and critical infrastructure. [Explanation of Symbols]

[0189] 1. Authentication and Certification System 10 processors 20 Memory section 30. Acting Agent (Artificial Intelligence (AI) Agent) 32 Other Artificial Intelligence (AI) Agents 40 External resources 41 Pre-stage device 50 Recording infrastructure 51 Servers 52 Distributed Servers 53 Distributed Ledger 60 (60a~60n) Validation Artificial Intelligence (AI) Models 70 Communication Networks 81 Action request information acquisition department 82 Hold Control Unit 83 Judgment result acquisition part 84 Integrated Determination Unit 85 Confirmation Control Unit 86 Execution Control Unit 87 Recording Control Unit 100 Action request information 101 Action Identification Information 102 Identifying the Acting Party 103 Act information 104 External Resource Identification Information 105 Information on the type of activity 106 External Resource Type Information 107 Risk Information 108 series 109 Conditions under which effects accumulate 110 Judgment result 111 Validation Model Identification Information 112 Identification information of the decision-making entity 113 Judgment time information 114 Judgment result signature information 115 Judgment Authority Information 116 Public Key Information 117 Verified Artificial Intelligence Registration Information 120 Integrated Rules 121 Approval Conditions 122 Quorum Conditions 123 Authorization judgment result 124 Permissible Impact Range Information 130 Confirmed permission information 131 Expiration Date Information 132 One-time identification information 133 Version Information identifying commit 134 135 Control Signature Generation Key 136 Control Signature Target Information 137 Control Signature 138 Deterministic control authority 139 Execute permission 140 Authorization trail information 141 Recording complete information 150 Conditional Delegation Information 151 Delegator Identification Information 152 Delegator Identification Information 153 Delegation Scope Information 154 Identifying information of the parent delegate 155 Expiration Information 156 Delegation System 160 Status Change Information 161 Compensation Act 162 Compensation control information 170 Potential Impact Information 171 Confirmed result information 180 Decryption Information 181 Encrypted data 190(190a~190n) Management area 191 Deterministic control region 200 approvers

Claims

1. An expert certification system that provides expert certification of target actions by an artificial intelligence (AI) agent, Equipped with a processor and memory, The processor comprises an action request information acquisition unit, a hold control unit, a determination result acquisition unit, an integrated determination unit, a confirmation control unit, an execution control unit, and a recording control unit. The aforementioned action request information acquisition unit acquires action request information that includes action identification information that identifies the target action that the artificial intelligence (AI) agent, which is the subject of the action, intends to perform on an external resource, action identification information that identifies the subject of the action, and action content information that indicates the content of the target action. The holding control unit holds at least one of the execution of the target act or the determination of the impact of the target act on the external resources until confirmation that the confirmed permission information is valid is obtained. The judgment result acquisition unit acquires a judgment result from each of the multiple verification artificial intelligence (AI) models indicating whether the target act conforms to at least one of the scope of authority granted to the acting entity or the execution conditions attached to said authority. At least one of the aforementioned multiple verification artificial intelligence (AI) models is operated by an entity different from the entity operating the aforementioned acting entity, or operates in a management domain different from the management domain in which the aforementioned acting entity operates. The integrated determination unit integrates the acquired multiple determination results based on predetermined integration rules and generates an approval determination result indicating whether the integrated result satisfies the pre-set approval conditions. The confirmation control unit, when the approval determination result indicates approval, generates or sets to an active state the confirmation permission information associated with the action request information and permits the confirmation of the impact on the external resources, based on confirmation control authority different from the execution authority of the target action granted to the acting entity; the confirmation permission information is not generated or set to an active state solely by the exercise of the execution authority by the acting entity; and the confirmation permission information is not generated or set to an active state when the approval determination result indicates disapproval. The execution control unit, if it confirms that the confirmed permission information is valid and corresponds to the action request information, releases the hold and executes the target action or determines the impact on the external resources; if the confirmation is not successful or the confirmed permission information does not exist, it does not execute the target action or determine the impact. The record control unit generates authorization trail information including the action identification information, information corresponding to a plurality of the judgment results, and the authorization judgment result, and records it in a record infrastructure capable of detecting tampering in association with the action identification information. A certification system characterized by the following features.

2. The aforementioned action includes the process of transmitting data to the aforementioned external resource, The holding control unit transmits the data to the external resource in an encrypted format that cannot be used without decryption information, and holds the decryption information in a state that cannot be used by the artificial intelligence (AI) agent that is the subject of the action. The execution control unit makes the decryption information available to the external resource if the confirmation that the valid authorization information is valid is successful, and does not make the decryption information available to the external resource if the confirmation is not successful. The appraisal and certification system according to feature 1.

3. The aforementioned actions include at least one of the following: connecting to the external resource, using a key, generating an electronic signature, issuing verifiable credentials (VC), transmitting data, delegating authority or processing to another artificial intelligence (AI) agent, processing to change the state of the external resource, processing to activate another artificial intelligence (AI) agent, or processing to control an external device. If the target act is an act for which the determination of the effects can be postponed after execution, the hold control unit provisionally executes the target act and postpones the determination of the effects; if the target act is an act for which the determination of the effects cannot be postponed after execution, the hold control unit postpones the execution of the target act until confirmation that the determination permission information is valid is established. The appraisal and certification system according to feature 1.

4. The deterministic control unit is implemented in a deterministic control domain where access rights to the deterministic control authority are separated from both the management domain where the artificial intelligence (AI) agent acting as the subject operates and the management domain where the multiple verification artificial intelligence (AI) models operate. The artificial intelligence (AI) agent that is the subject of the aforementioned action is not authorized to generate the confirmed permission information, change it to a valid state, or restore it from an invalid state. The confirmed permission information is maintained in a state in which the artificial intelligence (AI) agent acting as the entity is not permitted to modify it, and is accessible from the external resource or the execution control unit located upstream of the external resource. The appraisal and certification system according to feature 1.

5. The aforementioned action request information further includes external resource identification information that identifies the external resource, The record control unit records the authorization trail information on the record base and obtains record completion information indicating the completion of the record. The confirmation control unit, when the authorization determination result indicates authorization and the record completion information is obtained, generates the confirmation authorization information by assigning a control signature to the control signature target information, which includes the act identification information, the act subject identification information, a digest of the act content information, the external resource identification information, a digest of the authorization trail information, and the record completion information or a digest of the record completion information, using the control signature generation key associated with the confirmation control authority. The execution control unit determines the effect only if the verification of the control signature and the confirmation of the correspondence between the control signature target information and the action request information are successful. The appraisal and certification system according to feature 1.

6. The memory unit stores verification artificial intelligence registration information for each of the multiple verification artificial intelligence (AI) models, which associates verification model identification information, judgment authority information indicating the authority to make judgments, and public key information for verifying the judgment result signature information. The judgment result acquisition unit acquires the verification model identification information, the judgment subject identification information, the judgment time information, and the judgment result signature information in association with the judgment result, verifies the validity of the judgment authority information and the judgment result signature information based on the verification artificial intelligence registration information, and uses the judgment result in which the judgment authority information is valid and the verification of the judgment result signature information is successful for integration by the integration judgment unit. The appraisal and certification system according to feature 1.

7. The confirmation authorization information includes expiration information indicating the validity period for which the confirmation of the impact is permitted, one-time identification information that is limited to one use, and information identifying the state version of the external resource or the commit of processing in the external resource. The execution control unit confirms that the validity period is within, that the one-time identification information is unused, and that the status version or the information identifying the commit corresponds to the current state or process of the external resource, and then updates the one-time identification information from unused to used, and determines the impact on the external resource, all within the same transaction or an exclusively controlled process. The appraisal and certification system according to feature 1.

8. The aforementioned action request information acquisition unit acquires a series of target actions performed by the artificial intelligence (AI) agent, which is the subject of the action. Even if each of the multiple target actions falls within the scope of the authority granted to the acting entity, if the sequence of the multiple target actions satisfies the condition for the cumulative impact on the external resources, the judgment result acquisition unit acquires the judgment result from the multiple verification artificial intelligence (AI) models as a single target action. The appraisal and certification system according to feature 1.

9. The aforementioned subject actions include the delegation of authority or processing from the artificial intelligence (AI) agent that is the subject of the aforementioned actions to another artificial intelligence (AI) agent, The holding control unit generates conditional delegation information which includes delegate identification information, delegate identification information, delegation scope information, and information identifying the parent delegation, and which remains in an invalid state until confirmation that the confirmed permission information is valid is obtained. The execution control unit, upon confirmation that the confirmed authorization information is valid, changes the conditional delegation information to a valid state; if the confirmed authorization information later becomes invalid or the information on which the authorization determination result is based is changed, it traces the conditional delegation information to identify the delegation lineage originating from the conditional delegation information and generates invalidation information that invalidates the delegation included in the identified lineage. The appraisal and certification system according to feature 1.

10. If a predetermined time has elapsed for obtaining the determination results, if the number of determination results required for the approval conditions cannot be obtained, or if the validity of the obtained determination results cannot be confirmed, the determination control unit will treat the target act as unapproved and will not generate or set the determination approval information to a valid state. The appraisal and certification system according to feature 1.

11. The authorization conditions include a quorum condition requiring that predetermined determination results be obtained by at least two of the verification artificial intelligence (AI) models operated by different entities or operating in different management domains, The confirmation control unit shall not generate or set the confirmation authorization information to a valid state if the quorum conditions are not met. The appraisal and certification system according to feature 1.

12. The conditions for the cumulative impact on the external resource include at least one of the following: a condition relating to the number of times the target action is performed on the same external resource within a predetermined period; a condition relating to the cumulative amount of changes to the external resource or the amount of data transmitted to the external resource by the target action included in the series; or a condition relating to the depth of the chain of authorizations or delegations of processing to other artificial intelligence (AI) agents included in the series. The appraisal and certification system according to feature 8.

13. An expert certification system that provides expert certification of target actions by an artificial intelligence (AI) agent, Equipped with a processor and memory, The processor acquires action request information, which includes action identification information, action subject identification information, and action content information, regarding the target action that the artificial intelligence (AI) agent, which is the acting entity, intends to perform on an external resource. Until confirmation is obtained that the confirmed permission information is valid, at least one of the following will be withheld: the execution of the subject act or the determination of the impact of the subject act on the external resources. From each of multiple verification artificial intelligence (AI) models, a determination result is obtained indicating whether the target act conforms to at least one of the scope of authority granted to the acting entity or the execution conditions attached to said authority; the multiple determination results obtained are integrated based on predetermined integration rules, and an authorization determination result is generated indicating whether the result of the integration satisfies pre-set authorization conditions; If the approval determination result indicates approval, the confirming permission information is generated or set to an active state, which permits the confirmation of the impact on the external resources based on a confirming control authority different from the execution authority of the target act granted to the acting entity; if the approval determination result indicates denial, the confirming permission information is not generated or set to an active state. If the confirmed permission information is valid and it is confirmed that it corresponds to the action request information, the hold will be released and the target action will be performed or the impact will be determined; if the confirmation is not successful or the confirmed permission information does not exist, the target action will not be performed or the impact will not be determined. This system generates authorization trail information including the aforementioned action identification information, information corresponding to a plurality of the aforementioned determination results, and the authorization determination result, and records it in a recording infrastructure capable of detecting tampering in association with the aforementioned action identification information. The process of generating or setting the aforementioned definitive permission information to a valid state is performed in a definitive control domain where access rights to the definitive control authority are separated from both the management domain where the artificial intelligence (AI) agent acting as the subject operates and the management domain where the multiple verification artificial intelligence (AI) models operate. The artificial intelligence (AI) agent acting as the subject of the action is not granted the authority to generate the confirmed permission information, change it to an active state, or restore it from an invalid state. As a result, the confirmed permission information is not generated or set to an active state by the control within the management area in which the artificial intelligence (AI) agent acting as the subject of the action operates. A certification system characterized by the following features.

14. A program for causing one or more computers to function as an action request information acquisition unit, a holding control unit, a judgment result acquisition unit, an integrated judgment unit, a confirmation control unit, an execution control unit, and a record control unit for appraising and verifying target actions by an artificial intelligence (AI) agent, The aforementioned action request information acquisition unit acquires action request information, which includes action identification information, action subject identification information, and action content information, for the target action that the artificial intelligence (AI) agent, which is the act subject, intends to perform on an external resource. The holding control unit holds at least one of the execution of the target act or the determination of the impact of the target act on the external resources until confirmation that the confirmed permission information is valid is obtained. The determination result acquisition unit acquires a determination result from each of the multiple verification artificial intelligence (AI) models indicating whether the target act conforms to at least one of the scope of authority granted to the acting entity or the execution conditions attached to said authority, and at least one of the multiple verification artificial intelligence (AI) models is operated by an entity different from the entity that operates the acting entity, or operates in a management area different from the management area in which the acting entity operates. The integration determination unit integrates the multiple determination results based on predetermined integration rules and generates an approval determination result indicating whether the integration result satisfies the pre-set approval conditions. The confirmation control unit, when the approval determination result indicates approval, generates or sets to an active state the confirmation permission information associated with the action request information and permits the confirmation of the impact on the external resources, based on confirmation control authority different from the execution authority of the target action granted to the acting entity; the confirmation permission information is not generated or set to an active state solely by the exercise of the execution authority by the acting entity; and the confirmation permission information is not generated or set to an active state when the approval determination result indicates disapproval. The execution control unit, if it confirms that the confirmed permission information is valid and corresponds to the action request information, releases the hold and executes the target action or determines the impact on the external resources; if the confirmation is not successful or the confirmed permission information does not exist, it does not execute the target action or determine the impact. The record control unit generates authorization trail information including the action identification information, information corresponding to a plurality of the judgment results, and the authorization judgment result, and records it in a record infrastructure capable of detecting tampering in association with the action identification information. A program characterized by the following features.

15. The aforementioned action request information further includes at least one of the following: action type information indicating the type of the target action, external resource type information indicating the type of the external resource, or risk information relating to the target action. The judgment result acquisition unit determines, based on at least one of the action type information, the external resource type information, or the risk information, the combination of the verification artificial intelligence (AI) model for acquiring the judgment results and the number of required judgment results. The integrated determination unit applies the integrated rules and approval conditions corresponding to the determined combination and number. The program according to feature 14.

16. The authorization conditions include a quorum condition requiring that predetermined determination results be obtained by at least two of the verification artificial intelligence (AI) models operated by different entities or operating in different management domains, The confirmation control unit shall not generate or set the confirmation authorization information to a valid state if the quorum conditions are not met. The program according to feature 14.

17. If a predetermined time has elapsed for obtaining the determination results, if the number of determination results required for the approval conditions cannot be obtained, or if the validity of the obtained determination results cannot be confirmed, the determination control unit will treat the target act as unapproved and will not generate or set the determination approval information to a valid state. The program according to feature 14.

18. The integrated determination unit generates permissible impact range information indicating the range within which the determination of the impact on the external resources is permissible, based on the plurality of determination results or the result of the integration. The confirmation control unit associates the confirmation permission information with the allowable influence range information. The execution control unit determines the influence only within the range indicated by the permissible influence range information, and does not determine the influence beyond that range. The program according to feature 14.

19. The execution control unit acquires candidate impact information indicating the impact to be determined in the external resource, determines the impact if the candidate impact information corresponds to the content of the target action indicated in the action request information or the determination permission information, acquires determination result information indicating the determined impact in the external resource after determining the impact, and records the determination result information in the recording infrastructure in association with the authorization trail information. The program according to feature 14.

20. The integrated determination unit, after generating the confirmed permission information or determining the impact, detects a change in at least one of the following: the execution environment of the artificial intelligence (AI) agent that is the acting entity, the authority granted to the acting entity, the execution conditions, or the information on which the determination result is based; identifies the target act or confirmed permission information affected by the change based on the authorization trail information; and performs a re-evaluation of the identified target act using the multiple verification artificial intelligence (AI) models. The determination control unit, if the result of the re-evaluation does not satisfy the approval conditions, changes the determination permit information that has not yet been used to determine the impact to an invalid state, and generates status change information indicating non-conformity or expiration for the target act for which the impact has already been determined. The program according to feature 14.

21. When the state change information is generated, the execution control unit identifies a compensatory action to cancel, reduce, disable, or replace the already determined effect with another effect, outputs compensatory control information to execute or cause the external resource to execute the compensatory action, and records the result of the compensatory action in the recording infrastructure in association with the authorization trail information. The program according to feature 20.

22. An expert certification method for verifying an action performed by an artificial intelligence (AI) agent, which is executed by one or more computers, The one or more computers mentioned above A process for acquiring action request information, which includes action identification information, action subject identification information, and action content information, regarding the target action that an artificial intelligence (AI) agent, which is the acting entity, intends to perform on an external resource, The process includes a step of suspending at least one of the execution of the subject act or the determination of the impact of the subject act on the external resources until confirmation that the confirmed permission information is valid is obtained, A step of obtaining a determination result from each of multiple verification artificial intelligence (AI) models indicating whether the target act conforms to at least one of the scope of authority granted to the acting entity or the execution conditions attached to said authority, A step of integrating the multiple judgment results obtained based on predetermined integration rules and generating an approval judgment result indicating whether the integration result satisfies pre-set approval conditions, If the approval determination result indicates approval, the process involves generating or setting to an active state the definitive permission information, which is associated with the action request information and permits the determination of the impact on the external resources, based on a definitive control authority different from the execution authority of the target action granted to the acting entity; the definitive permission information is not generated or set to an active state solely by the exercise of the execution authority by the acting entity; and if the approval determination result indicates denial, the definitive permission information is not generated or set to an active state. If the confirmation that the confirmed permission information is valid and corresponds to the action request information is confirmed, the hold is released and the target action is performed or the impact on the external resources is confirmed; if the confirmation is not confirmed or the confirmed permission information does not exist, the target action is not performed or the impact is not confirmed. A step of generating authorization trail information including the aforementioned action identification information, information corresponding to a plurality of the aforementioned determination results, and the authorization determination result, and recording it in a recording infrastructure capable of detecting tampering in association with the aforementioned action identification information, Includes, At least one of the aforementioned multiple verification artificial intelligence (AI) models is operated by an entity different from the entity operating the aforementioned acting entity, or operates in a management domain different from the management domain in which the aforementioned acting entity operates. Method of authentication and certification.

23. The aforementioned target action includes a process of transmitting data to the external resource in an encrypted format that cannot be used without decryption information, and a step of maintaining the decryption information in a state that cannot be used by the artificial intelligence (AI) agent that is the entity performing the action. If the confirmation that the confirmed permission information is valid is successful, the decryption information is made available to the external resource; if the confirmation is not successful, the decryption information is not made available to the external resource. The method of certification of expert opinion according to claim 22, which includes the following:

24. The process of generating or setting the aforementioned definitive permission information to a valid state is performed in a definitive control area where access rights to the definitive control authority are separated from both the management area where the artificial intelligence (AI) agent acting as the subject operates and the management area where the multiple verification artificial intelligence (AI) models operate. The process of not granting the artificial intelligence (AI) agent, which is the subject of the action, the authority to generate the confirmed permission information, change it to an active state, or restore it from an invalid state, The process of maintaining the confirmed permission information in a state in which modification by the artificial intelligence (AI) agent acting as the entity is not permitted, and making it accessible from the external resource or an execution control process set up prior to the external resource; The method of certification of expert opinion according to claim 22, which includes the following:

25. The process of including external resource identification information that identifies the external resource in the action request information, A step of recording the authorization trail information on the recording infrastructure and obtaining recording completion information indicating the completion of the recording, If the authorization determination result indicates authorization and the record completion information is obtained, the process of generating confirmed authorization information by assigning a control signature to the control signature target information, which includes the act identification information, the act subject identification information, a digest of the act content information, the external resource identification information, a digest of the authorization trail information, and the record completion information or a digest of the record completion information, using the control signature generation key associated with the confirmed control authority. The process of determining the impact is performed only if the verification of the control signature and the confirmation of the correspondence between the control signature target information and the action request information are completed. The method of certification of expert opinion according to claim 22, which includes the following:

26. The process of including in the confirmed permission information: expiration date information, one-time identification information, and information identifying the status version of the external resource or the commit of processing in the external resource, A step of confirming that the expiration date information is within the validity period indicated, that the one-time identification information is unused, and that the status version or the information identifying the commit corresponds to the current state or processing of the external resource, A step of updating the aforementioned one-time identification information from an unused state to a used state, and determining the aforementioned impact on the external resource, by performing these actions in the same transaction or an exclusively controlled process, The method of certification of expert opinion according to claim 22, which includes the following:

27. The process involves acquiring a sequence of multiple target actions performed by the artificial intelligence (AI) agent, which is the subject of the aforementioned actions. Even if each of the aforementioned multiple target acts falls within the scope of the authority granted to the acting entity, if the sequence of the aforementioned multiple target acts satisfies the condition for the cumulative impact on the external resources, the sequence is treated as a single target act, and the determination result is obtained from the aforementioned multiple verification artificial intelligence (AI) models. The method of certification of expert opinion according to claim 22, which includes the following:

28. The aforementioned target actions include a process of receiving a delegation of authority or processing from an artificial intelligence (AI) agent that is the subject of the aforementioned action to another artificial intelligence (AI) agent, A step of generating conditional delegation information which includes delegate identification information, delegate identification information, delegation scope information, and information identifying the parent delegation, and which remains in an invalid state until confirmation that the confirmed permission information is valid is obtained. If the confirmation that the aforementioned final permission information is valid is successful, the process involves changing the aforementioned conditional delegation information to a valid state, If the aforementioned confirmed authorization information later becomes invalid or the information on which the authorization decision result was based is changed, the process of tracing the conditional delegation information to identify the lineage of delegation originating from said conditional delegation information, A step of generating revocation information that invalidates the delegation included in the identified system, The method of certification of expert opinion according to claim 22, which includes the following:

29. The conditions for the cumulative impact on the external resource include at least one of the following: a condition relating to the number of times the target action is performed on the same external resource within a predetermined period; a condition relating to the cumulative amount of changes to the external resource or the amount of data transmitted to the external resource by the target action included in the series; or a condition relating to the depth of the chain of authorizations or delegations of processing to other artificial intelligence (AI) agents included in the series. The method of certification of expert opinion as described in claim 27.

Images