Multi-agent decision causal chain automatic recording and responsibility attribution reasoning method and system
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGZHOU HONGCHEN LINGJING DIGITAL TECHNOLOGY CO LTD
- Filing Date
- 2026-05-13
- Publication Date
- 2026-08-07
AI Technical Summary
[0004]第二，跨维度数据无法自动串联
[0024]规避对物理时间戳的依赖：通过严格偏序关系定义事件间的逻辑因果，解决了分布式多智能体时钟不同步、消息乱序导致的因果误判问题，使得因果链结构唯一可确定。
Abstract
Description
Technical Field
[0001] This invention belongs to the field of multi-agent collaboration and distributed system technology, specifically relating to a method and system that models decision events of a multi-agent system as causal chains independent of physical timestamps, automatically detects causal contradictions, and derives responsibility attribution through a rule engine. This invention is applicable to scenarios with high requirements for decision traceability and accountability, such as industrial control, aerospace equipment collaboration, and large-scale edge computing. Background Technology
[0002] With the rapid development of intelligent agent technology, multi-agent collaboration is increasingly widely used in fields such as industrial control, aerospace, military, and large-scale distributed computing. When multi-agent collaboration completes a complex task and produces erroneous results, the following prominent technical challenges remain: how to trace the entire decision-making process, how to definitively assign responsibility to each participant, and how to generate a standardized audit report that can be submitted to regulators.
[0003] First, causal tracing in decision-making lacks mathematical guarantees that it is independent of physical timestamps. Existing technologies for multi-agent causal tracing often employ timestamp sorting, determining causal relationships by comparing the timestamps of events at each node. This method can lead to misjudgments of causality in distributed systems with asynchronous clocks and out-of-order message delays. Another existing solution uses probabilistic graphical models, such as Bayesian networks, to derive causality; however, the results are random and cannot be used as formal evidence for compliance audits. Existing technologies, such as Lamport's logical clock, only address the logical ordering of events and cannot determine causal triggering relationships between events. This invention, combining a strict partial order relationship with a predefined causal triggering rule base, achieves accurate determination of causal relationships without relying on any clock mechanism.
[0004] Second, cross-dimensional data cannot be automatically linked. In multi-agent systems, log data from dimensions such as behavior, permissions, communication, networking, and security are stored in a scattered manner. Existing solutions lack a unified traceability framework to automatically integrate this data into a complete chain of decision-making evidence. After an incident, multiple log tables must be manually reviewed to piece together clues, which is inefficient and prone to omissions.
[0005] Third, there is a lack of a structure-level mechanism for generating tamper-proof audit evidence. Existing log anti-tampering schemes only protect the content of individual log entries from modification, but do not protect the causal relationships between log entries from being tampered with. Malicious parties can delete or forge causal edges and still pass individual log hash verification. Summary of the Invention
[0006] The technical problem to be solved by this invention is: how to provide a method and system that does not rely on physical timestamps, automatically constructs decision causal chains, aggregates evidence data across dimensions, deterministically derives responsibility attribution based on predefined rules, and generates structurally immutable audit reports.
[0007] Terminology Definition
[0008] Event Log Item: The core data structure of this invention, corresponding to a decision event of an agent, records the complete context and legal basis of the operation. In the mathematical model, it corresponds to the concept of a morphism.
[0009] Causal relationship edge: Represents a logical causal triggering relationship between two event records. It is a strict partial order relationship that satisfies non-reflexivity, antisymmetry, and transitivity, and does not depend on physical timestamps. The strict partial order relationship is implemented through a predefined causal triggering rule base, including standard patterns such as task allocation triggering, permission verification triggering, communication handshake triggering, and security monitoring alarm triggering. Each triggering pattern corresponds to a unique logical judgment condition, ensuring the unidirectionality and non-cyclicity of the causal relationship.
[0010] Causal closure operator: An algorithm module implemented in a distributed system, used to automatically collect all event records that are causally related to the target event by traversing forward and backward along causal edges, and outputting the computation process of the minimum complete evidence set.
[0011] State composite result: refers to the final state observed at the target node after combining and calculating the state snapshot, data record, and behavior triggering condition of the event record item along a specific path in the causal chain graph. It is used for parallel path consistency detection.
[0012] Minimal fixed point: The stable state reached during the responsibility inference iteration process when the responsibility assignment result no longer changes, corresponding to a unique and reproducible responsibility attribution result.
[0013] Minimal complete evidence set: refers to the minimum set of data that can completely reconstruct the cause and effect of an event, which is automatically calculated by the causal closure operator. It includes all direct causal events and necessary indirect causal events, and has no redundant events. It is automatically determined by the monotonic convergence property of the causal closure operator.
[0014] Technical solution
[0015] This invention provides a method for automatically recording causal chains in multi-agent decision-making and reasoning about responsibility attribution, executed on a distributed agent device, comprising the following steps:
[0016] Step S1. Each intelligent agent device automatically generates an event record item when executing a decision event; the event record item includes an event identifier, a subject identifier, an event type, a digital signature, the rule basis involved in this decision, a rule version snapshot, a subject permission snapshot at the time of the event, and a topology context snapshot; the event record item is synchronously and permanently generated when the event occurs, and it is prohibited to add, delete, or tamper with it afterward;
[0017] Step S2. The system automatically detects currently unassociated upstream event records. If an upstream event record that meets the predefined causal triggering conditions exists, a causal relationship edge is established from that upstream event record to the current event record. When establishing the causal relationship edge, the validity of the digital signature of the upstream event record is automatically verified. Event records that fail the signature verification do not participate in the construction of the causal relationship edge. The predefined causal triggering conditions are strict partial order relations on the event set, satisfying non-reflexivity, antisymmetry, and transitivity, and do not depend on the order of physical timestamps to determine causality.
[0018] Step S3. When tracing the target event, the causal closure operator is used to traverse along the causal relationship edges in both the forward and backward directions to automatically collect all event records and their context snapshots that are causally related to the target event, and output the minimum complete evidence set of the target event; if an expected event that should exist but is missing is detected during the tracing process, a missing marker node is generated and the node is included in the minimum complete evidence set.
[0019] Step S4. Perform consistency checks on parallel paths in the causal chain graph: When there are multiple different morphic composite paths from the same initial event to the same target event, and the morphic composite results of each path do not satisfy the commutativity of the commutative graph—that is, when reaching the same target state along different paths, the state attributes, data records, or subsequent behavior triggering conditions are inconsistent—mark the relevant event record items as candidate sources of contradiction;
[0020] Step S5. Based on the predefined hierarchical responsibility rule base expressed using modal logic formulas, starting from the contradiction markers in Step S4, perform monotonic operator iterative derivation in reverse along the causal association edges to automatically determine the responsibility attribution type of each relevant event record item; when the iteration reaches the point where the responsibility assignment result no longer changes, converge to the minimum fixed point and output a unique and reproducible final responsibility attribution result; if the number of iterations exceeds a preset configurable maximum iteration depth threshold (e.g., 1000 times, which can be adjusted according to the actual deployment scenario) and still has not converged, then all relevant event record items on the loop path are automatically truncated and marked as system architecture responsibility;
[0021] Step S6. Map the topological structure of the causal chain graph and all event record items to a Merkle causal tree structure; the leaf nodes of the Merkle causal tree correspond to event record items, and the non-leaf nodes correspond to the aggregation of causal association edges; the integrity check code of each non-leaf node is calculated by combining the content check codes of all its child nodes and the structure check code used to characterize the connection topology between the child nodes; generate the final integrity check code based on the root node of the Merkle causal tree, and generate a standardized audit report containing the causal chain graph, responsibility attribution results, judgment basis, and rule clause references.
[0022] Beneficial effects
[0023] The present invention has the following beneficial effects:
[0024] Avoiding dependence on physical timestamps: By defining the logical causality between events through strict partial order relations, the problem of causal misjudgment caused by clock asynchrony and message out-of-order in distributed multi-agent systems is solved, making the causal chain structure uniquely determinable.
[0025] Automatic cross-dimensional evidence aggregation: By leveraging the monotonically convergent property of the causal closure operator on a complete lattice, all evidence related to the target event can be automatically calculated without the need for manual piecing together from logs in multiple modules, greatly improving tracing efficiency and ensuring the integrity and reusability of the evidence set.
[0026] The reasoning results are unique and reproducible: Based on the responsibility rules expressed in modal logic and the fixed-point iteration of monotonic operators, it is ensured that, given the same input and rule base, the responsibility attribution results are unique and reproducible, meeting the compliance and evidence requirements of high-reliability scenarios such as industry and aerospace.
[0027] Structure-level tamper-proof evidence storage: Merkle causal tree is used to store evidence of the entire causal chain. Any modification to the event content or causal topology will be detected, providing a higher level of security than existing single log anti-tampering. The generated audit report can be directly submitted to the regulator.
[0028] High engineering feasibility: This invention can run on ordinary distributed intelligent agent devices without the need for customized dedicated hardware, and can be deployed on general embedded and edge computing intelligent agent devices. Event log items support distributed redundant storage, and the rule base supports offline configuration and online updates, making it adaptable to practical deployment scenarios such as industrial and aerospace applications. Detailed Implementation
[0029] Example 1: Automatic Construction of Causal Chains and Time Independence
[0030] Agent A, in state S0, executes a task assignment decision and automatically generates a first event record. This record contains a snapshot of A's permissions (e.g., "Assignable image recognition task"), rule version (e.g., "V1.2"), digital signature, and timestamp (used only as an attribute, not for causal determination). Agent B subsequently receives the task and executes a capability matching decision, generating a second event record.
[0031] The system detects that the occurrence of the second event record satisfies the predefined causal triggering condition "task is assigned and the receiver executes," which satisfies the transitivity of partial order relations. Therefore, a causal relationship edge is automatically established from the first event record to the second event record. When establishing this causal relationship edge, the system automatically verifies the validity of the digital signature of the first event record. Even if the second event record arrives at the central log system before the first event record due to network latency, this logical causal edge can still be correctly established, eliminating the requirement for physical clock synchronization.
[0032] Example 2: Parallel Path Conflict Detection and Responsibility Attribution Derivation
[0033] A design task is triggered simultaneously by a root event and processed in parallel by two agents, A and B, before being aggregated by agent C. In the corresponding causal chain graph, there are two parallel paths from the root event to the aggregated event. The tracing system uses causal closure operators to traverse these two paths, extracting all events and their state snapshots along each path.
[0034] When comparing the morphic composite results of path 1 (processed by A) and path 2 (processed by B), it was found that the design version generated by path 1 was inconsistent with the version generated by path 2, causing the exchange graph to fail to satisfy exchangeability. The system automatically marked the relevant events on the two paths as candidate sources of conflict.
[0035] The responsibility inference engine starts and iterates monotonically along the marked path in reverse. The iteration formula is: Responsibility Iteration Operator F: R→R, where R is the set of responsibility attribution, F(r) = {r' | r' = rule base matching result(r)}, and it converges to the minimum fixed point when F(r) = r. The rule base is expressed using modal logic formulas, and example rules include: □(Mismatch in capability declaration ∧ Task execution error → Direct responsibility); ◇(Establishment of delegation relationship ∧ Inadequate supervision → Indirect responsibility). Here, □ indicates 'certainty', meaning that when the preconditions are met, the corresponding responsibility result will inevitably occur; ◇ indicates 'possibility', meaning that when the preconditions are met, the corresponding responsibility result may occur.
[0036] By matching the modal logic rule base, it was found that the event on path 2 was directly responsible because the capability declaration referenced by B was incompatible with the task requirements; while the root task initiator A was not at fault. After the iteration converged to the minimum fixed point, a unique and reproducible responsibility attribution result was output. Responsibility types include four categories: direct responsibility refers to a clear decision error in the event record that directly leads to a contradictory result; indirect responsibility refers to related errors such as supervision or delegation in the event record that indirectly lead to a contradictory result; environmental responsibility refers to errors caused by external environmental factors, not agent decision errors; and system architecture responsibility refers to errors caused by system design defects, not single agent errors.
[0037] Example 3: Merkle Causal Tree Tamper-Proof
[0038] Map all event records and causal links in the complete causal chain described above to a Merkle causal tree. Leaf nodes store the integrity checksum of each event record, calculated using the SHA-256 algorithm. Non-leaf nodes correspond to the aggregation of causal links, and their integrity checksums are calculated by combining the content checksums and structure checksums of all child nodes: Non-leaf node checksum = SHA-256(Child node 1 content checksum + Child node 2 content checksum + Structure information).
[0039] Any alteration to the event content or causal chain topology will result in a change to the final root checksum. Unlike existing technologies (such as blockchain and ordinary hash trees), the Merkle causal tree of this invention not only protects data integrity but also protects the integrity of the causal topology structure through a structural checksum, and can detect tampering with causal edges.
[0040] This invention employs distributed multi-replica (e.g., 3 replicas) redundant storage and guarantees data consistency based on the Paxos protocol. Physical damage to a single device only marks a break in a local chain of evidence, without causing the loss of the global causal chain; the system automatically retains the complete topology and checksums of the remaining nodes.
[0041] Example 4: Handling Circular Dependencies (Maximum Iteration Truncation)
[0042] When monitoring alarm events and task self-healing events are repeatedly triggered, a local loop forms in the causal chain graph. If the monotonic operator of the responsibility inference engine fails to converge after iterating beyond a preset configurable maximum iteration depth threshold, a truncation mechanism is triggered. The system locates all event nodes on the loop path and marks them uniformly as system architecture responsibilities, stopping further derivation of the loop to prevent exhaustion of computing resources.
[0043] Example 5: Handling of Missing Events
[0044] The system's rule base predefines the expected behavior that "a security monitoring agent must report and process the event within a specified time window after receiving an alarm." When tracing a security incident, the causal closure operator detects the absence of this expected processing event while constructing the evidence set. The system automatically generates a virtual node marked "expected missing" in the evidence set and includes it as part of the incident chain, determining the monitoring party's responsibility in rule-based reasoning.
[0045] Example 6: Temporary Delegation Authorization
[0046] Agent A temporarily delegates its "write to database" permissions (which it does not possess) to Agent B to perform a maintenance task. A data write error occurs during task execution. During tracing, the event log entries generated by A and B are linked through the delegation relationship tag. Based on this, the rule engine assigns responsibility: A bears indirect responsibility for initiating the delegation and failing to adequately supervise it, while B bears direct responsibility for operational errors during execution.
[0047] Example 7: Large-scale multi-subtask hierarchical incremental tracing
[0048] When the root task contains hundreds of subtasks and involves the collaborative execution of thousands of agent nodes, the system adopts a hierarchical incremental tracing scheme to reduce the computational overhead of full traversal: The system first divides the system into independent causal subgraphs according to the subtask dimension. The causal closure operator only performs incremental aggregation operations within each causal subgraph, without traversing all nodes. The exchange graph detection module only performs consistency checks on parallel paths with anomalies, without additional processing for normal paths. Causal relationships across subgraphs are processed uniformly in the final aggregation stage, effectively controlling the overall inference time complexity. Based on this mechanism, the system can support collaborative tracing of thousands of agent nodes. Under typical industrial deployment conditions, the subgraph incremental inference time can be controlled within 100 milliseconds, adapting to the efficient tracing requirements of large-scale distributed collaborative scenarios.
Claims
1. A method for automatically recording causal chains in multi-agent decision-making, executed on a distributed agent device, characterized in that, Includes the following steps: Each intelligent agent device automatically generates an event log entry when executing a decision event; The event log entry shall include at least the event identifier, subject identifier, digital signature, rule version snapshot, rule basis for this decision, subject permission snapshot and topology context snapshot at the time the decision occurred; The event record items are generated and solidified synchronously when the event occurs, and it is prohibited to add, delete or tamper with them afterward; the system automatically detects upstream event record items that are not currently associated. If there is an upstream event record item that meets the predefined causal triggering conditions, a causal association edge is established from that upstream event record item to the current event record item. When establishing causal edges, the validity of the digital signature of the upstream event record is automatically verified. Event records that fail the signature verification will not participate in the construction of causal edges. Digital signatures are used to prevent event log entries from being forged or tampered with, ensuring the integrity and authenticity of the causal chain; The predefined causal triggering condition is a strict partial order relation on the event set, satisfying non-reflexivity, antisymmetry, and transitivity, and does not use the order of physical timestamps as the basis for determining causality.
2. A multi-agent responsibility attribution reasoning method, The method according to claim 1 is characterized by performing the following steps: The target event is traced back by traversing the causal relationship edges using the causal closure operator, automatically collecting all event records that are causally related to the target event, thus forming the minimum complete evidence set for the target event; The minimum complete evidence set contains all direct causal events and necessary indirect causal events, and has no redundant events. It is automatically determined by the monotonic convergence property of the causal closure operator. Consistency checks are performed on parallel paths in the causal chain graph. When the same target state is reached along different composite paths, the inconsistencies between the state attributes, data records, or subsequent behavior triggering conditions are compared. If inconsistencies exist, the relevant event log entries are marked as potential sources of conflict. Based on a predefined hierarchical responsibility rule library, starting from the candidate source of the contradiction, the monotonic operator is iterated and deduced in reverse along the causal relationship edge to determine the responsibility attribution type of each event record item. When the iteration reaches the point where the responsibility assignment result no longer changes, it converges to the minimum fixed point and outputs a unique and reproducible final responsibility attribution result. The predefined hierarchical responsibility rule base is expressed using modal logic formulas; The types of liability attribution include direct liability, indirect liability, environmental liability, and system architecture liability.
3. A method for generating multi-agent compliance audit reports, characterized in that, Includes the following steps: The topological structure of the causal chain graph and all event record items are mapped to a Merkle causal tree structure, where the leaf nodes of the Merkle causal tree correspond to event record items and the non-leaf nodes correspond to the aggregation of causal association edges. The integrity check code of each non-leaf node is calculated by combining the content check codes of all its child nodes and the structure check code used to characterize the connection topology between the child nodes. The content check code and structure check code are calculated using a hash algorithm, including the SHA-256 algorithm. Based on the root node of the Merkle causal tree, the final integrity check code is generated; Generate a standardized audit report that includes the aforementioned causal chain diagram, the results of liability attribution, the basis for judgment, and the references to the rule clauses.
4. A multi-agent decision-making causal chain automatic recording and responsibility attribution reasoning system, used to execute the method described in any one of claims 1 to 3, characterized in that, include: An event logging and signature module deployed on each intelligent agent is used to generate the event logging items; The causal binding module is used to establish the causal association edge based on the predefined causal triggering conditions; The causal closure calculation and contradiction detection module is used to generate the minimum complete evidence set and label the contradiction candidate sources; The responsibility inference engine is used to execute the monotonic operator iteration and output the responsibility attribution result. The audit report generation and evidence storage module is used to construct the Merkle causal tree and generate the audit report. The rule base configuration module is used for offline configuration and online updating of the modal logic responsibility rule base; The data masking and redundant storage module is used for data masking and redundant storage of multiple copies of critical events during cross-organizational tracing. The desensitization is performed at the field level, and sensitive information is stored using AES encryption and decrypted as needed during auditing.
5. The method according to claim 2, characterized in that, The direct responsibility refers to a clear decision-making error in the event record that directly leads to the contradictory result; the indirect responsibility refers to related errors such as supervisory or delegation errors in the event record that indirectly lead to the contradictory result. Environmental responsibility refers to errors caused by external environmental factors, including non-agent decision-making errors. The system architecture responsibility refers to errors caused by system design flaws, not single agent errors; when the number of iterations of the monotonic operator exceeds a preset configurable maximum iteration depth threshold and fails to converge, it is automatically truncated and all relevant event records on the current loop path are marked as system architecture responsibility.
6. The method according to claim 2, characterized in that, When collecting the minimum complete evidence set, if there are missing events that were expected but did not occur, then a missing marker node is generated and included in the evidence set.
7. The method according to claim 1, characterized in that, When a retry event from the same source is detected, it is marked as an equivalence class; when establishing causal relationships or performing responsibility reasoning, the event record items in the equivalence class are merged.
8. The method according to claim 2, characterized in that, The causal relationship edge or event record item also includes a temporary delegation label; when performing the responsibility attribution inference, the responsibility is divided into the initiation and supervision responsibility of the delegating party and the direct execution responsibility of the delegated party based on the temporary delegation label.
9. The method according to claim 3, characterized in that, The standardized audit report is compatible with Syslog and JSON common log formats and supports cross-organizational compliant presentation and anonymized export.
10. The system according to claim 4, characterized in that, Each module adopts a microservice architecture, communicates through application programming interfaces, supports containerized deployment, and supports independent tracing and parallel reasoning of local subgraphs; the rule base configuration module supports structured format import and export, and ensures rule consistency through version number verification during online updates.