SYSTEM FOR PROCESSING LEGAL DOCUMENTS IN A DISTRIBUTED NETWORK INCLUDING MULTIPLE EDGE NODES

Description

BE2025 / 5547 2 other AI agents) to host; where the system is configured as a redundant network system; where the edge nodes are mutually connected (directly or indirectly) via a secure network (e.g. a peer-to-peer network); and where at least one of the edge nodes is equipped to receive a legal question and / or a legal task (e.g. in the form of a prompt, and / or an image, and / or an electronic document such as a PDF document, a text document, etc.); where at least one of the edge nodes includes a first AI agent (M1) (also called JADOP module, where JADOP stands for "Jurisdiction-Aware Dynamic Agent Orchestration Protocol") that is configured to (e.g. automatically) select a subset of AI agents to handle the legal question and / or task (e.g. based on certain relevance scores derived from the semantic analysis of the task); and where at least one e-nominated 10 is configured for the (e.g.real-time learning and / or updating of legal and regulatory knowledge of a multitude of legal domains (e.g.: commercial law, criminal law, tax law, family law, etc.); and where the selected AI agents generate task results without the legal data being sent outside the local node. This system can also be called a "distributed legal AI system".15 The system can be, for example, an Edge network that extends across multiple sites within one country, or that extends across multiple sites spread across at least two countries, or that extends across multiple sites spread across multiple continents, e.g. a globally distributed network. The modules can be hardware modules, or software modules, or a mix of hardware and software. Preferably, every edge node has all the functionality of "MainOrchestrator", but at any given moment, only one of the Edge Nodes of a predefined group of edge nodes (hereinafter also called "subsystem") effectively assumes the role of "MainOrchestrator".Preferably, the AI ​​agents (or "AI modules") are specialized in at least one of: a jurisdiction, a legal area, or a specific legal processing function. It is an advantage that this system is sovereign, scalable, and reliable. The architecture is designed to be and remain legally sound, even when workloads peak, jurisdictions change, and / or languages ​​switch. In one implementation form, everything is configured for learning (e.g., local learning) the knowledge and regulations of a multitude of legal domains at the national, and / or regional, and / or federal, and / or supranational level. 2025 / 5547 BE2025 / 5547 3 In one implementation form, the network system is configured as a self-healing architecture (e.g. as a beehive architecture, or a hive structure, or a grid structure) where coordinating functions are redundantly present. In one implementation form, the network system is configured as a self-healing architecture where coordinating functions are automatically replaced before they fail.5 (e.g.upon detecting a suspected infringement, or similar) In an implementation form, the system is configured such that the multitude of edgegenodes are physically located in at least two different countries. In an implementation form, the system is configured such that the system comprises multiple national subnetworks, which are linked together via secure interconnection points for international legal analysis and redundancy. Each subnetwork preferably has one edgegenodes acting as the "main orchestrator". As an example, a subnetwork can typically comprise one legal area, e.g., one country in Europe, or one state in the US. In an implementation form, the system is configured such that the multitude of edgegenodes are physically located in at least two different continents. In an implementation form, the system is configured in such a way that it comprises multiple continental subnetworks that are linked together via secure interconnection points for intercontinental legal analysis and redundancy.In one implementation form, the first AI agent (M1) is configured to select one or more agents based on jurisdiction, language, and subject matter upon an incoming legal query. The inventors are of the opinion that such a JADOP task allocation module, specifically tailored to legal tasks, does not yet exist, but offers significant advantages in the context of the present invention, such as not having to share all available (not yet anonymized) information metal lenodes, and / or not having to pointlessly translate all information into all possible languages. In one execution form, the first AI agent (M1) is further configured to also take into account one or more of the following aspects: (i) historical performance scores of agents in similar cases; (ii) the origin of jurisdiction of relevant model updates; (iii) current capacity and latency measurements.30 In this way, tasks can be assigned to that subset of AI agents whose combined processing time is the shortest and the expected accuracy is the highest.2025 / 5547 BE2025 / 5547 4 In one implementation form, at least one edge node of the system (and preferably all edge nodes) includes a second AI agent (M2) configured to verify clause compliance without disclosing clause content, e.g. by using "zero-knowledgeproofs". The second AI agent is also referred to herein as the "ZK-LVC" agent ("Zero-KnowledgeLegal ClauseVerification-module") or the "ZK-FKV" module ("Zero-Knowledge-FederatedKnowledge5 Verification"-module). In one implementation form, this second module is configured to verify that a model update, learning unit, or derived knowledge provided by a node or partner meets predefined legal knowledge criteria, without exposing the underlying documents, data content, or origin thereof to other nodes or partners.10 In one implementation form, this second module is configured to prove that a legal clause meets a specific statutory standard without exposing the content of the clause.In one execution form, the ZK-FKV module uses cryptographic zero-knowledge proof techniques applied to parameters or weights of the learning model, such that the contribution15 can be utilized in the network's learning process while maintaining the confidentiality and anonymity of the source. In one execution form, the "zero-knowledge proof" is executed via zk-SNARKs, zk-STARKs, or a similar protocol, with predefined legal knowledge criteria and security parameters.20 In one execution form, at least one edge node of the system (and preferably all edge nodes) includes a third AI agent (M3) configured to cluster incoming legal questions (or queries) and / or model updates. The third AI agent is also referred to here as the "ALRME" module ("AdaptiveLegal RedundancyMinimisationEngine").25 In an implementation form, the third module is configured to cluster incoming legal questions (or queries) and / or model updates based on legal semantic similarities and / or contextual metadata, such that redundant processing across multiple edge nodes is avoided and already known patterns or case law are not unnecessarily re-analyzed, thereby avoiding redundant processing across multiple nodes. 30 The clustering may optionally make use of a legal embedding space trained on jurisdiction-specific legal texts. In this way, clustering results can be shared without revealing the original content. 2025 / 5547 BE2025 / 5547 5 In an implementation form, the legal embedding space and clustering algorithms are updated via federated learning with the application of differential privacy, such that model improvements are shared without exposing the underlying confidential data or document content.In an implementation form, legal questions and / or model updates are not bundled or clustered only in one country5 or in one state, but across at least two or at least three languages, and / or across at least two countries or at least three countries or states, and / or across at least two continents. Searching for and bundling identical questions is simple, but bundling semantically and legally equivalent questions certainly is not. In an implementation form, at least one edge node of the system (and preferably all 10 edge nodes) includes a fourth AI agent (M4) configured to compare and align imprecents (e.g. multi-jurisdictional). The fourth AI agent is also referred to herein as the "CL-PAM" module ("Cross-Language Precedent Alignment Module"). The inventors are of the opinion that an AI agent that automatically harmonizes precedents while preserving the legal context does not yet exist.15 In one implementation form, the fourth module is configured to automatically align and link those legal precedents, rulings, or interpretations in multiple languages ​​based on substantive and / or procedural similarity, whereby the module can add annotations that refer to corresponding precedents in other jurisdictions, without exchanging the original confidential documents or source files. 20 In one implementation form, the fourth module is further configured to also take into account one or more of the following when aligning precedents: (i) local procedural differences; (ii) jurisdiction-specific interpretation rules; (iii) translation uncertainties, whereby these factors are weighted in the alignment score without exchanging the original confidential documents. 25 In one implementation form, the clustering uses a legal embedding space that is trained on jurisdiction-specific legal texts.In an implementation form, the clustering makes use of a legal embedding space that is trained on jurisdiction-specific legal texts, and where the legal embedding space is updated via federated learning with differential privacy.30 In an implementation form, the alignment of precedents also takes into account local procedural differences. In an implementation form, at least one edge node of the system (and preferably all edge nodes) includes a fifth AI agent (M5) that is configured to monitor (e.g. detect and / or harmonize) interpretation drift (e.g. within the multinational, supranational, international and / or global federated model). The fourth module is also referred to herein as the "FCDD" module ("FederatedComplianceDrift Detector"). This module ensures coherence across multiple e-odes in order to maintain model accuracy.5 In one implementation form, the fifth module is configured to detect deviations in legal interpretations, model output, or application criteria between edge odes, e.g.based on a comparison of model parameters and / or semantic outputs. Preferably, detected deviations are automatically corrected by a consensus mechanism, without disclosing the confidential training data of individual edgenodes.10 In an implementation form, the consensus mechanism for drift correction applies a weighted majority based on, for example, jurisdiction authority, the recency of the data used, and the degree of reliability of the participating edgenodes, without the underlying confidential training data being shared between edgenodes. In an implementation form, the edgenodes are provided to share new or updated information with other edgenodes (preferably in real-time or as soon as possible) via the secure (e.g., peer-to-peer) network. In one implementation form, at least two of the edge nodes, but preferably all edge nodes, comprise at least four of the aforementioned AI agents, more specifically, at least the first AI agent (M1), the second AI agent (M2), the third AI agent (M3), and the fourth AI agent (M4).20 In one execution form, at least two of the edgenodes, but preferably all edge nodes, comprise at least four of the aforementioned AI agents, more specifically, at least the first AI agent (M1), the second AI agent (M2), the third AI agent (M3) and the fifth AI agent (M5). In one execution form, at least two of the edgenodes, but preferably all edge nodes, comprise at least four of the aforementioned AI agents, more specifically, at least the first AI agent (M1), the second AI agent (M2), the fourth AI agent (M4) and the fifth AI agent (M5). In one execution form, at least two of the edgenodes, but preferably all edge nodes, comprise at least four of the aforementioned AI agents, more specifically, at least the first AI agent (M1), the third AI agent (M3), the fourth AI agent (M4) and the fifth AI agent (M5). In one execution form, at least two of the edgenodes, but preferably all 30 edge nodes, comprise at least four of the aforementioned AI agents, more specifically, at least the second AI agent (M2), the third AI agent (M3), the fourth AI agent (M4) and the fifth AI agent (M5).2025 / 5547 BE2025 / 5547 7 In an implementation form, at least two of the edge nodes, but preferably all edge nodes, comprise at least the first AI agent (M1), the second AI agent (M2), the third AI agent (M3), the fourth AI agent (M4) and the fifth AI agent (M5). It is an important advantage of this system that it simultaneously includes advanced load balancing, data sovereignty, cryptographic verification and multi-jurisdictional knowledge sharing.5 Preferably, the first AI module (JADOP) is only active on one edge node of each subsystem, but is present on the other Edge Nodes, and can or will be activated there if problems are detected with the edge node or the thief as the root node. In a form of execution, at least one of the oaths is provided for receiving a legal question and / or a legal task, stated in natural language, and where at least one of 10 oaths (e.g., the same invited person, or another invited person) is provided for to give an answer to the posed legal questions and / or the legal task.According to a second aspect, the present invention also provides a method for processing legal questions and / or commands, which comprises the following steps: a) providing a system according to one of the preceding claims; b) receiving a legal question and / or a legal command (e.g. in the form of a prompt, or a query, which may contain natural text, and / or an image, and / or an electronic document such as a PDF document, a text document, etc.) by one of the system's members (100); c) providing an answer (e.g. in the form of natural text, and / or an image, and / or an electronic document) by one of the members. This method can be performed by a system according to the first aspect, but that is not strictly necessary, and this method can also be performed by another system, e.g. a variant of the system according to the first aspect. Specific and preference-bearing aspects of the invention are included in the attached independent and dependent claims.Characteristics of the dependent claims25 can be combined with characteristics of the independent claims and with characteristics of other dependent claims as indicated and not merely as expressly presented in the claims. These and other aspects of the invention shall be evident from and clarified by reference to the form(s) of execution described below.30 Brief description of the figures FIG.1 is a schematic representation of a system based on the present invention, with edges set up in at least two countries. 2025 / 5547 BE2025 / 5547 8 FIG.2 is another schematic representation of a system based on the present invention. FIG.3 and FIG.4 show an example of a legal assignment relating to two jurisdictions, in the example France (FR) and the US state of Delaware (DE). FIG. 5 shows an example of a cross-border Legal Processing.5 The figures are schematic only and not exhaustive.Reference numbers in the claims may not be interpreted to limit the scope of protection. Detailed description of the forms of execution of the invention10 The present invention shall be described with regard to specific forms of execution and with reference to specific drawings, however the invention is not limited thereto but is limited only by the claims. It should be noted that the term "includes", as used in the conclusions, should not be interpreted as limited to the means described below; this term does not exclude any other elements or steps. It is thus to be interpreted as specifying the presence of the mentioned characteristics, values, steps, or components referred to, but does not exclude the presence or addition of one or more other characteristics, values, steps, or components, or groups thereof. Thus, the scope of the expression "a facility comprising means A and B" should not be limited to facilities consisting of only components A and B.It means that with regard to the present invention, A and B are the only relevant components of the device. Reference throughout this specification to “one form of execution” or “a form of execution” means that a specific mark, structure or characteristic described in connection with the form of execution is incorporated in at least one form of execution of the present invention. Thus, occurrences of the expressions “in one form of execution” or “in a form of execution” at various places throughout this specification do not necessarily all have to refer to the same form of execution, but may do so. Furthermore, the specific features, structures or characteristics may be combined in any suitable manner, as would be clear to an average skilled person on the basis of this disclosure, in one or more forms of execution.Similarly, it should be appreciated that in the description of exemplary implementation forms of the invention, various features of the invention are sometimes grouped together in a single implementation form, figure, or description thereof with the aim of 2025 / 5547 BE2025 / 5547 9 streamlining the disclosures and assisting in understanding one or more of the various inventive aspects. This method of disclosure should, however, not be interpreted as a reflection of an intention that the invention requires more features than explicitly mentioned in each claim. Rather, as the following claims reflect, inventive aspects lie in fewer than all the features of a single 5 previously disclosed implementation form. Thus, the claims following the detailed description are hereby explicitly included in said detailed description, with each on self-standing conclusion as a separate form of implementation of this invention.Furthermore, while some of the forms of execution described herein contain some, but not others, features included in other forms of execution, combinations of features of 10 different forms of execution are intended as falling within the scope of the invention, and constitute these different forms of execution as would be understood by the skilled person. For example, in the following claims any of the described forms of execution may be used in any combination. 15 The abbreviation AI stands for "Artificial Intelligence". The abbreviation GAI stands for "Generative Artificial Intelligence". The abbreviation CPU stands for "Central Processing Unit". The abbreviation GPU stands for "GraphicsProcessingUnit", or "graphics processing unit".20 The abbreviation TPU stands for "TensorProcessingUnit", or "tensor processing unit". The abbreviation FPGA stands for "FieldProgrammableGateArray", or "programmable logic array".It is a digital chip that can be programmed for specific purposes. The abbreviation ASIC stands for "Application-Specific Integrated Circuit". It is a chip that is specially designed for certain applications. The term "AI agent" is used in its common meaning. Simply put, it is a software program that can autonomously perceive its environment, make decisions, and perform tasks. The term "edgenode" or "node" can be translated as "edge node". Simply put, "edgenodes" or "edge servers" are powerful computers that are placed "at the edge" of a network where data is calculated, i.e., closer to the users. Edgenodes used in systems according to the present invention preferably include Parallel Processing Agents chosen from the following group: GPUs, CPUs, FPGAs, ASICs. 2025 / 5547 BE2025 / 5547 10 An "edge computing network" is a network where data processing and storage take place closer to the source of the data, instead of in a central data center.The term "load balancing" can be translated as "distributing the load". A "peer-to-peer" (P2P) network is a computer network in which all participating computers, also called peers, nodes, or nodes, are equal and both share and process information. Unlike client-server networks, there is no central server that has control. Each peer can function as both a client and a server. In the present invention, "Main Orchestrator" or "orchestration node" refers to an edge node that assumes the role of distributing incoming tasks or commands, specifically the JADOP module running on this edge node distributes incoming requests or commands to suitable AI agents available on the same edge node (as the one on which the JADOP runs), or on other edge nodes. The invention in question is in the field of artificial intelligence, but specifically in applications in the legal sector.The invention specifically concerns distributed multi-agent systems, federated learning, data sovereignty, secure knowledge exchange, and advanced load balancing techniques for performing legal tasks in one or more jurisdictions. As described in the background section, AI systems can significantly increase the productivity of legal professionals, but encounter three structural limitations: (i) centralization in external clouds, (ii) limited jurisdiction sensitivity, and (iii) absence of a robust mechanism to monitor consistency when multiple teams and talent are involved simultaneously. Often, only a mechanism for confidential clause validation without data sharing is missing. Semantic redundancy elimination across multiple organizations is also lacking. Furthermore, there is a need for multilingual precedent alignment and continuous monitoring of interpretation consistency, not only at the national level, but preferably also across national borders. The present invention breaks one or more of these boundaries, e.g.by providing a system with a distributed hive, beehive, or grid architecture, and with jurisdiction-aware agent orchestration, and preferably also with "federation learning" that keeps the source data local. This creates a system that can operate quickly, securely, reliably, and legally soundly, with demonstrable data separation and compliance with European and other regulations. Referring to Figures, 2025 / 5547 BE2025 / 5547 11 FIG. 1 is a schematic representation of a system according to the present invention, with a multitude of edge units that are connected to each other. In the example of Fig. 1, the system 100 comprises two subsystems, namely: a first subsystem 120a (or "first grid") with six Edges arranged in a first geographical area, e.g. a first state or country (in the specific example of Fig. 1: in Belgium, BE), and a second subsystem 120b (or "second grid") arranged in a second geographical area, e.g. a second state or country (in the specific example of Fig. 1: in France, FR).Naturally, the invention is not limited to this, and systems under the present invention comprise at least two edge genodes which may, for example, be located in one country, but may comprise many more than two edge genodes, e.g. at least ten edge genodes, or at least twenty, 10 or at least thirty, distributed over multiple countries and / or states, e.g. over at least two countries (e.g. BE and NL), or over at least three countries (e.g. BE and NL and FR), or over at least five countries (e.g. BE, NL, FR, DE and IT), or over at least two states (e.g. New York and Texas), or over multiple continents (e.g. a system distributed over at least two European countries and over at least two American states).15 Each edge genodes is configured to be one or more legal AI agents (English: "AI"). to host agents, more specifically specialized legal AI agents, preferably optimized for a specific jurisdiction, legal area, or processing function. However, the AI ​​agents present do not necessarily all need to be effectively active (i.e., running) at every moment.Preferably, each item comprises at least the following five legal AI agents:20 i) a first AI agent, herein also referred to as the JADOP module, where JADOP stands for "Jurisdiction-AwareDynamicAgentOrchestrationProtocol". This module is configured to select a subset of specialized AI agents for processing a posed legal question and / or an incoming legal task. The selection of AI agents can, for example, be based on jurisdiction, language, subject matter, and complexity. The JADOP can, for example, select the most suitable AI agents25 based on jurisdiction- and language-specific relevance scores (e.g., derived from the semantic analysis of the task), and / or taking into account historical performance (e.g., based on historical performance scores of agents in similar cases), and / or taking into account the origin jurisdiction of relevant model updates, and / or with current capacity and latency measurements.In certain execution forms, the orchestration point (which is the node executing JADOP30) distributes not only new tasks, but can also dynamically redistribute remaining tasks to nodes with available capacity. ii) a second AI agent, hereinafter also referred to as the ZK-LVC agent, where ZK-LVC stands for "Zero-KnowledgeLegalClauseVerification-module", or also referred to as the ZK-FKVmodule, where ZK- 2025 / 5547 BE2025 / 5547 12 FKV stands for "Zero-Knowledge-FederatedKnowledgeVerification"-module. This module is configured to demonstrate that a clause meets the applicable legal requirements, but without disclosing the content, e.g. by using cryptographic techniques such as zero-knowledge proofs. Or in other words, this module verifies that contributions meet predefined legal knowledge criteria, without disclosing the underlying documents or origin. The federated learning mechanism operates via the ZK-FKV module, whereby model updates between nodes can be verified and shared without real data or origin information being transferred.This guarantees data sovereignty confidentiality. iii) one-third AI agent, hereinafter also referred to as ALRME module, where ALRME stands for "AdaptiveLegalRedundancyMinimisationEngine". This module is configured to cluster 10 incoming legal questions (or queries) and / or model updates. This module can, for example, perform semantic analysis on incoming tasks (e.g. worldwide), so that legally identical tasks, regardless of language or wording, can be bundled into a single execution, in order to avoid duplication as much as possible. iv) a fourth AI agent, herein also referred to as the CL-PAM module, where CL-PAM stands for "Cross-LanguagePrecedentAlignmentModule". This module is configured to compare and align precedents (e.g., multi-jurisdictional). This module can, for example, translate, compare, and harmonize precedents across multiple jurisdictions and languages, while preserving citations and legal context. v) a fifth AI agent, herein also referred to as the FCDD module, where FCDD stands for "Federated ComplianceDriftDetector".This module is configured to detect and / or harmonize interpretation drift. This module can, for example, monitor interpretation patterns within the federated learning model. New case law that influences interpretations can be integrated worldwide. However, the present invention is not limited to this, and edgenodes can also contain 25 other AI agents or modules, e.g., local AI agents specialized by legal fields / or task type. Some examples will be discussed in Fig. 3. As mentioned above, however, these five AI agents do not necessarily have to be active on every Edge Node. This is particularly the case for the JADOP agent, which is preferably present in every Edge Node (so that every Edge Node is ready to take over the role of "main orchestrator"), but which is preferably active on only a single Edge Node of each subsystem, namely on the Edge Node that is fulfilling the role of "main orchestrator" (or simply put: "task distributor", or "orchestration node") at that moment.A large system with multiple subsystems preferably also has multiple orchestration nodes, configured to monitor the load of other nodes (e.g., 2025 / 5547 BE2025 / 5547 13 all nodes of the same subsystem), distribute new tasks, and / or dynamically shift remaining tasks to nodes with available capacity. The system is set up as a redundant network system, which essentially means that if one of the nodes fails (for whatever reason), the system can in principle simply continue operating without data loss.5 Preferably, at least some nodes of a subsystem, but preferably all nodes, are equipped with an emergency power supply, schematically represented by a battery symbol. Emergency power supply systems are known in standard technology. They can, for example, contain a multitude of batteries (also known as UPS, which stands for "Uninterruptible Power Supply") and / or one or more emergency power generators, e.g. diesel generators.10 Preferably, each unit shall not only have redundant power supplies, but also cooling systems, data center-grade physical and logical security layers, including physical access control, and optionally air gaps, military-grade encryption and continuous penetration tests. This physical and logical isolation is configured to prevent performance loss caused by external attacks or unauthorized access attempts.15 Preferably, denodes are interconnected via a secure peer-to-peer network, allowing collaboration and knowledge sharing without central storage of confidential data. In some implementation forms, model updates are shared between denodes via federated learning with differential privacy security aggregation through such a secure peer-to-peer network.20 The System100 can, for example, be configured as a "hive architecture," a "grid architecture," or a "beehive architecture," i.e., an architecture where tasks are distributed and a failover mechanism is provided in the event a node fails.Preferably, each node of such a system functions largely autonomously (apart from assigning orders), and preferably, coordinating core functions (of the system or subsystem) exist in multiple copies. Preferably, such core functions are proactively replaced before failure, e.g., upon detection of a cyberattack or a prolonged absence of power supply. This ensures the system remains operational as long as possible, even in the event of a complete failure of a single node, and the system automatically recovers when infrastructure becomes available again. In other words, the so-called beehive architecture means that all core functions (such as the orchestrator) are redundantly present on multiple nodes, and that nodes operate largely autonomously. At any given moment, one node assumes the role of "Main Orchestrator" (task distributor), but every node possesses the JADOP module and can take over that role in the event of failure.2025 / 5547 BE2025 / 5547 14 In an implementation form, the takeover of the core function (JADOP) by another is proactively organized, whereby coordinating functions are taken over preventively as soon as a problem is suspected, e.g. in the event of an impending cyberattack or power outage. In this way, the system is very robust against disruptions. 5 In a preferred implementation form, the system is composed of separate continental subnetworks, each consisting of multiple regional edge node clusters. These continental grids are initially operational independently and are later connected to one another via secure interconnection points to form a single intercontinental network. This allows cross-jurisdiction and cross-continent analyses to be performed, whereby all core modules function over the combined agent and knowledge base. The hive-of-beehive-of-grid architecture is preferably also applied at this intercontinental level, so that in the event of the failure of an entire continent, the remaining continents automatically take over core functions and initiate recovery procedures.15 In a specific implementation form, every edge node is configured for (e.g. real-time or "on the fly") local learning based on the full set of laws and regulations applicable in its jurisdiction (i.e. depending on the location where the edge node is physically set up). This includes both single and layered legal systems, for example the distinction between "state law" and "federal law" in the United States, or the combination of national law and supranational law for countries in Europe (such as EU regulations). The edge nodes are preferably configured to gather information on multiple legal domains, such as: contract law, tax legislation, import and export tariffs, competition law, environmental legislation, and other relevant legal or regulatory frameworks. The edge nodes preferably carry out their learning processes entirely locally on source data that never leaves the specific node. The edge nodes are preferably equipped to share abstracted knowledge with other edge nodes. They can, for example,deliver model updates consisting exclusively of anonymized, non-traceable parameters without personal names, concrete contract texts, or other identifiable information. This abstracted knowledge is made available via federated updates to other enodes or grids, enabling real-time cross-border and cross-continental consultation. In a sample scenario, a company in France wishing to conclude a contract with a company in the US state of Delaware can automatically and in real-time consult the relevant knowledge from the US grid (or subsystem) via the European grid (or subsystem). 2025 / 5547 BE2025 / 5547 15 In this process, the contract in Europe is assessed against French national legislation, EU rights, and applicable European trade rules, while in the United States it is assessed against state law (Delaware), federal law, and relevant US customs and import tariffs.Thus exists an integrated, but decentralized legal analysis that covers all relevant layers and domains, without confidential source data leaving the original jurisdiction.5 Referring back to system100 of FIG.1, at least one of the provisions is to receive a legal question and / or a legal task, e.g. in the form of a prompt or a query, which may contain natural text, and / or an image, and / or an electronic document such as a PDF document, a text document, etc. This question or prompt may, for example, 10 originate from a user device110 (e.g. computer, smartphone, PDA, tablet, etc.) at the request of, for example, a Belgian lawyer who asks a question to system100 via the Main Orchestrator101a of the Belgian subsystem120a, or at the request of a French lawyer who asks the system100 via the Main Orchestrator101g of the French subsystem120b. Preferably, all omissions of each subsystem include AI agents configured for processing (e.g. collecting, interpreting, analyzing, abstracting, updating, etc.)knowledge of regulations regarding each of the legal domains (e.g. commercial law, criminal law, tax law, family law, etc.), but preferably only one AI agent per subsystem and per legal domain is active at any given time. This prevents, for example, two different enodes from simultaneously collecting and processing information on family law in a specific country, but distributing different abstract information to the other enodes because, for example, they had not yet processed the same source documents at the moment of forwarding. Referring back to system 100 of FIG.1, it is schematically represented that system 100 receives legal updates, e.g. from user devices, affiliated lawyers, or affiliated courts, or official government agencies, and the like. Such legal updates may include, e.g., new legislative texts, or recent rulings by courts, and the like. FIG.2 is a schematic representation of an exemplary subsystem, as part of system 100 of FIG.1. Where the main objective of FIG.1 to give an overview of the entire system100, is the intention of FIG. 2 rather to explain some sub-aspects, such as e.g.: 2025 / 5547 BE2025 / 5547 16 - the repeated obtaining140 of legal updates, e.g. related to new laws and / or regulations, recent court cases, judicial rulings, etc.) - the dissemination150 of (processes an anonymized) updates by the "Main Orchestrator" of a subsystem120 to other edgenodes of the subsystem, preferably to all other edgenodes of the subsystem;5 - the subsystem120 is very easily scalable160 or expandable by adding additional edgenodes; -subsystem120 is preferably configured for "automatic failover"170, e.g. by activating the JADOP agent on another node. Preferably, in the event of node failure, tasks are automatically redistributed by the failover mechanism10, and preferably also without performance loss; -subsystem120 is preferably updated repeatedly (e.g. continuously), e.g.as part of the "beehive mechanism", so that in the event of a failure or malfunction of one edge node, the system can simply continue working without, or with as little loss of data as possible;15 -Although not shown in Fig. 2, node 101a preferably also includes an emergency power supply; As stated, Fig. 2 shows only a few aspects of subsystem 120. System 100 of Fig. 1 can be seen as a multitude of subsystems 120, which are interconnected by means of secure connections (not shown) for international20 and / or intercontinental linkage. FIG. 3 and FIG. 4 show a schematic representation of how a "legal assignment" received by one of the "Main Orchestrators" (e.g. the French Main Orchestrator), and which relates to two jurisdictions, in the example France (FR) and the American state of Delaware (DE),25 is divided into two sub-assignments, one (on the left) for the French subsystem, one (on the right) for the Delaware subsystem, assuming that the state of Delaware has its own subsystem. More specifically, FIG. shows3. An example scenario of cross-border contract analysis between a French company and an American company in the state of Delaware. The assignment is divided between an edge node 301a of the French subsystem, where node 301a is provided with AI agents (or modules) 312 to 316 for French law, EU law, tax legislation, import / export tariffs, and contract law; and an edge node 301b of the Delaware subsystem node, where node 301b is provided with AI agents (or modules) 312 to 316 for state law, federal law, tax legislation, commercial law, and import / export tariffs. These two nodes are communicatively (albeit indirectly) connected for this assignment by means of a virtual federated connection 330 that exchanges anonymized model parameters. FIG. 4 shows how an assignment (e.g. from a Belgian lawyer) that relates to multiple jurisdictions (e.g.the aforementioned contract between a French company and a 5 company from Delaware arrives at the JADOP module of a first edge node, and how this assignment is assigned and forwarded to AI agents 312 to 316 of an Edge node in France for retrieving legal information regarding France, and how this assignment is assigned and forwarded to AI agents 322 to 326 of an Edge node in Delaware for retrieving legal information regarding Delaware. Each of these AI-10 agents provides information (e.g. a document) that contains specific information for this jurisdiction. Edge node 301 can, if desired (e.g. depending on the specific query or assignment wording), combine the information provided by AI agents 312 to 316 into one document before sending back to the JADOP module 451 that gave the order. Similarly, the aforementioned 301b can, if desired, combine the information provided by the AI ​​agents 322 to 326 into a single document before sending back to the JADOP module 451. FIG.5 shows an example of another cross-border legal processing of data. More specifically, Fig. 5 shows an example scenario where an office in New York simultaneously receives inquiries about the same international ESG regulations20 with a compliance department in Tokyo. ALRME recognizes semantic similarity or equivalence, bundles the tasks, and only one analysis is performed, which is subsequently applicable in both contexts. In the specific example, JADOPAI activates agents in both jurisdictions (in the example: New York and Tokyo). CL-PAM harmonizes precedents. ZK-LCV validates clauses without disclosing content. FCDD monitors consistency.25 The following are a few more statements: •Now that the main principles of the present invention have been explained, the following statements can be better understood and appreciated: •In the present invention, "Learning" takes place locally on source data that remains within the relevant Edge30 nodes; only anonymized model parameters are shared with other Edge30 nodes.This allows the grid to collectively become smarter without confidential documents or personal data ever leaving the organization. 2025 / 5547 BE2025 / 5547 18 • Implementations in accordance with the present invention offer one or more of the following benefits: a distributed network system that is independent of external cloud infrastructures, applying load balancing in a jurisdiction-aware manner, e.g. by distributing jurisdiction-dependent tasks among specialized AI agents, and mechanisms for fully secure knowledge exchange via federated learning.5 • Implementations of the present invention offer a node-level security level equivalent to that of existing data centers. • Implementations of the present invention can verify and share model updates without exposing confidential information (such as names of private individuals or companies).•By making use of Edgenodes, redundancy, and optionally contingency power supplies, systems or subsystems under the present invention can maintain high processing speeds and availability, even in the event of failure of multiple Edgenodes. •Certain implementation forms of the present invention combine a Hive / Grid architecture that provides load balancing and failover, with secure federated knowledge exchange. This delivers a legal AI infrastructure that is scalable, multilingual, secure, and jurisdiction-aware. •Within the system, model updates are shared via federated learning between multiple edge nodes of a subsystem using the ZK-FKV module, which verifies that contributions meet predefined legal knowledge criteria, without disclosing the underlying documents or origin.20 •The beehive architecture gives the system self-healing capabilities. Coordinating functions exist in multiple copies and are proactively taken over, e.g., based on high load, before failure occurs.The design is hierarchical: local enodes form regional clusters, and clusters form national and subsequently continental subnetworks. These continental grids operate autonomously and are linked via secure interconnection points to form an intercontinental network. Even in the event of a large-scale failure in a single region, the whole remains operational; after restoration, the affected part is automatically reintegrated without loss of knowledge or consistency. • Systems based on the present invention use "multilayer federation learning". The system models legislation and regulations on multiple layers simultaneously. In Europe, this comprises national law and supranational EU law; in the United States, the layering of state law and federal law. Additionally, domains such as contract law, tax legislation, competition law, environmental law, and import-export rules are processed in parallel. Each edge learns from local sources, preferably in real-time or on the fly; only abstracted, non-reducible parameters are shared with other edges of the grid.This enables cross-border consultation without compromising confidentiality. •From the description above it should by now be clear that the ultimate version of a system according to the present invention is an Internationally Distributed LegalGrid, with a multitude of jurisdiction-aware AI agents, with jurisdiction-aware task routing,5 (task distribution across different agents), with multilayer federation learning, with zero-knowledge clause validation, with semantic redundancy minimization, with multilingual precedent alignment, and with interpretation drift monitoring, within one coherent grid of edge nodes, combined.