Self-generating cognitive operating system scheduling method based on hyperbolic module space with geodesic boundary

Abstract

The application discloses a self-generating cognitive bottom-layer operating system based on a hyperbolic module space with a geodesic boundary and a scheduling method, and belongs to the technical field of artificial intelligence operating systems. The system adopts a three-layer original self-generating architecture, takes a topological-metric double-layer separation architecture as a core, realizes self-cognition, self-optimization and self-evolution through a kernel-level self-generating system E-M-C ternary recursive closed loop, and is originally provided with core functions such as endogenous geodesic scheduling, topological isolation memory, three-level endogenous safety and directional endogenous evolution. The application constructs the operating system itself into a self-generating cognitive entity, reconstructs an AIOS architecture from a paradigm level, solves the problem of mispositioning of traditional OS and self-generating intelligent characteristics, and makes the hardware utilization rate in a multi-task scene reach more than 85%, and the steady-state running time is increased by 2.5 times. The application can be widely used as an original computing base of embodied robots and industrial agents.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence operating systems, self-generated intelligent agent system software, and AI chip architecture technology, and particularly to a self-generated cognitive underlying operating system and scheduling method based on hyperbolic mode space with geodesic boundaries. Background Technology

[0002] With the development of general artificial intelligence and embodied intelligence technologies, existing AI operating systems based on the von Neumann architecture are no longer able to meet the core requirements of self-generated cognitive systems, becoming a key bottleneck restricting the large-scale deployment of self-generated intelligent agents. The self-generated cognitive system is a specific form of self-generated systems in the field of cognitive computing. It is a self-generated system with the endogenous generation and iteration of cognitive abilities as its core attribute. Based on satisfying the operational closure and steady-state self-maintenance of self-generated systems, it autonomously completes a complete cognitive closed loop of information representation, reasoning and decision-making, rule extraction, and knowledge accumulation, using the information field (C) as its core carrier. Cognitive rules and processing paradigms are generated endogenously by the system, rather than being entirely preset by humans. The boundaries of cognitive abilities are not locked by initial rules and can grow autonomously during operation, distinguishing it from traditional artificial intelligence systems that rely on manually preset models and rule bases. The self-generating system, theoretically based on the self-creation theory proposed by Chilean scientists Humberto Maturana and Francisco Varela, is a unified entity composed of a recursive network of internal production processes. Through its own internal operational closure and endogenous evolution, the system continuously reproduces all its constituent elements and boundaries. The self-generating intelligent agent is the concrete form of the self-generating system at the physical entity level. It is a self-generating system equipped with sensing and execution units, capable of interacting with the physical environment. It possesses characteristics of autonomous steady-state maintenance, endogenous capability growth, and multi-agent autonomous collaboration, distinguishing it from traditional intelligent terminal devices that rely on centralized cloud management and have fixed functions.

[0003] Existing technologies suffer from two major levels of fundamental flaws:

[0004] Level 1: Architectural misalignment defects of traditional general-purpose AI operating systems. Most existing AI operating systems are based on trimmed and packaged versions of traditional general-purpose OSes, essentially still using a binary instruction stream task scheduling model. This is incompatible with the core characteristics of self-generated systems—cognition and existence as one, operational closure and endogenous evolution as isomorphic. They can only simulate cognitive capabilities at the application layer, unable to support self-generated steady state and autonomous evolution from the computational level. This leads to unstable operation of self-generated intelligent agents, easily causing actuator malfunctions and incorrect computational results.

[0005] The second level: the paradigmatic limitations and defects of existing hyperbolic geometry AI solutions. Some studies have attempted to introduce hyperbolic geometry into the field of AI representation and scheduling, but they only use hyperbolic space as a distance metric and feature representation tool. All scheduling and management rules are still preset manually. The system itself does not have the self-awareness and self-iteration attributes, and is essentially still a tool-type system that is "organized by others". Without establishing a two-layer architecture of topology-metric separation, it is impossible to achieve kernel-level self-optimization of "using intelligence to serve intelligence".

[0006] The third level: There is a paradigm conflict in the task access mechanism. The task access of the existing operating system adopts the direct instruction injection method, which allows external tasks to directly interfere with the kernel scheduling logic, thus breaking the operational closure of the system. It cannot be adapted to the endogenous steady-state architecture of the self-generated system, and it also lacks a standardized geometric unified access interface. Summary of the Invention

[0007] To address the aforementioned shortcomings of existing technologies, this invention provides a self-generated cognitive underlying operating system and scheduling method based on a hyperbolic mode space with geodesic boundaries. It reconstructs the AI ​​operating system architecture from a paradigm perspective: using a hyperbolic mode space with geodesic boundaries as a unified geometric foundation, separating the topological element parameter layer and the metric representation layer as the core architecture, and using the Weil-Petersson metric as the global metric standard, the operating system itself is constructed as a complete self-generated system energy mechanism E-matter structure M-information field state C three-element self-generated cognitive entity. Through a kernel-level self-generated closed loop, it achieves full-link self-cognition, self-optimization, and self-evolution in resource scheduling, memory management, security protection, and version upgrades, truly realizing the self-generated operating paradigm of "using intelligence to serve intelligence," and providing native underlying computing support for the next generation of general artificial intelligence.

[0008] The self-generating system employed in this invention is based on a monistic two-state three-body core architecture: a monistic ontology is constructed with operational identity as the sole core convergence objective throughout its entire lifecycle. It possesses a self-referential and other-referential dual-modal operating mechanism, continuously and recursively coupled by three irreducible functional primitives: information field state C, energy mechanism E, and matter structure M. These primitives jointly generate and maintain the system's stable operation, exhibiting core capabilities of endogenous evolution, autonomous decision-making, environmental adaptation, and steady-state self-maintenance. Under the premise of stable topological boundaries and environmental constraints, this closed-loop recursive mapping satisfies the Lipschitz compression condition, and a unique globally asymptotically stable steady-state fixed point exists within its corresponding complete state space. The system can autonomously converge to this steady state from any legal initial state. The system itself is both the executor of the operating process and its product, capable of autonomous existence without the need for continuous external injection of control rules, unlike other organizational systems that rely entirely on externally preset instructions and are controlled by external entities.

[0009] To achieve the above objectives, this invention adopts a three-layer native self-generating architecture design, in which the operating system itself constitutes a complete self-generating cognitive entity. The core technical solution is described in five aspects:

[0010] In a first aspect, the present invention provides a scheduling method for an autogenous cognitive operating system based on a hyperbolic mode space with geodesic boundaries.

[0011] The corresponding method is deployed on a heterogeneous computing hardware platform including FPGA, NPU, and CPU. The operating system itself acts as a self-generating cognitive entity, constructing a two-layer architecture that separates the topology meta-parameter layer and the metric representation layer. It achieves self-cognition, self-optimization, and self-evolutionary endogenous scheduling through a kernel-level EMC ternary recursive closed loop. The core includes four execution steps, each of which is described in detail below:

[0012] Step S1: Kernel Initialization

[0013] This step is the core preparation stage for system cold start, completing the construction of the geometric computing base, the deployment of hardware acceleration units, and the configuration of the secure operating baseline. It is a prerequisite for all subsequent scheduling and evolution operations. The execution entity is the CPU core, and the core operations are divided into four parts:

[0014] Global Modulus Space Topological Element Parameter Initialization: A global hyperbolic modulus space with geodesic boundaries is constructed. The CPU kernel executes the topological element parameter initialization of this space. The topological element parameter layer stores the baseline equivalent genus and the baseline geodesic boundary number, completing the baseline parameter configuration of the topological element parameter layer. The hyperbolic modulus space with geodesic boundaries refers to a complete hyperbolic surface parameterized space with two types of topological parameters: genus number and geodesic boundary cycle number. It is used to uniformly represent the cognitive rules and existence state of the system, and can naturally bear endogenous characteristics such as hierarchical nesting, boundary steady state, and semantic exponential divergence. In this patent, all modulus spaces are engineering equivalent, implemented using a real hyperbolic space approximation, and are not strictly continuous modulus spaces in the Riemannian geometric sense. The topological element parameter layer stores two core baseline parameters: equivalent genus and geodesic boundary number. The equivalent genus refers to the baseline topological complexity parameter stored in the topological parameter layer. It is an integer baseline value corresponding to the system's basic cognitive level, and its value determines the upper limit of the number of independent functional closed loops the system can support. The default initial equivalent genus for server-level standard configurations is 2, corresponding to a mid-range cognitive ability level, which can cover the parallel requirements of eight typical tasks. The geodesic boundary number corresponds to the number of interface channels for interaction between the system and the external environment, and together with the genus, determines the scale of the metric space.

[0015] Metric Representation Layer Carrier Construction: The metric representation layer adopts a d-dimensional Poincaré sphere model as an approximate computation carrier. The d-dimensional Poincaré sphere model refers to a d-dimensional conformal sphere model in hyperbolic space, which is the standard numerical implementation carrier of hyperbolic geometry. A two-dimensional special case is the Poincaré disk, which can maintain angular invariance through conformal mapping, adapting to the hierarchical representation requirements of cognitive semantics. The metric space dimension and topological parameters satisfy an approximate relationship. Where g is the equivalent genus and n is the number of geodesic boundaries, the size of the metric space is determined by topological parameters, not the other way around, ensuring the topological dominance of the architecture. In the d-dimensional Poincaré sphere model, two points and The general formula for calculating geodetic distance is: This formula is an approximate geodesic distance calculation form under the Weil-Petersson metric, which can be directly extended to any dimension and serves as a unified benchmark for distance measurement across the entire system.

[0016] Hardware-based geometric operator implementation: The Weil-Petersson metric calculation kernel and the Riemann Langevin geodesic flow operator are embedded into the FPGA logic unit, achieving hardware acceleration of core geometric operations. The Weil-Petersson metric refers to a natural Riemannian metric defined on modulus space, possessing coordinate-independent isometry invariance, eliminating the need for manually setting similarity rules, and ensuring the objectivity and consistency of calculations from a mathematical foundational level. The geodesic distance calculation adopts a three-stage fully pipelined hardware architecture, sequentially completing coordinate reading and mapping, WP metric tensor matching, and geodesic distance integral output. At the same operating frequency, the number of clock cycles per step is reduced by more than 80% compared to the serial implementation in general-purpose CPU software. The CPU software implementation requires approximately 15 clock cycles, while this solution only requires 3. Combined with the parallel pipeline characteristics of FPGA, the throughput per unit time is increased by more than 5 times compared to general-purpose CPUs. The Riemann-Lanczwan geodesic flow operator refers to a generalized geodesic gradient flow dynamics operator with intrinsic inertial characteristics and dissipative damping terms, which is the core dynamic unit driving the evolution of system and task states. In this patent, it is an engineering deterministic approximation version. The inertial characteristics are reflected by the second-order dynamic equations themselves, and the fluctuation characteristics are simulated by the equivalent simulation of initial condition perturbation and environmental noise terms. The core adopts the geodesic gradient flow form with dissipative damping, and the complete form can be extended to include Riemann Brownian motion fluctuation terms.

[0017] Safety baseline and closed-loop parameter configuration: Configure the operational consistency safety baseline. The safety baseline is set to an operational consistency strength of not less than 0.7. If it is lower than this threshold, the system self-healing calibration will be triggered. Set core safety parameters such as topology fluctuation threshold and state deviation threshold as the judgment benchmark for subsequent verification. Configure the initial parameters of the kernel EMC self-generated closed loop, including the initial energy distribution ratio, rule learning rate, damping coefficient, etc., to complete the initial startup configuration of the closed loop.

[0018] The two-tier architecture features a bidirectional interaction mechanism, ensuring the consistency between topology and metrics:

[0019] Forward mapping from topology to metric: When the topology parameters change, the metric space dimension is adjusted synchronously according to the dimensional approximation formula. The smooth state transition is completed through orthogonal dimensionality increase or dimensionality reduction projection. The single state offset does not exceed 1% of the current feature scale, avoiding system oscillation caused by topology changes.

[0020] Metrics-to-topology feedback: The metric layer outputs continuous cohomology calculation results at preset intervals, feeding back and updating the floating-point estimates of effective functional genus; the continuous deviation judgment window is 100 cycles. When the deviation ratio within the window exceeds 80% and the phase transition or elastic control conditions are met, the baseline values ​​of the topology meta-parameter layer are formally updated, completing the closed-loop iteration of the architecture and avoiding architecture jitter caused by instantaneous disturbances. The system characteristic scale. Defined as the average geodesic distance from all state points to steady-state points in the global modulus space, it is updated every 1000 cycles after cold start and serves as a unified benchmark for all distance class thresholds.

[0021] Step S2, Task Process Mapping

[0022] This step serves as the entry point for task access. It uses the pointer mapping operator to achieve a geometrically normalized mapping from external AI tasks to the internal topological computation space. This is the foundation for multi-task isolation and parallel scheduling. The entire process adheres to the principle of operational closure; external inputs only alter the subsystem boundary conditions and do not interfere with the core kernel operating rules. Operational closure is a core organizational attribute of a self-generating system, meaning that all external inputs only indirectly affect the internal evolution by changing system boundary conditions, without modifying the kernel dynamics rules and core topological structure. Operational closure is the prerequisite for the long-term stable maintenance of operational identity, and operational identity is the concrete verification standard for operational closure at the system's operational state level.

[0023] The core logic of this step is to uniformly map various AI cognitive tasks and environmental inputs into independent submodule space processes, allocate initial genus quotas, geodesic boundary interface resources, and initial state coordinates to each process, and establish a one-to-one correspondence between tasks and subspaces. The mapping operator uniformly maps the functional requirements, performance constraints, and environmental perturbations of external tasks into standard geometric parameters of the corresponding submodule space, and on this basis, completes the allocation of three types of resources:

[0024] Genus quota allocation: Based on the cognitive complexity level of the task, an initial equivalent genus quota is allocated to each subspace to determine the cognitive complexity that the task can bear; simple control tasks are allocated a genus 1 quota, and complex reasoning tasks are allocated a genus 2 or higher quota. The quota limit shall not exceed 60% of the global equivalent genus to avoid a single task occupying too much topology resources.

[0025] Geodesic boundary interface allocation: A corresponding number of geodesic boundary interfaces are allocated to each subspace as channels for interaction between the task and the external environment and other tasks; input and output data are encoded and mapped inside and outside the module space through the boundary interfaces, and the number of boundaries matches the interaction requirements of the task.

[0026] Hardware computing power weight allocation: An initial hardware computing power weight is allocated to each subspace, corresponding to the quota ratio of three types of hardware resources: FPGA, NPU, and CPU. The initial weight is positively correlated with task priority and computing demand, and is dynamically adjusted according to the scheduling strategy.

[0027] Each sub-module space process has independent state coordinates, evolution parameters, and memory blocks, which are naturally isolated by geodesic boundaries. There are no direct data paths between tasks, which ensures the cognitive isolation of multiple tasks from the topology level and avoids cognitive crosstalk between tasks.

[0028] Step S3: Endogenous geodesy priority scheduling

[0029] This step is the core execution phase of the method. It combines hardware-accelerated computation with iterative, endogenous rule-based processing to achieve intelligent scheduling and parallel evolution of multiple tasks. Simultaneously, the kernel EMC ternary closed-loop operates synchronously, enabling autonomous optimization of scheduling capabilities. The execution process consists of four stages: computation, decision-making, execution, and closed-loop iteration.

[0030] Hardware parallel computing: The free energy potential and global geodesic distance of each task subspace are calculated in parallel using an FPGA hardware pipeline. The free energy potential represents the current unexpectedness and deviation of the task, and is a core quantitative indicator of the task's urgency; the global geodesic distance represents the convergence distance from the task state to the global steady state point, and is a core quantitative indicator of the task's convergence difficulty. The formula for calculating the free energy potential function is: in The regularization coefficient is calibrated through grid search to balance the weights of unexpectedness and topological constraints; unexpectedness is perceived. That is, the approximate WP geodesic distance between the current state module and the predicted state module; the larger the distance, the higher the unexpectedness; topological regularization term. This refers to the relative deviation between the equivalent genus of the local state and the global genus, used to constrain the local state from deviating from the global topological baseline.

[0031] Endogenous strategy decision-making: Combining the scheduling strategy generated endogenously from the C-dimensional rule base, the basic calculation results are corrected to construct a dynamic priority queue. The basic priority is calculated using a weighted average of three factors: free energy, geodesic distance, and deadline, with the initial weight values ​​determined endogenously by the global genus. The C-dimensional module autonomously learns the task load patterns through inductive reasoning, generates dynamic correction coefficients, and fine-tunes the priorities to ensure that the scheduling strategy continuously adapts to the actual load scenario.

[0032] Hardware resource allocation and execution: Heterogeneous hardware computing resources are allocated according to geodesic gradient scheduling rules to drive the parallel evolution of multi-task subspaces. The geodesic gradient scheduling rule prioritizes tasks with high free energy and farthest from steady state, enabling them to converge rapidly towards steady state along the geodesic gradient; high-priority tasks can preempt hardware resources from low-priority tasks to ensure the real-time performance of urgent tasks. The global evolution process strictly satisfies operational identity constraints to ensure the stability of the system's core structure. Task state evolution follows the Riemann-Langevan geodesic dynamics equations, specifically in the form: in As an evolutionary path, For the Levi-Civita communication corresponding to the WP metric, Let be the free energy potential function; The velocity damping coefficient corresponds to the resistance in the opposite direction of velocity and is the core source of damping for system convergence. This coefficient is calibrated through grid search of 50 different scenarios, with the weighted sum of steady-state convergence speed and overshoot as the optimization objective. For Hessian dissipation coefficient, this term is curvature-dependent dissipation. It can adaptively adjust the damping magnitude according to the local potential energy curvature, increasing the damping in steep potential energy regions and decreasing the damping in gentle regions. Compared with fixed linear damping, it can improve the convergence speed by more than 20% and reduce the overshoot. The coupling coefficient. When Furthermore, when the Hessian dissipation term is dominant, the system has global dissipation, satisfying the Lipschitz convergence criterion.

[0033] The complete definition of an interactive traction item is: The free energy increment field corresponding to the input of the environment and the task. The environmental coupling coefficient is typically 0.1, and its physical meaning is the modulation effect of external tasks and sensing signals on the system's evolution direction.

[0034] The above is the deterministic skeleton equation; in scenarios with environmental noise, random fluctuation terms satisfying the definition of Riemann-Brownian motion can be additively injected on the right side to form the complete Riemann-Langevan equation, which can be tailored according to accuracy requirements in engineering implementation.

[0035] Kernel-level self-generated closed-loop iteration: During the scheduling process, E-dimensional energy management, M-dimensional parallel computation, and C-dimensional rule iteration constitute a kernel-level self-generated closed loop, continuously and autonomously optimizing the scheduling strategy. Among them, the hardware abstraction layer scheduling uses M-dimensional parallel computation as the main execution link, and the E-dimensional energy management module outputs power constraints and energy efficiency priorities to it. The two have a constraint-execution collaborative relationship. The core scheduling logic belongs to the M-dimensional dimension, and the energy constraints belong to the E-dimensional dimension.

[0036] As a preferred implementation, the self-generation of C-dimensional rules adopts a self-referential-other-referential bimodal dynamic weight adjustment mechanism. The weights are endogenously adjusted according to the intensity of the identity of the operation, and the rule update intensity is adaptively constrained by the steady state of the system. Candidate rules must be implemented only after the survival is pre-verified. This forms a full-dimensional synergy with the dynamic control of the self-referential-other-referential energy ratio in E-dimensional rules, jointly ensuring the endogenous balance of the bimodality.

[0037] The kernel EMC closed loop is a three-level cascaded mapping: the energy allocation mapping E, the state evolution mapping M, and the rule update mapping C are sequentially combined to form the closed-loop mapping. According to the composition property of Lipschitz mappings, the total Lipschitz constant after combining multiple Lipschitz mappings is equal to the product of the Lipschitz constants of each sub-mapping. Therefore, the total closed-loop Lipschitz constant satisfies... If and only if At that time, the closed-loop system converges globally, and the system reaches steady-state stability.

[0038] Step S4: Global Identity Self-Check

[0039] This step is an inherent security safeguard of the system. It ensures the stability of the system topology and state through periodic checks, promptly handles abnormal tasks, and iterates security policies, serving as a core barrier for the long-term stable operation of the system. The execution entity is the CPU scheduling kernel, triggered according to a preset cycle. The core process includes three stages: verification, handling, and feedback.

[0040] Identity verification operation: Verifies the topological invariance and state stability of the global module space and each task subspace, employing a dual criterion. Operation identity strength. The formula is a weighted average of topological stability and state stability, and is as follows: in For effective inequality fluctuations, The geodesic distance from the current state to the steady-state point. Based on the equivalent inequality, The system characteristic scale is defined by a value ranging from [0,1], with higher values ​​indicating stronger system homogeneity. The topological invariance check core monitors the fluctuation range of the effective functional genus estimate and the jump in the rank of the first-order homology group; the state stability check core monitors the geodesic distance from the state point to the globally steady-state fixed point.

[0041] Anomaly Handling: For abnormal tasks that fail verification, the hardware memory protection unit, in conjunction with the triggered isolation mechanism and topology calibration, prevents the anomaly from spreading and affecting the entire system. The memory protection unit refers to the storage area protection circuit integrated into the hardware memory management unit, which can achieve hardware-level real-time interception of memory access through configuration of area registers.

[0042] The system employs a tiered approach to handling anomalies: minor anomalies are prioritized for self-healing through parameter fine-tuning; moderate anomalies involve task isolation and topology calibration; and severe anomalies trigger task termination and global calibration to ensure overall system security.

[0043] Strategy Iteration Feedback: Verification data is synchronously fed back to the C-dimensional rule base, and the anomaly identification strategy and self-healing parameters are iteratively updated, so that the system's ability to identify and handle anomalies continues to improve with operation, realizing the endogenous growth of security capabilities.

[0044] As a preferred implementation, the endogenous geodesy priority scheduling adopts a five-layer asynchronous scheduling architecture. The priority calculation adopts a second-order mechanism that combines the basic formula with endogenous correction. The weight coefficient is endogenously determined according to the system genus and dynamically adjusted through autonomous learning.

[0045] As an optional implementation, it can be configured with extended functions such as endogenous predictive cognitive memory dynamic management, heterogeneous hardware abstraction layer endogenous scheduling, standardized system call interface, three-level endogenous security in-depth defense, resource recycling and rule accumulation, system endogenous directional evolution, bidirectional elastic topology control, and distributed multi-system self-coordination.

[0046] Secondly, this invention provides a self-generating cognitive operating system based on hyperbolic mode space with geodesic boundaries.

[0047] The corresponding system is deployed on a heterogeneous computing hardware platform that includes FPGA, NPU and CPU. It adopts a three-layer native self-generated architecture, which consists of a geometric kernel layer, an endogenous cognitive service layer and an application interface layer from bottom to top. The operating system itself constitutes a complete EMC three-element self-generated cognitive entity.

[0048] The geometric kernel layer provides a unified geometric computing foundation for the entire system, supporting a two-layer separation architecture for topological parameters and metric representations. The intrinsic cognitive service layer forms a kernel-level self-generating closed loop through three core modules: E-dimensional, M-dimensional, and C-dimensional, enabling the intrinsic growth of control capabilities. The application interface layer encapsulates standardized system call APIs, supporting various AI applications. All modules communicate through the kernel's internal bus, forming a bottom-up capability output and a top-down call feedback.

[0049] As a preferred implementation, the endogenous cognitive service layer is extended with functional modules such as cognitive memory management, rule storage management, system evolution control, hardware abstraction scheduling, distributed collaboration, and elastic topology control, covering the core capabilities of the operating system across all dimensions.

[0050] Thirdly, this invention provides a self-generating cognitive AI chip.

[0051] The corresponding chip integrates the aforementioned self-generated cognitive underlying operating system and has a built-in geometric kernel hardware acceleration circuit, including a geodesic distance pipeline, a free energy gradient acceleration core, and a dedicated identity verification core. The three types of operators adopt a fully pipelined parallel architecture, which greatly improves the execution efficiency and throughput per unit time of core geometric calculations.

[0052] Fourthly, the present invention provides a self-generating intelligent agent device.

[0053] The corresponding equipment includes a sensing unit, the aforementioned self-generated cognitive AI chip, and an execution unit. The signal output terminal of the sensing unit is connected to the sensing input interface signal of the self-generated cognitive AI chip, and is used to collect external environmental information and equipment operating status information and transmit them to the self-generated cognitive AI chip. The control input terminal of the execution unit is connected to the control output interface signal of the self-generated cognitive AI chip, and is used to receive and execute the control commands output by the self-generated cognitive AI chip. The self-generated cognitive AI chip, through its built-in self-generated cognitive underlying operating system, relies on a two-layer architecture that separates the topology meta-parameter layer and the metric representation layer, as well as a kernel-level EMC three-element self-generated closed loop, to achieve device steady-state autonomous maintenance, endogenous evolution of system capabilities, and parallel scheduling of multiple cognitive tasks.

[0054] Fifthly, the present invention provides a computer-readable storage medium.

[0055] The corresponding storage medium stores an operating system program, which, when executed by a processor, implements all the steps of the above-mentioned self-generated cognitive operating system scheduling method based on hyperbolic mode space with geodesic boundaries.

[0056] System Architecture and Core Modules. This operating system adopts a three-tier native self-sustaining architecture design. The following is a detailed description of each layer's modules:

[0057] First layer: Geometric kernel layer

[0058] The geometric kernel layer serves as the mathematical foundation of the operating system, replacing the traditional kernel scheduling core. All computations are performed within a hyperbolic modulus space with geodesic boundaries, employing a two-layer architecture that separates the topological parameter layer and the metric representation layer. The topological parameter layer stores core topological parameters such as equivalent genus and geodesic boundary number, defining the system's cognitive levels and essential capability boundaries. The metric representation layer uses a d-dimensional Poincaré sphere model as an approximate computational carrier, undertaking specific geometric operations and state evolution; the dimension and topological parameters satisfy an approximate relationship. Topological parameters determine the size of the metric space, not the other way around.

[0059] The two-tier architecture features a bidirectional interaction mechanism, ensuring consistency between topology and metrics:

[0060] Forward mapping from topology to metric: When the topology parameters change, the metric space dimension is adjusted synchronously according to the dimensional approximation formula. The smooth state transition is completed through orthogonal dimensionality increase or dimensionality reduction projection. The single state offset does not exceed 1% of the current feature scale, avoiding system oscillation caused by topology changes.

[0061] Reverse feedback from metric to topology: The metric layer outputs continuous cohomology calculation results according to a preset period, and feeds back the floating-point estimate of the effective functional genus; the continuous deviation judgment window is 100 periods. When the deviation ratio in the window exceeds 80% and the phase change or elastic control conditions are met, the baseline value of the topology meta-parameter layer is formally updated to complete the closed-loop iteration of the architecture and avoid architecture jitter caused by instantaneous disturbances.

[0062] The global module space is stored in the kernel as structured data, containing two types of data: topology meta-parameter layer and metric representation layer. The topology meta-parameter layer data includes the global equivalent genus baseline value and the number of geodesic boundaries. The metric representation layer data includes discrete grid dimensions, grid coordinate arrays, an approximate Weil-Petersson metric tensor array, a subspace process pointer array, the current number of subspaces, the global free energy value, and the coordinates of endogenous steady-state fixed points. Each submodule space process corresponds to independent topology resources and state data, including subspace genus, number of boundaries, grid block index, computing power weight, priority, and current state coordinate fields.

[0063] The geometry kernel layer contains the following core modules:

[0064] Modular space initialization module: Constructs a global hyperbolic modular space with geodesic boundaries, completes the initialization of topological parameters such as genus and boundary, and establishes a unified geometric carrier with a two-layer architecture;

[0065] WP Metric Calculation Module: Configures Weil-Petersson metric parameters and provides basic geometric calculation capabilities such as geodesic distance, Levi-Civita connection, and Hessian matrix, providing a unified measurement benchmark for the entire system. The geodesic distance calculation adopts a three-stage fully pipelined hardware architecture, sequentially completing coordinate reading and mapping, WP metric tensor matching, and geodesic distance integral output. At the same operating frequency, the number of clock cycles per step is reduced by more than 80% compared to the serial implementation in general-purpose CPU software. Combined with the parallel pipeline characteristics of FPGA, the throughput per unit time is increased by more than 5 times compared to general-purpose CPU.

[0066] Riemann Langevin Geodesic Flow Engine: It realizes geodesic gradient flow dynamics calculations that include intrinsic inertial characteristics, dissipative damping terms, and interactive traction terms, driving the evolution of system and mission states, and is the core power engine for system operation.

[0067] Identity verification kernel: It has a built-in dual judgment standard for operation identity and periodically checks the global and local topological stability, which is the core barrier for the inherent security of the system.

[0068] Second layer: Endogenous cognitive service layer

[0069] The intrinsic cognitive service layer is the core carrier of the operating system's inherent attributes. Corresponding to the core capabilities of traditional OS such as processes, memory, scheduling, and storage, it is entirely implemented based on a geometric kernel and achieves self-optimization through a kernel-level EMC self-generated closed loop. This kernel-level EMC self-generated closed loop refers to the operating system's own intrinsic operating paradigm, encompassing three dimensions: E-dimensional energy management, M-dimensional parallel scheduling, and C-dimensional rule self-generation. These three dimensions form a recursive iterative closed loop, enabling the autonomous growth of the operating system's management capabilities without the need for external rule injection. Specifically, the E-dimensional layer, or evolutionary energy layer, is the energy dimension of the operating system's self-generated closed loop, responsible for the dynamic allocation of computing power and energy consumption, energy efficiency monitoring and optimization, providing energy drive for system operation; the M-dimensional layer, or parallel computation layer, is the execution dimension of the operating system's self-generated closed loop, responsible for multi-task parallel scheduling, operator hardware mapping, and state evolution execution, serving as the carrier for cognitive rules; and the C-dimensional layer, or information creation layer, is the rule dimension of the operating system's self-generated closed loop, responsible for rule accumulation, pattern learning, policy generation, and self-diagnosis, serving as the intrinsic source of system intelligence.

[0070] The three core dimensions constitute a kernel-level self-generating closed loop. The operating logic is as follows: the C-dimensional rule self-generation module generates control rules and optimization strategies; the E-dimensional energy control module executes the dynamic allocation of energy and computing power according to the rules; the M-dimensional parallel scheduling module drives the parallel evolution of multiple tasks according to the rules; and the operating status and results are fed back to the C-dimensional module to realize rule iteration, forming continuous endogenous iteration. The control capability of the operating system improves autonomously with operation.

[0071] The E-dimensional energy management module employs a marginal energy efficiency gradient balancing strategy to achieve dynamic power allocation. The specific engineering implementation logic is as follows:

[0072] Energy allocation algorithm: The power allocation vector satisfies the optimal condition that the marginal energy efficiency of each channel is equal, that is... The allocation vector is solved iteratively by gradient descent to ensure that the task output per unit energy input is maximized.

[0073] Dynamic adjustment of self-pointing to other-pointing energy ratio: Setting the energy allocation ratio between self-pointing maintenance and other-pointing tasks. When the intensity of operation is the same hour, The semantic entropy increases by 0.02 per cycle, prioritizing the system's own steady state; the semantic entropy uses a normalized value, and the rate of decrease is the relative decrease within a unit cycle. When the semantic entropy decrease rate is less than 0.001 per cycle, The value decreases by 0.01 per cycle, allocating resources more towards task execution, with the value constrained within the range of [0.2, 5.0]. The semantic entropy is calculated for the rule base using the following formula: The rules are divided into four main categories based on their functions: scheduling, security, memory management, and coordination. For the first Information entropy of class rules The semantic entropy represents the mutual information between rule classes, corresponding to the strength of the association and coupling between different functional rules; the higher the semantic entropy value, the more dispersed the rule base and the lower the cognitive integration. The rate of decrease is the amount of semantic entropy reduction per unit period, which characterizes the speed at which the system's understanding deepens.

[0074] Energy efficiency quantification method: energy utilization efficiency A sliding window of 100 periods was used as the core evaluation indicator for energy allocation optimization; cognitive energy consumption. Using the actual power consumption of the hardware as the conversion benchmark, with the unit being joules, the total energy consumed to complete cognitive computing is calculated by weighting the runtime and rated power consumption of each hardware unit.

[0075] The C-dimensional rule self-generation module employs structured rule representation and incremental inductive algorithm to achieve endogenous rule iteration. The specific engineering implementation logic is as follows:

[0076] The rules are represented in a structured manner: all rules are stored in the form of triplets, namely (conditional feature vector, action policy vector, confidence weight), which correspond one-to-one with the storage structure of the rule base. The conditional feature vector corresponds to the task scenario and system state characteristics, the action policy vector corresponds to operations such as scheduling, allocation, and disposal, and the confidence weight corresponds to the validity of the rule verification. Each rule corresponds to a steady-state attractor in the modulus space. The conditional feature vector of the rule corresponds to the coordinate range of the attractor's attraction domain, the action policy vector corresponds to the geodesic flow direction at the attractor, and the confidence weight corresponds to the potential energy depth of the attractor. The rule base as a whole constitutes the free energy topography of the modulus space.

[0077] Inductive generation algorithm: The incremental DBSCAN clustering algorithm is used to cluster the "feature-action" sample pairs in the history of the operation. The center of each cluster corresponds to a general rule. The confidence weight is the average task success rate of the samples in the cluster. The clustering results are updated incrementally when new samples are continuously added, without the need for full recalculation.

[0078] Rule conflict resolution: When the cosine similarity of the antecedents of two rules is greater than 0.8 and the cosine similarity of the consequents is less than 0.3, it is determined to be a rule conflict; the rule with the higher confidence weight is retained as the main rule, and the low-weight rule is downgraded to a special case rule, which only takes effect in specific sub-scenarios, thus ensuring the consistency of the rule base.

[0079] Based on this, expand the supporting functional modules:

[0080] The cognitive memory management module uses discrete grid points in the modular space as cognitive memory units. It employs a WP volume-weighted paging mechanism to achieve dynamic allocation, reclamation, and defragmentation, supporting endogenous predictive scheduling and boundary-isolated access control. Cognitive memory refers to topological storage resources based on discrete grid points in the global modular space, corresponding to the storage carrier of cognitive states and rules, distinct from traditional linear byte memory. A single grid point stores the state vector, rule weights, and evolutionary gradient information corresponding to its location in the modular space, representing the complete data of a micro-level cognitive state in the cognitive system. The physical storage size of a grid point is determined by the dimension of the state vector; one unit of WP volume corresponds to one standard discrete grid point, and a single grid point corresponds to 256 bytes of physical memory.

[0081] Rule storage management module: Using steady-state attractors as storage units, it realizes the persistent storage, retrieval and reuse of cognitive rules, corresponding to the file system function of traditional OS, and provides data support for the self-generation of C-dimensional rules.

[0082] System evolution control module: Monitors global free energy accumulation and topological invariants, and combined with C-dimensional bottleneck identification results, triggers system-level topological phase transitions to achieve endogenous directional version upgrades without the need for external patches.

[0083] The effective functional genus refers to the topological representation of the number of independent, self-sustaining functional closed loops in a system, calculated using the continuous cohomology method of the state graph. It is a continuous floating-point estimate that characterizes the essential emergent nature of the functional level, distinct from the static integer definition of pure topological genus. For the operating system state graph, the filtering scale... The threshold for the strength of association between corresponding nodes, expressed in normalized association degree; the strength of association between nodes is calculated by weighting normalized communication frequency, resource dependency, and functional coupling, with a value range of [0,1]. A higher value indicates a stronger association between the two kernel entities. As the threshold is gradually increased from 0, edges with a correlation strength higher than the threshold are gradually added to the graph, corresponding to the "birth" of topological features; as the threshold is further increased... The disappearance of a topological feature corresponds to its "death," and the difference between the birth and death thresholds is the lifespan of that feature. A longer lifespan in a topological feature corresponds to a more stable functional closed loop in the system, and the effective functional genus obtained after weighting can characterize the true functional complexity of the system. Specifically, it is calculated by weighting the first-order homology group rank under multi-scale filtering by lifespan length, as shown in the formula: in Filtering scale The first-order Betti number, i.e. the rank of the first-order homology group, is an integer value; The lifetime length of this topological feature is given, and the final output is a continuous estimate with one decimal place.

[0084] The joint closed-loop Lipschitz constant refers to the quantitative criterion for the steady-state convergence of the system, which is decomposed into E / M / C three-dimensional components. The closed loop is stable and convergent when the product of the three factors is less than 1; among them The Lipschitz constant for energy distribution mapping is measured by the ratio of input power perturbation to output parallelism variation; The Lipschitz constant for the state evolution mapping is measured by the ratio of the initial state perturbation to the steady-state output deviation. The Lipschitz constant for the rule update mapping is measured by the ratio of feedback perturbation to rule weight change. In engineering practice, the finite perturbation method is used for online calculation. A small-amplitude white noise perturbation with a relative amplitude of 1% is applied continuously for 10 times, and the average value of the output change is used to calculate the Lipschitz constant for the corresponding dimension. This ensures measurement accuracy while avoiding the impact of perturbation on the normal operation of the system.

[0085] Distributed collaboration module: Supports the construction of joint module space by multi-system boundary coupling, forming higher-order self-generated cognitive entities, realizing endogenous collaborative scheduling and consistency maintenance without a central node, and supporting communication-free implicit collaboration based on shared physical fields.

[0086] The joint Chern number refers to the topological invariant of the joint module space, used to determine the essential emergent degree after multi-system coupling. It is calculated using an engineering approximation method that combines the eigenvalues ​​of subsystems with coupling corrections. This calculation method is an equivalent quantification method for determining topological emergence, not a strictly complex vector bundle-based Chern number; its physical meaning is a weighted sum of the subsystem topological complexity and boundary coupling contributions, used to quantitatively characterize the topological globality of the joint system, with the numerical increment corresponding to the intensity of topological emergence. The eigenvalues ​​of each independent subsystem are also considered. Due to its equivalent inequality With boundary number The mapping is obtained, and the mapping relationship is as follows: This mapping relationship was obtained based on the calibration fit of 20 sets of topological emergence samples, and the goodness of fit was... It can effectively characterize the topological complexity of the subsystem itself and is an engineering equivalent approximation for emergence determination.

[0087] The joint Chern number of the joint modulus space is the sum of the eigenvalues ​​of each subsystem plus a boundary coupling correction term: in The number of coupling boundaries, The total number of boundaries for all subsystems. This is the coupling correction coefficient, calibrated based on 20 sets of topological emergence experiments, characterizing the contribution weight of boundary coupling to the overall topological complexity. Essential topological emergence is determined when the joint Chern number increment is greater than 0.8 and the effective genus increment is greater than 0.

[0088] Elastic Topology Control Module: Autonomously predicts system load and environmental non-stationarity, performs dynamic pruning and expansion of topology-level resources, and adapts to load fluctuation scenarios. The steady-state upper threshold is set at 1.5 times the steady-state free energy benchmark, with continuous triggering cycles. This parameter is based on load fluctuation statistics to avoid falsely triggering topology adjustments. The trigger boundary adopts a dual-response mode: normal load fluctuations only adjust computing power weights and scheduling priorities, which are parameter-level adjustments and do not change the topology structure; topology-level upgrades or downgrades are only triggered when the environmental time series long-range correlation index Hearst exponent H > 0.7 and the free energy is higher than the steady-state upper threshold for N consecutive periods, thus avoiding frequent topology jitter. The Hearst exponent is calculated based on the load time series using the rescaled range method and is used to characterize the long-range correlation of load fluctuations.

[0089] Each module communicates through the internal kernel bus. The geometry kernel layer outputs basic geometry calculation capabilities upwards, the intrinsic cognitive service layer calls the geometry kernel capabilities to achieve self-management and external services, and the application interface layer encapsulates system capabilities to provide standard calls upwards.

[0090] The generation process of the endogenous adaptation rules for the self-referential and other-referential dual-modal weights adopts a self-referential and other-referential dual-modal collaborative operation mechanism: the self-referential mode is responsible for steady-state verification, red line constraints and survival control, while the other-referential mode is responsible for new rule exploration, environmental adaptation and pattern extraction. The weights of the two are endogenously and dynamically adjusted according to the intensity of the system's operational identity, and together with the self-referential and other-referential ratio at the E-dimensional energy level, they constitute a dual-modal full-dimensional endogenous regulation system.

[0091] The bimodal weights satisfy the global normalization constraint: Under normal operating conditions, the dual-mode operates synchronously in each iteration cycle without mutual exclusion pause logic; only when the Level 3 safety circuit breaker is triggered, the rule update function of the other-pointing mode is temporarily paused to prioritize the steady state of the system.

[0092] The formula for endogenous adjustment of weights is: ,correspond ;in The sigmoid activation function is used. The steepness coefficient is determined by the system's endogenous time scale; in this embodiment, it is set to 10. The minimum operating weight for the dual-mode system is set to 0.05 to ensure that the dual-mode system maintains its basic operational capabilities throughout its entire lifecycle. The high steady-state threshold is set to 0.9, which corresponds to a state where the system is sufficiently steady.

[0093] The operating logic is as follows: When the operational identity strength is higher than the high steady-state threshold, the weight of the self-referenced mode approaches 0.95, and the system maintains steady state primarily through the self-referenced mode, while the other-referenced modes operate synchronously in a low-weight exploratory state, iterating the rules slightly without disrupting steady state; when the operational identity strength is lower than the low steady-state threshold... He pointed out that the modal weights are automatically increased, enhancing the ability to explore rules and restore order, and accelerating the system's return to a steady state.

[0094] All candidate rules generated by the self-referential modality must undergo four levels of pre-compliance verification before entering the final update stage to ensure that the rule iteration does not break the red line of operational identity and that the system's sustainability is non-decreasing: (1) Rigid red line verification: Verify that the cosine similarity between the segments corresponding to the core steady-state constraints in the candidate rules and the base rules with full lifecycle read-only locking is ≥0.999, and directly eliminate invalid candidate rules that break the core architecture; (2) Endogenous generation of adoption threshold: Based on the self-referential modality weight of the current cycle, dynamically generate the rule adoption threshold: The higher the stability of the system, the higher the rule adoption threshold, to avoid invalid disturbances under steady state; (3) Survival prospective verification: Based on the "rule change - survival change" mapping model of the system endogenous fitting, the prediction operation consistency index of the candidate rule after the implementation of the N cycles is calculated by multi-step prospective prediction. The candidate rule is retained only when the minimum value in the prediction cycle is not lower than the current value and higher than the adoption threshold; (4) Optimal candidate screening: Multi-objective non-dominated sorting is performed on the candidate rules that pass the verification. The optimization objective priority is: maximize the prediction survival index → ​​minimize the rule change amplitude → minimize the prediction self-consistency deviation rate, and screen out the unique legal optimal candidate rule.

[0095] The final rule update uses an incremental smoothing approach, with varying update intensity. The weights of the following factors and the steady-state adaptive constraints of the system are strictly limited to the range of 0 to 0.2 to avoid rule jumps. The formula is as follows: During engineering implementation, the input of the exponent term is subjected to range limiting processing within the range of [-20, 0] to avoid the risk of numerical overflow under extreme working conditions.

[0096] The final rule update formula is: ,in For the current rule tensor, The selected optimal candidate rule attractor is used. Smooth incremental updates ensure the continuity and stability of system evolution, satisfying the non-decreasing constraint of operational identity throughout the process.

[0097] Third layer: Application interface layer

[0098] The application interface layer provides standardized system call APIs, encapsulating eight types of cognitive reasoning operators into a unified interface. Upper-layer AI applications do not need to concern themselves with the underlying geometric implementation, and can call the system's native cognitive capabilities, achieving complete decoupling between the application and the kernel.

[0099] As a preferred implementation, this operating system adopts an endogenous geodesic priority scheduling algorithm to replace the static preset scheduling rules of traditional OS, thereby achieving self-optimization of the scheduling strategy.

[0100] The scheduling architecture employs a five-layer asynchronous scheduling model, divided according to real-time requirements from highest to lowest. This five-layer cycle division is based on statistical calibration of the real-time requirements of over 100 types of cognitive tasks, covering more than 95% of AI application scenarios.

[0101] Hardware synchronization layer: period 10μs-1ms, carries hard real-time control tasks, is embedded in FPGA, has the highest priority, and can preempt all lower-level resources;

[0102] System control layer: with a period of 1ms-10ms, it is responsible for system time control and kernel management tasks, and runs on a dedicated hardware timing unit;

[0103] Real-time perception layer: with a cycle of 10ms-100ms, it carries perception processing and real-time inference tasks and is deployed on NPU+FPGA heterogeneous units;

[0104] Maintenance management layer: cycle time from seconds to minutes, responsible for tasks such as rule accumulation, memory management, and health monitoring, running on the CPU;

[0105] Long-term optimization layer: with a cycle time of seconds to 24 hours, it carries tasks related to scheduling strategy iteration and version evolution. It runs offline on the CPU and does not affect the real-time link.

[0106] Task prioritization employs a second-order calculation mechanism combining a "basic formula + endogenous correction":

[0107] Basic priority calculation formula: in The current free energy of the task represents the urgency of the task. WP represents the geodesic distance from the task state to the global steady state, indicating the difficulty of task convergence. This is a normalized value for the remaining deadline of the task, based on the total task duration. 0.9 corresponds to 90% of the total time remaining. The initial baseline weight is determined endogenously with the system genus. , , .

[0108] Endogenous correction mechanism: The C-dimensional module of the system learns the load patterns, value density and operation modes of tasks through inductive reasoning, generates scheduling correction coefficients, and dynamically fine-tunes the basic priority; the longer the system runs, the more the scheduling strategy fits the actual load scenario, and the resource utilization and task execution efficiency improve autonomously.

[0109] In this embodiment, the global steady-state fixed point is entirely generated endogenously by the system and does not require manual setting. The endogenous generation and update rules are as follows: Cold start phase: Based on the distribution of modular points in the initial task set, the DBSCAN density clustering algorithm is used, with the neighborhood radius set to 0.1 times the feature scale and the minimum number of samples set to 5. The center of the highest density cluster is identified as the initial steady-state fixed point, which fits the natural center of the semantic distribution; Running phase: Updated once every 1000 cycles, the mean of the long-term stable distribution of the system within the cycle is taken as the new steady-state point, and the offset of a single update does not exceed 1% of the current feature scale to ensure smooth evolution; After hierarchical transition: Based on the orthogonal dimensionality increase result of the low-genus steady-state point, the high-genus steady-state point is obtained through short-time evolution convergence to achieve smooth hierarchical transition.

[0110] The complete scheduling execution flow is as follows: task creation → subspace resource allocation → FPGA parallel computation of free energy and geodesic distance → basic priority calculation → C-dimensional intrinsic correction → priority queue generation → hardware resource allocation according to priority → task execution → context saving → task switching / termination → runtime data feedback to the rule base. The task context includes fields such as subspace genus, state coordinates, computing power quota, and runtime slice. Context switching is completed by the hardware scheduling unit, with a switching overhead of less than 1μs. High-priority tasks can preempt hardware resources from low-priority tasks. Tasks of the same priority are sorted by geodesic distance, with those closer being scheduled first. Tasks interact through a lock-free circular message queue to avoid latency jitter caused by lock overhead.

[0111] As an optional implementation, this operating system adopts an endogenous predictive cognitive memory management mechanism based on WP volume weighting to replace the passive reactive memory paging of traditional OS, and adapts to the topological characteristics of cognitive data.

[0112] Memory pool initialization: The global module space is discretized into standard grid units. Each grid unit corresponds to a cognitive memory space of fixed WP volume, forming a global memory pool. One unit of WP volume corresponds to one standard discrete grid unit, and one grid unit corresponds to 256 bytes of physical memory.

[0113] Endogenous predictive allocation: The C-dimensional module of the system autonomously predicts memory access patterns through abductive reasoning. The engineering implementation of the abductive reasoning is as follows: using historical memory access sequences as observation results and task load patterns as candidate causes, the posterior probability of each load pattern under a given access sequence is calculated through a Bayesian network, and the load pattern corresponding to the highest posterior probability is selected as the prediction basis, and continuous grid blocks are proactively reserved in advance; when a task requests memory, the corresponding number of continuous grid blocks are allocated according to the dimension and volume of the cognitive data to be stored, and the allocation algorithm prioritizes low-fragmentation areas.

[0114] Isolation Access: The memory blocks of each task subspace are naturally isolated by geodetic boundaries, and cross-task access must be authorized through system calls. Each subspace's memory block corresponds to an independent hardware memory protection unit region register configuration. When a task attempts to access an unauthorized region, the hardware MPU directly intercepts it at the bus level and returns an access exception, without the need for software verification, achieving nanosecond-level real-time blocking.

[0115] Active Reclaiming and Scrambling: After a task releases memory, grid blocks are reclaimed. The system autonomously assesses the fragmentation rate and periodically performs topology defragmentation, merging free blocks through grid rearrangement, maintaining memory utilization above 85%. The fragmentation rate is defined as the ratio of the total size of non-contiguous free blocks to the total free space size. Free blocks with 3 or more consecutive grid points are considered fragmented blocks. This threshold is based on the average request size of cognitive tasks. The trigger threshold is when the fragmentation rate of free blocks exceeds 20%. The execution logic is: mark consecutive free blocks → migrate valid grid data → update page table pointers and MPU configuration → merge free blocks.

[0116] The formula for mapping grid points to physical memory addresses is: in For the physical base address of the cognitive memory pool, For the global numbering of grid points, The number of physical bytes corresponding to a single grid point. This is the offset within the page. Paging table entries contain fields such as grid block start number, block size, subspace ID, access permissions, physical address pointer, and MPU locale index.

[0117] The memory management system also includes built-in memory over-limit protection: when a task's memory usage exceeds the quota, a boundary shrinking mechanism is triggered to prioritize the reclamation of non-core cognitive data, ensuring the operation of core tasks and preventing system crashes caused by memory overflow.

[0118] As an optional implementation, this operating system has a built-in unified hardware abstraction layer that shields the differences between heterogeneous hardware at the bottom and provides a unified geometric operator execution interface at the top, thereby realizing the intrinsic intelligent scheduling of hardware resources.

[0119] Operator-hardware mapping rules: Hard real-time low-latency operators (state propagation, potential field mapping) are embedded in the FPGA pipeline with a single-step latency of <1μs; matrix-intensive operators (flow field reconstruction, holographic compression) are scheduled to NPU tensor units to maximize parallel computing power; global scheduling and management operators run on the CPU general-purpose kernel to take into account flexibility.

[0120] Endogenous dynamic load balancing: The system learns hardware load patterns autonomously and monitors the load rate of each hardware unit in real time. When the load of a certain type of hardware exceeds 80%, it automatically migrates non-critical operators to idle hardware units to ensure the overall throughput of the system.

[0121] Priority preemption mechanism: High-priority hard real-time tasks can directly preempt the hardware resources of low-priority tasks, with context switching overhead of <1μs, ensuring zero jitter for hard real-time tasks.

[0122] In multi-task parallel cognitive computing scenarios, through hardware abstraction layer scheduling and intrinsic optimization, the overall hardware resource utilization rate of the system can reach more than 85%, which is 35 percentage points higher than the traditional AI scheduling scheme, and it continues to autonomously optimize as the system runs.

[0123] As an optional implementation, this operating system encapsulates eight core cognitive computing categories into standardized system call APIs, allowing upper-layer applications to invoke native cognitive capabilities without needing to understand the underlying geometric theory. These interfaces can be called independently or in combination, and the system can adaptively select the appropriate type based on the application scenario, not limited to all listed types.

[0124] All eight types of cognitive reasoning operators are implemented based on a unified geometric paradigm of "geodesic operation + free energy gradient", sharing the same Weil-Petersson metric calculation kernel and global module space carrier, and belong to different functional branches under the same inventive concept:

[0125] Deductive reasoning call: Input initial state and rules, output deterministic deduction results;

[0126] Inductive reasoning invocation: Input observation samples, output universal cognitive rules;

[0127] Abductive reasoning call: Input result state, output optimal cause explanation;

[0128] Analogical reasoning call: Input two domain states, output cross-domain migration attributes;

[0129] Causal reasoning invocation: Input two events, output the causal relationship determination and strength;

[0130] Probabilistic inference call: Input prior distribution and observations, output posterior confidence;

[0131] Large number normalization call: Input a large number of micro-behaviors, output macro-emergent rules;

[0132] Three-Life Cognition Invocation: Input three types of heterogeneous subsystems and output higher-order emergent structures.

[0133] All interfaces adopt a unified input / output format, ensuring compatibility with different hardware configurations. Applications can be developed once and run on all platforms, achieving complete decoupling between the application and the underlying hardware and kernel version. The system autonomously analyzes interface call patterns and optimizes the hardware mapping efficiency of high-frequency operators. Testing shows that the latency error for eight types of operator calls is controlled within 10%, and the hardware mapping logic remains consistent.

[0134] As an optional implementation, this operating system constructs a three-tiered, intrinsically self-diagnostic security protection system to replace the external antivirus and preset rule protection of traditional OSes, ensuring system security at the geometric kernel level. The three-tiered mechanism can operate independently or work together to form a defense-in-depth system.

[0135] Process-level verification and cognitive memory management's boundary isolation mechanism work together: boundary isolation provides the first hardware-level protection barrier, preventing illegal memory access by abnormal tasks; process-level verification provides the second software-level protection barrier, monitoring the stability of the task's own state. The two barriers are independent of each other, forming a defense-in-depth structure.

[0136] Process-level protection and self-healing: Each task subspace independently performs operation identity verification, which is performed once every 10 running cycles. Tasks that fail the verification are given priority to self-heal through parameter fine-tuning. If they cannot be recovered, they are isolated and terminated on the spot, and the abnormal state is not spread. The operation identity refers to the attribute that the core topology and operating rules of the system remain essentially unchanged, and is judged by the dual standards of topological invariance and state stability.

[0137] Global-level protection and self-healing: Global topology verification is performed every 100 cycles to calculate the global effective functional genus estimate and core topology invariants. When the effective functional genus estimate fluctuates by more than 5%, self-healing operations such as topology calibration and task migration are triggered. In extreme cases, full system calibration is initiated. The first-order homology group rank is used as the topology benchmark verification item. When a jump occurs, full system calibration is triggered.

[0138] Endogenous immune self-diagnosis: This feature uses a self-referential operator to perform a secondary representation of the system's operational state, autonomously identifying unknown abnormal patterns and potential risks, generating and executing corresponding self-healing strategies. It can defend against unknown types of cognitive attacks and abnormal disturbances without requiring updates to external virus databases or preset rules. The self-referential operator is implemented using a dual-space mapping engineering method: firstly, the system's operational state vector is mapped through a pre-trained self-referential mapping matrix. Projecting onto the self-referential evaluation space yields a quadratic representation vector. The 32-dimensional original state vector contains 16 core state categories, including global free energy, effective genus, hardware load rate, task throughput, rule base size, and anomaly rate. Each category contains two statistics: mean and fluctuation, totaling 32 dimensions. The self-referential mapping matrix is ​​pre-trained offline based on a massive number of normal / abnormal operating samples before the system leaves the factory. After deployment, the basic parameters are fixed, and the classification threshold is only fine-tuned through rule accumulation; the matrix itself is not updated online. Next, calculation... The Mahalanobis distance between the anomaly and the historical normal state representation set is used to determine an unknown anomaly when the distance exceeds 3 times the standard deviation. Finally, a three-level self-healing strategy of parameter fine-tuning, task isolation, and topology calibration is generated based on the anomaly localization results. The strategy generation adopts a hybrid mechanism of rule base case matching and gradient optimization.

[0139] The operating system state graph construction rules are as follows: nodes correspond to kernel entities such as system processes, memory pages, and system calls, and edges correspond to relationships such as inter-process communication, resource dependencies, and call relationships; the first-order homology group rank is implemented in an engineering manner using a graph theory independent circuit counting approximation algorithm, corresponding to the number of independent functional closed loops in the system.

[0140] The system health assessment employs a two-tiered calculation mechanism: a basic formula plus self-reinforcing supplementation.

[0141] Basic sustainability: , , These are the weighting coefficients;

[0142] Fundamental development capabilities: ;

[0143] Basic health status: ;

[0144] Self-indicating augmentation diagnostics: The self-indicating operator performs a secondary evaluation of the entire system state, identifies hidden anomalies that cannot be covered by the basic formula, and supplements and corrects the health results.

[0145] Self-healing trigger rules: When the system health is below 0.7, the lowest priority non-core tasks are automatically terminated, topology resources are reclaimed, and global topology calibration is triggered; when the health is below 0.5, the system is rolled back to the most recent security snapshot.

[0146] This security system is completely built-in, requiring no external security software or patch injection. The system's average fault-free runtime is more than 2.5 times longer than that of traditional OSes, and the recognition rate of unknown anomalies is increased to 94%.

[0147] As an optional implementation, this operating system supports system-level intrinsic directional evolution upgrades, replacing the external flashing and passive updates of traditional OSes. System-level intrinsic directional evolution corresponds to a leapfrog qualitative improvement in the system's cognitive level, which is fundamentally different from the quantitative load adaptation of bidirectional elastic topology control. The triggering conditions, magnitude of change, and execution frequency of the two are all different.

[0148] Phase transition triggering mechanism: A triple judgment mechanism is adopted, and the system upgrade is triggered when all of the following conditions are met: (1) The global free energy sliding integral exceeds 5 times the steady-state value. Based on the statistics of 50 sets of system evolution samples, the phase transition probability exceeds 90% when the cumulative free energy reaches 4.8 times the steady-state value. 5 times is taken as the conservative judgment threshold; (2) The net increment of effective functional genus is >0, and the topological complexity meets the standard; (3) The joint closed-loop Lipschitz constant is stably less than 1 for 100 consecutive periods, and the system convergence meets the standard. The system's C-dimensional self-exponential operator simultaneously conducts capability bottleneck assessment, identifies the capability dimensions that need to be upgraded, and achieves precise targeted upgrades rather than generalized genus improvements.

[0149] Seamless Upgrade Process: The upgrade is performed in the background long-term optimization layer, employing a double-buffer hot-swap mechanism to achieve atomic switching: In the preparatory phase, a high-genus modular space replica is built in the idle buffer → the replica's steady-state is verified → the global modular space pointer is atomically switched during system idle time slots → old space resources are reclaimed. The upgrade process does not interrupt foreground business operations, and all applications seamlessly gain higher-order cognitive capabilities after the phase transition is complete.

[0150] Version compatibility mechanism: After the upgrade, the system retains the low-genus compatibility mode, old version applications can run normally, and new applications can call high-level capabilities, achieving full backward compatibility.

[0151] The system supports bidirectional reversible evolution: when the load is consistently low, it can reverse the topology and downgrade to reclaim redundant resources and reduce energy consumption; when the load increases, it can upgrade and expand, satisfying operational consistency constraints throughout the process. The system's intrinsic evolution corresponds to a step-by-step improvement in the system's cognitive level, which is a low-frequency, irreversible, or major version upgrade.

[0152] This mechanism enables the endogenous iteration of the operating system. The system's capabilities grow autonomously and directionally with runtime and task accumulation, without the need for manual intervention or external version injection. It is a direct manifestation of the core characteristics of the self-generating system at the operating system level.

[0153] As an optional implementation, this operating system possesses topology-level dynamic resource adaptation capabilities. The system autonomously predicts global load trends and environmental non-stationarity. When the system is in a low-load steady state for a long period, it reclaims hardware resources by pruning redundant subspaces and reducing global genus. When the load surges and the free energy continuously exceeds the steady-state upper threshold, it activates higher-order topology resources by expanding subspaces and increasing task coupling strength, thereby achieving topology-level dynamic resource adaptation.

[0154] Two-way elastic control corresponds to the load adaptation of the system's operating state. It is a high-frequency, reversible adjustment of resource allocation that does not change the core cognitive level of the system. Unlike the step-like qualitative change of the system's endogenous evolution, elastic control is a quantitative adaptation of parameters and topology scale. The two have fundamental differences in triggering conditions, change amplitude, and execution frequency.

[0155] The trigger boundary adopts a dual-response mode: Normal load fluctuations only adjust computing power weights and scheduling priorities, which are parameter-level adjustments and do not change the topology; topology-level upgrades or downgrades are only triggered when the environmental time-series long-range correlation index, the Hurst exponent H, > 0.7, and the free energy is higher than the steady-state upper threshold for N consecutive cycles, thus avoiding frequent topology jitter. The steady-state upper threshold is set at 1.5 times the steady-state free energy benchmark, with consecutive trigger cycles... This parameter is based on load fluctuation statistics calibration to avoid falsely triggering topology adjustments.

[0156] As an alternative implementation, multiple intelligent agents equipped with this operating system can construct a joint module space through boundary coupling, forming a higher-order self-generated cognitive entity and achieving distributed endogenous collaboration.

[0157] The system calculates the joint Chern number and effective functional genus dual topological invariants of the joint modulus space in real time. Essential emergence is determined when both indices meet the emergence threshold, and endogenously generates cooperative scheduling rules and a global steady-state attractor. It supports communication-free implicit cooperation based on a shared physical field: each agent collects disturbance signals from the common physical field of the environment (such as vibration and sound fields) through its own sensing unit, encodes the frequency components of the vibration field as boundary excitation phases and the amplitude as excitation intensities, and maps them to boundary geodesic flow driving terms after Fourier transform, with a mapping coefficient of 0.1, achieving endogenous sensing without communication signals. All nodes share the same physical field reference and spontaneously converge to the joint steady-state attractor through geodesic flow evolution, without explicit data packet interaction. The cooperative process does not require central node control, and the joint system also follows the EMC ternary self-generated closed loop, possessing endogenous steady-state and evolutionary capabilities.

[0158] Consistency Guarantee Mechanism: The global steady-state attractor of the joint modulus space serves as the consensus benchmark. Each node calculates its own state increment locally and broadcasts it. Global consistency is determined by a geodesic distance threshold, and iterative corrections are made when deviations exceed the threshold. Fault Tolerance Mechanism: In the event of a single node failure, the corresponding subspace is automatically removed from the joint modulus space, the global steady state is recalculated, and the remaining nodes automatically redistribute tasks without the need for central node coordination.

[0159] Compared with the prior art, the present invention has the following paradigm-level technological advancements:

[0160] (1) Paradigm-level breakthrough: From externally organized tools to self-generated cognitive entities. For the first time, the operating system itself is constructed as a complete EMC ternary self-generated cognitive entity, realizing closed-loop operation of kernel-level self-cognition, self-optimization, and self-evolution, breaking through the capability ceiling of traditional operating systems with manually preset rules; with a hyperbolic modulus space with geodesic boundaries as the unified computing base and a topology-metric two-layer separation architecture as the core support, all system functions are natively implemented based on geometric dynamics, matching the core characteristics of self-generated system cognition-existence identity and closed operation. In continuous operation scenarios, the average fault-free runtime of the system is 2.5 times longer than that of traditional AI operating systems, the cognitive drift rate is reduced by 90%, and the control capability grows autonomously with operation.

[0161] (2) Topology-level multi-task isolation and intrinsic optimization of resource efficiency. Physical isolation of tasks is achieved through sub-module space boundaries, which significantly reduces the cognitive crosstalk problem of multi-tasks; the intrinsic geodesic priority scheduling and hardware abstraction layer work together to improve the hardware resource utilization rate to more than 85% compared with the traditional static priority scheduling scheme in the scenario of 8-task parallel cognitive computing, and the performance degradation of multi-task parallelism is controlled within 10%, with an overall computing efficiency improvement of 40%, which continues to improve with the system's autonomous optimization.

[0162] (3) Intrinsic security and directed evolution reduce external dependence. The three-level intrinsic self-diagnostic security system can autonomously identify unknown anomalies without external patches, significantly improving system robustness; directed topological phase transition enables the intrinsic version upgrade of the operating system, and its capabilities grow autonomously and in a directed manner as it runs, without the need for manual flashing or version maintenance. It is an intelligent operating system with complete self-generating attributes.

[0163] (4) Full-scene adaptation and unified support for full-stack AI applications. The geometrical normalization of external input is achieved through the other-pointing mapping operator. All external tasks and environmental disturbances only indirectly affect the system evolution by changing the boundary conditions of the sub-module space, without breaking the operational closure of the kernel EMC ternary closed loop. While supporting dynamic access of multiple tasks, it provides a prerequisite guarantee for the long-term steady-state maintenance of operational uniformity. At the same time, it forms a standardized task access interface and has engineering scalability. Attached Figure Description

[0164] Figure 1 This is a schematic diagram of a three-layer architecture of a self-generated cognitive underlying operating system provided in an embodiment of the present invention, which shows the hierarchical structure and core module composition of the geometric kernel layer, the endogenous cognitive service layer and the application interface layer.

[0165] Figure 2 This is a schematic diagram of multi-task sub-module space isolation and scheduling provided in an embodiment of the present invention, illustrating the distribution, boundary isolation, and scheduling relationship of multiple independent task sub-spaces within the global module space;

[0166] Figure 3 This is a schematic diagram of geodesic priority scheduling dynamics provided in an embodiment of the present invention, illustrating the correspondence between multi-task priority queues, gradient evolution directions, and steady-state attractors;

[0167] Figure 4 A schematic diagram of the endogenous evolution phase transition process of a system provided in an embodiment of the present invention illustrates the complete endogenous upgrade process from a low-genus version to a high-genus version;

[0168] Figure 5 This is a schematic diagram of the heterogeneous hardware abstraction layer scheduling provided in one embodiment of the present invention, illustrating the mapping relationship and scheduling logic of various operators to FPGA, NPU and CPU. Detailed Implementation

[0169] The present invention will be further described in detail below with reference to specific embodiments. The following embodiments are all implemented under the premise of the technical solution of the present invention, and detailed implementation methods and specific calculation processes are given. However, the protection scope of the present invention is not limited to the following embodiments.

[0170] Benchmark testing platforms: The high-end server platform is configured with an FPGA Xilinx VU13P + NPU Ascend 910 + CPU Xeon Gold 6258R; the edge platform is configured with an FPGA Artix-7 + ARM Cortex-M7. All comparison schemes run on the same hardware platform, with the same workload and the same total resource quota, differing only in the core algorithm, thus eliminating interference from other variables.

[0171] FPGA resource terminology explanation: Look-Up Table (LUT) is a basic logic unit of FPGA, and Block Random Access Memory (BRAM) is a built-in memory unit of FPGA.

[0172] Example 1: Industrial-grade deployment and cold start implementation of operating system

[0173] This embodiment fully illustrates the entire process of industrial-grade deployment, compilation, burning, and cold boot of a self-generated cognitive underlying operating system, covering two typical application scenarios: server-level heterogeneous platforms and edge embedded platforms. It verifies that the system can complete the entire boot process from power-on to steady-state operation on standard hardware.

[0174] Industrial-grade hardware platform configuration

[0175] (1) Server-level heterogeneous platform

[0176] Main CPU: Intel Xeon Gold 6258R, 28 cores and 56 threads, 2.7GHz, responsible for global scheduling, rule iteration, and system management;

[0177] FPGA accelerator card: Xilinx Alveo U280 (VU13P chip), responsible for embedding the geometry calculation kernel, geodesic distance pipeline, and free energy gradient acceleration kernel;

[0178] NPU Tensor Unit: Ascend 910, 1.8GHz, responsible for high-dimensional matrix operations and large-scale state evolution parallel computing;

[0179] Memory configuration: 64GB DDR4 main memory, 16GB FPGA onboard memory, and 32MB NPU on-chip cache;

[0180] Interconnect bus: The CPU and FPGA are interconnected via PCIe 4.0 x16 with a bandwidth of 32GB / s; the CPU and NPU are interconnected via PCIe 4.0 x16.

[0181] Hardware auxiliary unit: integrates memory protection unit (MPU), hardware timer, DMA controller, and supports nanosecond-level interrupt response.

[0182] (2) Edge embedded platform

[0183] Main control MCU: ARM Cortex-M7, 216MHz, with 256KB SRAM and 1MB Flash, responsible for system management and peripheral interaction;

[0184] FPGA logic unit: Xilinx Artix-7 XC7A35T, 33280 LUTs, 1.8Mb BRAM, with a solidified simplified geometry calculation core;

[0185] Peripheral interfaces: UART, SPI, GPIO, ADC sampling interface, connecting the sensing unit and the execution unit;

[0186] Hardware auxiliary units: built-in MPU storage protection unit, watchdog timer, and low-power management module.

[0187] System compilation and deployment process

[0188] (1) Server version compilation

[0189] Geometry kernel RTL compilation: Using the Vivado 2023.1 toolchain, the three core operators—the Weil-Petersson metric calculation kernel, the Riemann Langevin geodetic flow engine, and the identity verification kernel—are compiled into FPGA bitstream files with a target clock frequency of 200MHz. The compiled bitstream file is approximately 128MB in size, occupying approximately 126,000 LUTs and 128 BRAM blocks, with a resource utilization rate of approximately 18%, and ample space is reserved for future expansion.

[0190] System kernel compilation: Using the GCC 12.2 cross-compilation toolchain, the operating system kernel written in Rust is compiled into an executable image, which includes all modules such as geometry kernel driver, process management, memory management, rule base, and system call interface; the compiled kernel image size is about 16MB, and the runtime memory usage is about 128MB.

[0191] Deployment and programming: The FPGA bitstream is dynamically loaded to the accelerator card via PCIe; the system kernel image is deployed in the CPU's main memory and is booted by the bootloader; it supports hot reload and online upgrades, and can update the geometric operator logic without downtime.

[0192] (2) Edge version compilation

[0193] Simplified kernel synthesis: The simplified geometry kernel is synthesized, placed, and routed using the Vivado toolchain, with a target clock speed of 100MHz, requiring approximately 1800 LUTs and 4 BRAM blocks, resulting in a resource utilization rate of approximately 5%.

[0194] Firmware compilation: The kernel firmware is compiled using ARM-GCC 10.3 to generate a .hex burning file with a total size of approximately 128KB;

[0195] Programming and Deployment: The FPGA bitstream and MCU firmware are programmed into the corresponding Flash memory chips via the SWD interface, and automatically loaded and run after power-on.

[0196] Cold Start Full Timing and Initialization Process

[0197] After the system is powered on, the cold start is completed in the following sequence, which fully corresponds to all operations in step S1 of claim 1. The total startup time is about 1.2s for the server version and about 80ms for the edge version.

[0198] Phase 1: Hardware Reset and Basic Initialization (Time taken: 200ms for server version, 10ms for edge version)

[0199] When the system is powered on, the hardware reset signal is released, and the CPU / MCU starts executing the Bootloader from the reset vector address.

[0200] Complete clock configuration, memory initialization, bus enumeration, and peripheral self-test; the server side synchronously completes PCIe link training and FPGA card enumeration.

[0201] Load the operating system kernel image into memory and jump to the kernel entry function.

[0202] Phase 2: Initialization of Topology Meta-parameter Layer (Time taken: 50ms for server version, 5ms for edge version)

[0203] The CPU kernel reads the default configuration parameters: the server version defaults to global equivalent genus g=2 and geodesic boundary number n=4; the edge version defaults to g=2 and n=2.

[0204] Construct a topology meta-parameter data structure in memory to store core parameters such as baseline genus, number of boundaries, version number, and security baseline threshold;

[0205] Calculate the dimension of the metric space: Server version 32-dimensional alignment; edge version The 16-dimensional alignment is used; the dimensions are aligned upwards to powers of 2, which adapts to the SIMD vector operation instruction set and memory bit width requirements of CPU and NPU, which can significantly improve parallel computing efficiency, while reserving dimensional margin to support subsequent topology upgrades.

[0206] Configure operational consistency safety baseline: set genus fluctuation threshold of 5%, state deviation threshold of 0.2 times the feature scale, and Lipschitz convergence threshold of 1.0, and write to the safety configuration register.

[0207] Phase 3: Initialization of the metric representation layer and geometry kernel (Time taken: 600ms for server version, 40ms for edge version)

[0208] Metric space construction: Discrete grid storage space for the d-dimensional Poincaré sphere model is allocated in memory. The server version allocates 10,000 standard grid points, corresponding to 2.5MB of physical memory; the edge version allocates 1,024 grid points, corresponding to 256KB of physical memory. The geodesic distance interpolation calculation error is less than 0.1% at this grid density, which meets the accuracy requirements of industrial-grade cognitive tasks and is a calibration result that represents a trade-off between computational accuracy and storage overhead.

[0209] Grid point coordinate initialization: Grid point coordinates are initialized using a uniform sampling method, and the magnitude of all grid point coordinates is controlled within 0.9. This setting reserves a boundary buffer area to avoid numerical overflow and computational singularities when the state evolution approaches the boundary of the Poincaré sphere, thus ensuring the numerical stability of long-term operation.

[0210] WP Metric Tensor Pre-calculation: The FPGA loads the bit stream and completes the register configuration of the geodesic distance pipeline, free energy calculation unit, and identity verification unit; the Weil-Petersson metric tensor lookup table values ​​in the neighborhood of discrete grid points are pre-calculated and hardened to the FPGA on-chip BRAM. During real-time calculation, the table is directly looked up and interpolated, which can reduce the amount of floating-point operations by more than 40% and significantly improve the efficiency of geometric calculation.

[0211] The Riemann Langevin engine configuration includes: writing dynamic parameters such as damping coefficient α=0.2, Hessian dissipation coefficient β=0.15, and coupling coefficient γ_c=0.1; configuring an integration step size Δt=0.01, and using the fourth-order Runge-Kutta method for numerical integration; this step size is calibrated based on numerical stability tests, with a cumulative energy error of less than 0.1% per 1000 cycles, balancing computational accuracy and operational efficiency; the fourth-order Runge-Kutta method has higher accuracy than the first-order Euler method and is adapted to the nonlinear characteristics of geodesic flow evolution.

[0212] Phase 4: EMC self-generated closed loop and functional module initialization (Time taken: 300ms for server version, 20ms for edge version)

[0213] E-dimensional energy management initialization: Set the initial self-pointing to other-pointing energy ratio ρ_SE=1.0, the energy efficiency optimization iteration step size to 0.01, and the sliding window length to 100 cycles;

[0214] M-dimensional parallel scheduling initialization: Initialize the time slice configuration of the five-layer scheduling architecture, build a lock-free circular message queue, initialize the priority queue; configure the hardware abstraction layer operator-hardware mapping table;

[0215] C-dimensional rule self-generation initialization: Load the basic rule library. The server version initially contains 32 basic rules, including 12 scheduling rules, 10 security rules, and 10 memory management rules. These are all manually preset seed rules, covering all basic system operation control logic, and serving as the initial baseline for endogenous iteration. The edge version initially contains 16 simplified rules. Initialize DBSCAN clustering parameters and rule conflict resolution thresholds.

[0216] Auxiliary module initialization: The cognitive memory management module initializes the free grid list, page table, and MPU region configuration; the system evolution control module initializes the phase change monitoring window and threshold; and the distributed collaboration module initializes the boundary interface parameters.

[0217] Phase 5: Steady-state verification and startup completion

[0218] Trigger the first global identity check and calculate the initial feature scale. Compared with the initial free energy reference;

[0219] Verify the ready flags of all modules to confirm there are no hardware errors or initialization exceptions;

[0220] Set the system operation flag, enter a steady-state idle state, and wait for task access; cold start complete.

[0221] Startup success criteria

[0222] The operating system is considered to have successfully started and entered a runnable state if all of the following conditions are met:

[0223] All hardware modules are ready with the ready flag set to 1, and there are no bus errors or memory check errors.

[0224] The topological meta-parameters have passed validation, and the coordinates of all grid points in the metric space are valid.

[0225] The FPGA geometric kernel loop closure test passed, and the geodetic distance calculation error was less than 1e-6.

[0226] The initial free energy is in the reference range of [0.20, 0.24], and the operational identity strength I_exist > 0.9;

[0227] The scheduler operates normally in idle state, with no deadlocks or abnormal interruptions.

[0228] Minimal feature set verification

[0229] After startup, execute the minimum functional verification test cases to confirm that the system's core capabilities are available:

[0230] Create one test task, allocate a g=1 genus quota, and verify that the subspace mapping is successful;

[0231] The task free energy and geodesic distance were calculated, and the deviation between the FPGA pipeline calculation results and the software simulation results was verified to be less than 0.1%.

[0232] Perform 1000 cycles of task evolution to verify that the state converges to the steady-state attractor and the convergence trajectory conforms to geodesic gradient flow dynamics.

[0233] Trigger exception injection to verify the effectiveness of MPU interception and process-level self-healing mechanisms;

[0234] The task was destroyed, and the successful resource reclamation was verified, with no memory leaks.

[0235] Steady-state operating baseline parameters

[0236] The core baseline parameters of the system under no-load steady-state conditions are as follows, serving as the benchmark for all subsequent functional tests: Global equivalent genus: g=2; Estimated effective functional genus: 2.0±0.02; Steady-state free energy baseline: 0.22; System characteristic scale: 0.35; Operational identity strength: 0.94, calculated by substituting the effective genus fluctuation and steady-state geodesic distance under no-load conditions into the operational identity strength formula; Joint closed-loop Lipschitz constant: 0.21; No-load hardware resource utilization: CPU<5%, FPGA<10%, NPU<2%.

[0237] Example 2: Performance Comparison and Verification of Multi-Task Geodesy Scheduling

[0238] This embodiment verifies the multi-task performance of the geodesic priority scheduling algorithm and makes a horizontal comparison with the static priority scheduling and time-slice round-robin scheduling of traditional OS to verify the scheduling efficiency and isolation advantages.

[0239] Test settings

[0240] Hardware platform: High-end server heterogeneous platform;

[0241] System configuration: Server-grade standard configuration, with a default initial equivalent genus g=2, corresponding to a mid-level cognitive ability, which can cover the parallel needs of 8 types of tasks;

[0242] Test task: Run 8 types of cognitive tasks simultaneously, including 2 hard real-time control tasks, 3 real-time perception tasks, 2 reasoning tasks, and 1 background maintenance task;

[0243] Comparison of schemes: Scheme A is traditional static priority scheduling, Scheme B is time-slice round-robin scheduling, and Scheme C is the geodesic priority scheduling of this invention;

[0244] Evaluation metrics: system throughput, average task latency, inter-task crosstalk rate, and hardware resource utilization.

[0245] Complete process and step-by-step calculation of single task scheduling

[0246] Taking a hard real-time obstacle avoidance task as an example, the complete scheduling and execution steps and numerical calculations are as follows:

[0247] Task creation: The upper-layer application initiates a task call, and the CPU kernel assigns a unique process ID to the task;

[0248] Subspace Mapping: The geometric normalization access and subspace resource partitioning of the task are completed through the other-pointing mapping operator. First, the task characteristics are extracted: the functional complexity of this hard real-time obstacle avoidance task corresponds to a normalized value of 1.0, with 2 equivalent interaction interfaces and a high priority; after operator mapping, the corresponding initial genus g=1, 2 geodesic boundaries, and high-priority initial state coordinates are configured synchronously with corresponding computing power weights, and independent subspace regions are divided in the global module space; the entire mapping process only modifies the subspace boundary parameters, does not affect the kernel operation rules, and strictly meets the operation closure requirements.

[0249] Priority calculation: Free energy calculation: Let the geodesic distance WP between the predicted state and the current state be denoted as . The relative deviation between local and global genus Substituting into the free energy formula, we get Geodesic distance calculation: Under a special case of the 2D Poincaré sphere model, namely the disk approximation, the current state coordinates of the task. Global steady state point Substituting into the simplified two-dimensional form of the d-dimensional general geodesic distance formula, we get Basic weight calculation: Global equivalent genus g=2, resulting in , , Normalized value of remaining deadline for the task Based on the total task duration, the remaining time corresponds to 90% of the total time; basic priority: Endogenous Correction: The endogenous correction coefficient is calculated by the C-dimensional module based on the task's real-time level and value density. This task is a hard real-time, high-value type, so the correction coefficient is increased by 3.3%, i.e., 1.033; Final Priority .

[0250] Resource allocation: The hardware abstract scheduling module allocates the hard real-time operators of the task to the FPGA pipeline and the perception operators to the NPU unit according to the priority. The scheduling management logic runs on the CPU.

[0251] Task execution: The geodesic gradient flow engine drives the subspace state evolution, and the state stability is checked every 10 cycles;

[0252] Task completion: After the task is completed, subspace topology resources and hardware quotas are reclaimed, and effective steady-state rules are stored in the rule base. See Table 1 for scheduling test results.

[0253] Table 1: Comparison of Multi-Task Scheduling Test Results

[0254] System throughput (tasks / second) 128 156 217 39.1% Average task latency (ms) 12.7 8.3 4.2 49.4% Inter-task crosstalk rate 18.3% 22.7% 1.2% 93.4% Hardware resource utilization 48% 57% 86% 50.9%

[0255] Note: All performance improvements are calculated relative to the more efficient time-slice round-robin scheduling scheme, for example, throughput improvement. The extent of delay reduction .

[0256] Test results show that the geodesic scheduling of the present invention achieves significant improvements in four dimensions: throughput, latency, isolation, and resource utilization. The sub-module space boundary isolation effectively solves the crosstalk problem of cognitive tasks, verifying the core advantages of the operating system scheduling architecture.

[0257] Example 3: Verification of Cognitive Memory Management Mechanism

[0258] This embodiment verifies the performance of cognitive memory dynamic management, and compares it with the traditional linear memory paging mechanism to verify memory utilization, isolation and defragmentation effects.

[0259] Test settings

[0260] The total capacity of the global memory pool is 10,000 grid units, corresponding to 2.5MB of physical memory, matching the memory scale of the test scenario; free blocks with 3 or fewer consecutive grid units are judged as fragmented blocks, and this threshold is based on the average request size of cognitive tasks.

[0261] The simulation involved 100 tasks randomly allocating and releasing memory, running 10,000 times consecutively. The memory allocation size followed a Poisson distribution (mean 5 units), the release time followed an exponential distribution (mean 100 periods), and the access locality conformed to the 80 / 20 rule. These distribution parameters were statistically fitted based on memory access logs from 100,000 real AI tasks, and are consistent with the memory access characteristics of cognitive tasks.

[0262] Comparison: Traditional partner system memory management, and the WP volume-weighted memory management of this invention;

[0263] Evaluation metrics: memory utilization, fragmentation rate, cross-task access interception rate;

[0264] Unit definition: 1 unit of WP volume corresponds to 1 standard discrete grid point, and a single grid point corresponds to 256 bytes of physical memory.

[0265] Step-by-step calculation of memory allocation and reclamation

[0266] Taking a single task requesting 4 units of cognitive memory as an example, the execution steps are as follows:

[0267] Receive memory request, parse the required volume as 4 units, corresponding to 4 consecutive grid points, totaling 1KB of physical memory;

[0268] Traverse the list of free grid points, find a free block with 4 consecutive grid points, and prioritize low-fragmentation areas;

[0269] Mark the corresponding grid point as occupied, update the page table entry, fill in the subspace ID and access permissions, and synchronously configure the corresponding MPU region register;

[0270] Returns the grid block start number and physical address pointer, completing the allocation. Upon task release, the corresponding grid is marked as idle, added to the idle list, and the corresponding MPU configuration is cleared. When the fragmentation rate exceeds 20%, defragmentation is triggered, migrating valid grids, merging consecutive idle blocks, and updating the global page table and MPU configuration. See Table 2 for detailed test results.

[0271] Table 2: Cognitive Memory Test Results

[0272] Average memory utilization 62% 87% 40.3% Final fragmentation rate 21% 4.2% 80.0% Illegal cross-task access interception rate 92% of the data is verified by software. Hardware MPUs inherently intercept 100% of attacks. Complete block

[0273] Numerical calculation explanation: During the test period, an average of 8700 units were used, and the average utilization rate was... The test concluded with a total of 1280 units of free space, of which 53.76 units were non-contiguous free space, resulting in a fragmentation rate of [missing information]. .

[0274] The tests verified the advantages of cognitive memory management: topology-level isolation combined with hardware MPU achieves 100% illegal access blocking, weighted paging and defragmentation significantly improve memory utilization, and adapt to the topological characteristics of cognitive data.

[0275] Example 4: Verification of Level 3 Intrinsic Security Protection

[0276] This embodiment verifies the protective effect of the three-level intrinsic security system by simulating abnormal task injection and cognitive attacks to verify the protective and self-healing capabilities.

[0277] Test settings

[0278] Ten types of anomalous tasks are injected, with typical categories including parameter out-of-bounds type (state coordinates exceeding the boundary by 10%), topology destruction type (forcibly modifying genus parameters), and cognitive interference type (injecting random state noise); five types of known cognitive attacks are simulated.

[0279] Threshold calibration instructions: The 5% fluctuation threshold is set based on three times the standard deviation of steady-state operation statistics; two self-healing cycles and a 5% adjustment step size are the optimal parameters for grid search, balancing self-healing speed and system stability;

[0280] Comparison of solutions: traditional OS firewall + antivirus software solution, and the three-level intrinsic security solution of this invention;

[0281] Evaluation metrics: anomaly interception rate, system crash rate, and anomaly recovery time.

[0282] Complete process and numerical calculation of anomaly interception

[0283] Taking a topology-destruction anomaly as an example, the handling process and numerical values ​​are as follows:

[0284] The hardware MPU first intercepts the illegal cross-space memory access attempts of the task to prevent the anomaly from spreading, and the interception time is less than 1μs;

[0285] The process-level check examines the task subspace every 10 cycles. The initial equivalent genus estimate for this task subspace is 1.0, and after anomaly injection, the estimate is 1.08, with fluctuation range... If the value exceeds the 5% threshold, the detection time is approximately 0.3ms.

[0286] First, a level one self-healing mechanism is triggered, and the fine-tuning target is the subspace dissipation coefficient. With regularization coefficient The step size was adjusted to 5% of the initial value, once per cycle; after two cycles, it took about 1ms, but the task inequality fluctuation still did not fall back to the threshold.

[0287] The process-level isolation termination is triggered, the hardware memory management unit cuts off the memory access permissions of the subspace, and reclaims the corresponding resources. The processing time is about 0.5ms.

[0288] Global-level verification and synchronization confirmed that the global topology was unaffected, and the system continued to operate normally. The average total recovery time for a single anomaly was approximately 1.8ms, composed of: hardware MPU interception <1μs, process-level detection 0.3ms, self-healing attempt 1ms, and isolation handling 0.5ms. See Table 3 for detailed test results.

[0289] Table 3: Safety Protection Test Results

[0290] Known anomaly interception rate 98% 100% Complete interception Unknown attack interception rate 35% 94% 168.6% System crash rate 12% 0% Zero crashes Mean recovery time for abnormalities 1200ms 1.8ms 99.85%

[0291] Note: All improvements are calculated relative to traditional security solutions.

[0292] Test results show that the intrinsic security system, through the dual barriers of hardware isolation and software verification, can identify unknown anomalies, prevent crashes through topology-level isolation, and achieve millisecond-level recovery through a self-healing mechanism. Its protection capabilities far exceed those of traditional external security solutions, and it does not require continuous updates to the virus database.

[0293] Example 5: Complete Verification of Intrinsic Evolution and Upgrade of the System

[0294] This embodiment fully demonstrates the entire process of an operating system's intrinsic upgrade from g=2 to g=3, verifying the effects of seamless upgrade and leapfrog capability enhancement.

[0295] Test settings

[0296] Initial system: global genus g=2, number of boundaries n=4, running continuous learning tasks, continuously inputting new scene data;

[0297] Monitoring indicators: global free energy accumulation, effective functional capacity, and task processing accuracy;

[0298] Upgrade process: The upgrade is executed in the background while the baseline task continues to run in the foreground, monitoring the business interruption time.

[0299] Test task: Target recognition in complex scenes, test set of 5000 samples;

[0300] Graph construction explanation: Nodes correspond to core functional entities, and edges correspond to the links between functional entities. Before the phase transition, the system contains 12 core functional entities and 13 functional links. After the phase transition, 4 new functional entities and 5 links are added, which is consistent with the topological change of the genus step.

[0301] Full upgrade process example

[0302] Phase transition trigger monitoring: A time window of 1000 cycles is set, with a single step size of 1 ms; the steady-state free energy of 50 consecutive samples in the standard cognitive task set with g=2 genus is 0.218, with 0.22 taken as the baseline value, and a 95% confidence interval of [0.20, 0.24]; the phase transition threshold is 5 times the baseline value, i.e., 1.1. When the system runs to the 12800th cycle, the free energy of the sliding window is approximately integrated by discretization and summation: If the threshold of 1.1 is exceeded, phase transition verification is triggered.

[0303] Topology verification: The first-order homology group rank of the global state graph is calculated using the continuous homology algorithm. The state graph before the phase transition contains 12 nodes, 13 edges, and 1 connected component. The joint closed-loop Lipschitz constant was calculated using the finite perturbation method. This value is the average measured value from the benchmark platform. Under a 1% power perturbation, the average relative change in parallelism is 0.6%, corresponding to... A 1% state disturbance corresponds to a steady-state deviation of 0.5%, therefore... A 1% feedback disturbance corresponds to a 0.7% change in the rules, resulting in... Joint value The measurement error is less than 5%. After the phase transition state ends, the state diagram has 16 nodes and 18 edges, with a rank... The effective functional inequality net increment Δg=1>0, and the joint closed-loop Lipschitz constant is consistently less than 1 for 100 consecutive cycles. Both criteria are met, confirming that it can be upgraded.

[0304] Atomic switching: During the system idle time slot in cycle 13200, a dual-buffer hot-swap mechanism is adopted: high-genus replica construction verification takes 120ms (executed offline in the background), idle time slot pointer switching takes 0.7ms (business-aware interruption time), and old space reclamation takes 2ms (executed in the background); the foreground business is unaware of it and there is no interruption.

[0305] Capability verification: After the upgrade, the accuracy of target recognition in complex scenes increased from 82% to 94%, the cognitive capacity increased by 65%, all existing applications ran normally, and the compatibility was 100%; the improvement comes from the stronger abstract representation capability brought by the higher-order topology.

[0306] Comparative conclusion: Traditional OS version upgrades require system restarts, with an average downtime of over 30 minutes, and require manual patch installation; the endogenous upgrade process of this invention is seamless and requires no manual intervention, with capabilities growing autonomously as it runs, thus revolutionizing the traditional operating system upgrade model.

[0307] Example 6: Edge Operating System Deployment Verification

[0308] This embodiment verifies the feasibility of implementing the operating system on edge embedded devices and tests its performance and functional integrity in low-resource environments.

[0309] Test platform

[0310] Hardware: ARM Cortex-M7 216MHz + small FPGA Artix-7, 256KB memory, integrated hardware MPU unit;

[0311] System configuration: Edge-end configuration equivalent genus Boundary number Substitute into the dimension formula A 4-dimensional computation buffer is retained to accommodate temporary computation variables and boundary expansion. The buffer dimension can be further clarified to temporarily store intermediate computation variables, avoid boundary overflow during numerical computation, and ensure numerical stability. A 16-dimensional Poincaré sphere model is adopted as a low-dimensional approximation to balance accuracy and resource consumption. Non-core modules are trimmed, while core functions such as kernel scheduling, security verification, and basic inference are retained.

[0312] Test task: Parallel execution of single-agent obstacle avoidance and path planning.

[0313] Resource usage and test results

[0314] Memory usage breakdown: Geometric kernel data structure 32KB, scheduling management module 48KB, basic rule base 32KB, system stack and buffer 16KB, totaling 128KB. The memory size of each module is the actual usage value after compilation on the ARM platform; the remaining 50% of memory is for application use.

[0315] FPGA resource allocation: 800 LUTs for geodesic distance pipeline, 600 LUTs for free energy calculation, and 400 LUTs for identity verification, totaling 1800 LUTs; data storage occupies 4 BRAMs to adapt to embedded edge resource constraints.

[0316] Performance: Stable parallel operation of dual tasks with an average scheduling latency of 0.42ms and no task crosstalk;

[0317] Long-term stability test: After 72 hours of continuous operation of obstacle avoidance and path planning dual tasks, with random parameter perturbation injected once per hour, the steady-state retention rate was 100%, no cognitive drift or system crash occurred, and the failure self-healing success rate was 100%;

[0318] Functional completeness: Supports the accumulation of endogenous rules and lightweight version evolution, meeting the operating system requirements of edge-generated intelligent agents.

[0319] Example 7: Verification of a Multi-Agent Distributed Cooperative Operating System

[0320] This embodiment verifies the distributed collaborative capability of multiple systems. Three intelligent agents equipped with this operating system construct a joint system through boundary coupling to achieve endogenous collaborative scheduling.

[0321] Test settings

[0322] Three identical intelligent agents, each with an initial genus g=2 and a boundary number n=2, collaborate to perform a transport task. Each subsystem provides two interaction boundaries, for a total of six boundaries across the three systems. There are three pairs of coupled boundaries, i.e., ;

[0323] Comparison of schemes: centralized scheduling scheme with central nodes, and the distributed endogenous collaborative scheme of this invention;

[0324] Evaluation metrics: Task completion time, collaboration fluency, and decentralization robustness;

[0325] Collaboration fluency is an objective indicator, calculated by weighting three dimensions: number of task pauses, path smoothness, and collaboration waiting time, with a maximum score of 10.

[0326] Coupling strength weight explanation: The weights are determined based on the analytic hierarchy process. Mutual information has the highest contribution to emergence, so it has the largest weight. The consistency test coefficient CR < 0.1.

[0327] Collaboration and Fault Tolerance Processes and Algorithms

[0328] The three systems construct an initial joint module space through interactive boundary matching;

[0329] Coupling strength calculation: Mutual information ratio 0.8 (weight 0.4), boundary adhesion 0.7 (weight 0.3), closed-loop convergence 0.75 (weight 0.3), weighted average yields... The measured average value was taken as 0.74. This threshold was calibrated based on the ROC curves of 20 multi-agent collaborative experiments, which included 12 positive samples (real emergence) and 8 negative samples (pseudo-emergence). The threshold of 0.74 was selected corresponding to the maximum point of Youden's index, and the overall accuracy reached 94%.

[0330] Joint Chern Calculus: Eigen Chern Numbers of Individual Subsystems The total number of the three subsystems is 9; the number of coupling boundaries is... Total number of boundaries Substituting into the formula, we get The joint Chern number increment is 0.75, close to the 0.8 threshold. Combined with the strong evidence of an effective genus increment of 1.0, it is comprehensively judged as an essential topological emergence, an endogenous generation of cooperative rules; the effective functional genus of the joint system increases from 6.0 to 7.0, a net increment of 1.0.

[0331] Tasks of each agent are scheduled based on a joint steady-state attractor, with no central node controlling the entire process;

[0332] Single-node failure test: When one agent is manually interrupted, the corresponding subspace is automatically removed from the joint module space. The remaining two agents recalculate the global steady state and automatically redistribute tasks. The recovery time is 0.8 seconds, and the tasks continue to execute.

[0333] Supplementary implicit cooperative testing without communication: Three agents were placed in a shared vibration field environment without any wired / wireless data interaction. They only collected common field disturbances through their own vibration sensors and encoded them as mode space boundary excitation terms. Test results showed that the cooperative task completion time was 13.4 seconds, only 8.9% longer than the explicit communication scheme, with a cooperative accuracy of 91.2%, verifying the engineering feasibility of implicit cooperative testing without communication. Detailed test results are shown in Table 4.

[0334] Table 4: Multi-task collaborative test results

[0335] Average task completion time 20.5s 12.3s 40.5% Collaborative fluency score 6.3 / 10 9.1 / 10 44.4% Recovery time after single node failure 5.2s 0.8s 84.6%

[0336] Note: All increases are calculated relative to the central node scheduling scheme.

[0337] The tests verified the distributed collaboration capability, which is more efficient in endogenous collaboration, and there is no bottleneck of central node. The failure of a single node does not affect the global operation, and the robustness is significantly improved.

[0338] Example 8: Verification of the self-optimization effect of scheduling strategy

[0339] This embodiment verifies the inherent self-optimization capability of the operating system scheduling strategy. By comparing the initial preset scheduling strategy with the inherently optimized strategy after 10,000 system cycles, the self-optimization effect is verified. The scheduling baseline cycle is 1ms, and 10,000 cycles correspond to a total runtime of approximately 10 seconds, which can complete the full learning and convergence of the load mode.

[0340] Test settings

[0341] Hardware platform: High-end server heterogeneous platform;

[0342] Test scenario: Mixed load scenario, including 8 types of tasks such as hard real-time control, real-time perception, inference, and background maintenance, with the load fluctuating dynamically over time;

[0343] Comparison of schemes: Scheme A is an initial preset static weight scheduling strategy, while Scheme B is an endogenous optimization scheduling strategy after the system has run for 10,000 cycles;

[0344] Evaluation metrics: system throughput, average task latency, hardware resource utilization, and task timeout rate;

[0345] Algorithm parameter description: The load feature vector contains 5 dimensions: computation density, memory access frequency, real-time level, duration, and variance; 8 load pattern cluster centers are set to match the load features of 8 typical cognitive tasks.

[0346] Self-optimization process and details

[0347] Initial phase: The system uses a baseline weight formula for scheduling, and the C-dimensional module synchronously collects task execution data, including task load distribution, execution time, priority matching degree, and resource idle rate;

[0348] Learning Phase: The C-dimensional inductive reasoning module uses an online K-means clustering algorithm to classify task load patterns, setting 8 load pattern cluster centers to match the load characteristics of 8 typical cognitive tasks. The matching degree between the 8 cluster centers and the preset 8 task types reaches 92%, and the clustering effect meets expectations. It identifies periodic high load periods and sudden peak characteristics. Based on the classification results, the scheduling weight correction coefficient is dynamically adjusted through Bayesian optimization. The optimization target is system throughput, and the optimization parameters are the correction coefficients of the three weights, ranging from [0.8, 1.2]. The optimal solution is generated by iterating 50 times based on the Gaussian process surrogate model. The initial exploration rate of the optimization process is set to 0.3, which is gradually reduced to 0.05 with iterations to balance global optimization and local convergence efficiency. After 50 iterations, the improvement of the optimization target is less than 0.1%, indicating convergence. Balancing exploration and utilization, the load fluctuation pattern and task value density distribution are extracted to generate scheduling weight correction coefficients and priority adjustment strategies.

[0349] Optimization Phase: The system automatically loads the intrinsically generated optimization strategy, dynamically adjusting the basic priority formula and allocating resources towards high-value, high-volatility tasks; the optimized weight adjustment coefficients are 1.1, 0.9, and 1.05, respectively, and the actual weights... , , .

[0350] Iteration Phase: The optimization strategy continues to run, feedback data is continuously collected, and strategy parameters are iteratively fine-tuned to achieve continuous endogenous growth in scheduling capabilities. Test results are detailed in Table 5.

[0351] Table 5: Comparison of Scheduling Strategy Test Results

[0352] System throughput (tasks / second) 217 243 12.0% Average task latency (ms) 4.2 3.6 14.3% Hardware resource utilization 86% 92% 6.98% Task timeout rate 3.2% 1.1% 65.6%

[0353] Note: All increases are calculated relative to the initial preset strategy.

[0354] Test results show that the operating system can achieve continuous performance improvement without external intervention by autonomously optimizing the scheduling strategy through an endogenous closed loop, verifying the core advantages of the self-generating paradigm of "intelligent services for intelligent".

[0355] The bimodal rule iteration mechanism is compared and verified. This verification is used to test the optimization effect of the self-referenced and other-referenced bimodal dynamic weight adjustment mechanism on the rule iteration convergence speed, steady-state stability and out-of-bounds risk. The verification conditions are completely consistent with the test platform and load scenario mentioned above in this embodiment. Only the operation mechanism of C-dimensional rule generation is changed and other variable interference is excluded.

[0356] Comparison scheme: The baseline version adopts a fixed self-reference / other-reference weight ratio of 1:1, without dynamic adjustment logic; the preferred version adopts the aforementioned self-reference-other-reference dual-modal endogenous weight adjustment mechanism, with supporting pre-compliance verification and incremental smooth update rules.

[0357] Operating parameters: Total operating cycle 10000 cycles, dual-mode steepness coefficient Minimum running weight High steady-state threshold Low steady-state threshold Rule look-ahead verification period ;

[0358] Evaluation metrics: rule iteration convergence period (number of periods to reach the steady-state rule set for the first time), steady-state operation identity fluctuation amplitude, rule iteration out-of-bounds number of times (number of times operation identity falls below 0.7), and throughput gain of the final scheduling strategy.

[0359] The first step in verifying the execution process is as follows: Both systems load the same initial seed rule base of 32 rules and default parameter configurations, and simultaneously start the same set of 8 mixed load tasks; the second step is: the baseline version runs rule iterations with fixed weights, and the optimized version runs according to the formula. Dynamically adjust bimodal weights, rule update intensity according to Adaptive constraints; Step 3: Sample the global operation identity strength, rule base update volume, and task throughput every 100 cycles, and record the number of convergence cycles that first reach the steady-state rule set; Step 4: Statistically count the number of times the operation identity falls below the 0.7 threshold due to rule updates throughout the process, as well as the relative throughput gain of the scheduling strategy at the end of the run.

[0360] The results show that after 720 convergence cycles of rule iteration, in the weak steady-state phase of the system, the weight of other-pointed modes is automatically increased, accelerating rule iteration and order restoration, shortening the convergence cycle by 40%. In the sufficient steady-state phase of the system, the intensity of other-pointed updates is automatically reduced, and the self-pointed steady-state control is strengthened, reducing the steady-state fluctuation amplitude by two-thirds. Furthermore, no rule iterations exceeded the steady-state threshold throughout the entire process. This mechanism significantly improves the safety and stability of rule iteration without sacrificing the final optimization effect, achieving an endogenous balance between exploration capability and steady-state maintenance capability, and verifying the engineering value of the unary two-state three-body architecture at the rule evolution level.

[0361] Example 9: Validation of Intrinsic Safety Self-Diagnosis Capability

[0362] This embodiment verifies the ability of endogenous immune self-diagnosis to identify unknown anomalies and compares the protection effects of the preset rule security scheme and the endogenous self-diagnosis scheme against novel unknown anomalies. The five unknown anomalies are implicit state drift attacks and adversarial sample attacks that did not appear in the training set. They do not trigger the first two levels of preset verification and can only be identified by the self-pointing operator.

[0363] Test settings

[0364] Five new types of cognitive attacks and anomaly patterns, which were not pre-defined, were injected and are not within the scope of the initial security rule base.

[0365] Comparison of schemes: Scheme A is a traditional preset rule security scheme, while Scheme B is the inherent self-diagnostic security scheme of this invention;

[0366] Evaluation metrics: Unknown anomaly identification rate, system crash rate, and average anomaly recovery time;

[0367] Distribution test explanation: The self-pointing quantity in the normal state has a p>0.05 according to the Shapiro-Wilk test, which conforms to the multivariate normal distribution. Therefore, three times the standard deviation is used as the anomaly judgment threshold, corresponding to the 99.7% confidence interval.

[0368] Self-diagnosis process and engineering details

[0369] After the exception injection, the process-level and global-level checks failed to identify the exception (because there were no preset rules).

[0370] The third-level endogenous immune module performs a secondary representation of the system's full state through self-reference operators; self-reference mapping matrix By contrastive learning pre-training, the training data consists of normal / abnormal state sample pairs with a dimension of 32×16, that is, the 32-dimensional original state is mapped to a 16-dimensional self-referential space. The training objective is to increase the distance between the self-referential spaces of the two types of samples.

[0371] Anomaly detection: Calculate the Mahalanobis distance between the quadratic representation vector and the historical normal state representation set. In the example, the mean of the self-pointing vectors of the normal state. Covariance matrix Abnormal sample vector Calculated Mahalanobis distance If the value exceeds 3 times the standard deviation threshold, it is judged as an unknown anomaly;

[0372] Autonomously locates anomalous subspaces and generates targeted isolation and calibration strategies;

[0373] After anomaly handling is completed, the self-directing features of the anomaly state are extracted as rule antecedents, and the corresponding handling strategy is used as consequents. The initial confidence weight is 0.7, which is gradually increased after subsequent verification. These features are then added to the rule base to update the anomaly recognition knowledge base, achieving endogenous growth in immunity. Test results are detailed in Table 6.

[0374] Table 6: Safety Diagnostic Test Results

[0375] Unknown anomaly detection rate 32% 92% 187.5% System crash rate 18% 0% Zero crashes Mean recovery time for abnormalities 1500ms 2.1ms 99.86%

[0376] Note: All increases are calculated relative to traditional preset rule schemes.

[0377] Test results show that the endogenous self-diagnostic system can overcome the limitations of preset rules, autonomously identify unknown anomalies and generate handling strategies, and its immunity continuously increases with operation, making it a truly self-generating security protection system.

[0378] The above embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the core concept and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A scheduling method for an autogenous cognitive operating system based on a hyperbolic mode space with geodesic boundaries, characterized in that, The method is deployed on a heterogeneous computing hardware platform including FPGA, NPU, and CPU. The operating system itself acts as a self-generated cognitive entity, constructing a two-layer architecture that separates the topological parameter layer and the metric representation layer. It achieves endogenous scheduling of self-cognition, self-optimization, and self-evolution through a three-element recursive closed loop of the self-generated system kernel-level energy mechanism E-matter structure M-information field state C. The method includes the following steps: S1, Kernel Initialization: The CPU kernel performs global topological parameter initialization of the hyperbolic modulus space with geodesic boundaries. The topological parameter layer stores the baseline equivalent genus and the baseline geodesic boundary number. The metric representation layer uses a d-dimensional Poincaré sphere model as an approximate calculation carrier. The Weil-Petersson metric calculation kernel and the Riemann Langevin geodesic flow operator are solidified into the FPGA logic unit. The operational identity safety baseline and kernel EMC self-generated closed loop initial parameters are configured to complete the startup loading of the geometric kernel. S2, Task Process Mapping: The various AI cognitive tasks are mapped to independent sub-modulus space processes, and each process is assigned... Initial genus quotas, geodesic boundary interface resources, and hardware computing power weights establish a one-to-one correspondence between tasks and subspaces; S3, Endogenous geodesic priority scheduling: Through parallel calculation of the free energy potential of each task subspace and the global geodesic distance via FPGA hardware pipeline, combined with the scheduling strategy generated endogenously by the C-dimensional rule base, a dynamic priority queue is constructed, and heterogeneous hardware computing resources are allocated according to the geodesic gradient scheduling rules to drive the parallel evolution of multiple task subspaces and globally satisfy the operational identity constraint; During the scheduling process, E-dimensional energy control, M-dimensional parallel computation, and C-dimensional rule iteration constitute a kernel-level self-generated closed loop, continuously and autonomously optimizing the scheduling strategy; S4, Global identity self-verification: Through the CPU scheduling kernel, identity verification operations are triggered at preset cycles to verify the topological invariance and state stability of the global module space and each task subspace. For abnormal tasks that fail the verification, the hardware memory protection unit cooperates to trigger isolation mechanisms and topological calibration to prevent the spread of anomalies from affecting the global system; Verification data is synchronously fed back to the C-dimensional rule base to iteratively update the anomaly identification strategy.

2. The scheduling method for an autogenic cognitive operating system based on a hyperbolic mode space with geodesic boundaries as described in claim 1, characterized in that, In step S1, the Weil-Petersson metric calculation kernel and the Riemann Langevin geodesic flow operator are embedded into the FPGA logic unit; the geodesic distance calculation adopts a three-level fully pipelined hardware architecture, which sequentially completes coordinate reading and mapping, WP metric tensor matching, and geodesic distance integral output; the WP metric tensor lookup table values ​​of the discrete grid neighborhood are pre-calculated and embedded into the FPGA on-chip BRAM; the Riemann Langevin geodesic flow operator is an engineered deterministic geodesic gradient flow operator with intrinsic inertia and dissipative damping, and the fluctuation characteristics are simulated equivalently by initial condition perturbation and environmental noise terms.

3. The scheduling method for an autogenous cognitive operating system based on a hyperbolic mode space with geodesic boundaries as described in claim 1, characterized in that, In step S1, the dimension of the metric representation layer satisfies the engineering approximation relation. The calculation result is rounded up to the nearest power of 2 for dimension alignment, adapting to the SIMD vector operation instruction set and memory bit width requirements of the CPU and NPU; among which For global equivalent inequality, This represents the number of geodesic boundaries.

4. The scheduling method for an autogenic cognitive operating system based on a hyperbolic mode space with geodesic boundaries as described in claim 1, characterized in that, In step S2, the normalized access of external tasks and environmental inputs is completed through the other-pointing mapping operator: the other-pointing mapping operator uniformly maps the functional requirements, performance constraints and environmental disturbances of external tasks to the initial genus, geodesic boundary number and state coordinates of the corresponding submodule space, and transforms all external inputs into standard geometric quantities in the module space; the external inputs only affect the internal evolution by changing the boundary conditions of the subsystem, without directly interfering with the operation of the kernel EMC ternary closed loop, maintaining the operational closure attribute of the system, and ensuring the steady-state basis of operational uniformity.

5. The self-generated cognitive operating system scheduling method based on hyperbolic mode space with geodesic boundaries according to claim 1, characterized in that, In step S3, the endogenous geodesy priority scheduling adopts a five-layer asynchronous scheduling architecture, which is divided into hardware synchronization layer, system control layer, real-time perception layer, maintenance management layer and long-term optimization layer according to the computation real-time performance from high to low. Each layer interacts through a lock-free circular asynchronous message queue, and high-priority tasks can preempt low-level hardware resources. The basic formula for calculating task scheduling priority is: ,in For mission free energy, The geodesic distance from the mission to the global steady state. The remaining deadline time is the normalized value; the weighting coefficients are dynamically adjusted by the system through inductive reasoning and autonomous learning of task load patterns. The initial baseline value is determined endogenously based on the equivalent genus, and the adjustment rule is as follows: , , , It is the current global equivalent genus.

6. The scheduling method for an autogenic cognitive operating system based on a hyperbolic mode space with geodesic boundaries according to claim 1, characterized in that, The method also includes an endogenous predictive cognitive memory dynamic management step, whereby cognitive memory refers to topological storage resources based on discrete grid points in the modular space: the discrete grid point resources in the global modular space are used as a cognitive memory pool, and a WP volume-weighted paging mechanism is used to achieve dynamic allocation and reclamation; the system autonomously predicts memory access patterns through abductive reasoning, actively reserves continuous grid point blocks and performs topological fragmentation, and memory utilization is dynamically adjusted through the boundary-volume ratio; memory access strictly follows subspace boundary isolation, cross-task memory access requires authorization through system calls, each subspace corresponds to an independent hardware memory protection unit area configuration, and illegal access is directly intercepted by hardware to avoid cognitive data crosstalk.

7. The scheduling method for an autogenous cognitive operating system based on a hyperbolic mode space with geodesic boundaries according to claim 1, characterized in that, The method also includes an intrinsic scheduling step in the heterogeneous hardware abstraction layer: all hardware computing resources are abstracted into unified geometric operator execution units, and automatically mapped to corresponding hardware according to operator characteristics: hard real-time low-latency operators are embedded in the FPGA pipeline, matrix-intensive operators are scheduled to the NPU tensor acceleration unit, and global scheduling and management operators run on the CPU kernel; the system autonomously learns hardware load patterns, dynamically adjusts operator mapping strategies and load balancing thresholds, hardware resource allocation is dynamically adjusted according to task priority and load prediction results, and global hardware resource utilization is autonomously optimized and improved as the system runs.

8. The scheduling method for an autogenic cognitive operating system based on a hyperbolic mode space with geodesic boundaries according to claim 1, characterized in that, The method also includes a standardized system call interface step: encapsulating eight types of geometric operation operators—deductive reasoning, inductive reasoning, abductive reasoning, analogical reasoning, causal reasoning, probabilistic reasoning, large number normalization reasoning, and three-life cognitive calculation—into a standard system call API. These operators can be called independently or in combination, and the system adaptively selects them according to the application scenario, without being limited to the enumerated types. Upper-layer AI applications do not need to pay attention to the underlying geometric implementation; they can complete full-spectrum cognitive calculations by calling the interface, thus decoupling the application from the kernel. The system autonomously statistically analyzes the interface call patterns and optimizes the hardware mapping efficiency of high-frequency operators.

9. The scheduling method for an autogenic cognitive operating system based on a hyperbolic mode space with geodesic boundaries according to claim 1, characterized in that, In step S4, the global identity self-verification adopts a three-level intrinsic security defense-in-depth system. The three-level mechanism can operate independently or work together: The first level is process-level verification and self-healing, where each task subspace independently verifies its own effective functional genus estimate and state stability. Anomalies are prioritized for self-healing through parameter fine-tuning, and if recovery is not possible, the process terminates in place. The second level is global-level verification and self-healing, which verifies the effective functional genus estimate and core topological invariants of the global module space. If the effective functional genus estimate fluctuates by more than 5%, self-healing operations such as topology calibration and task migration are triggered. In extreme cases, full system calibration is initiated. The first-order homology group rank serves as the topology benchmark verification item, and a transition triggers full system calibration. The third level is endogenous immune self-diagnosis, which uses self-indicating operators to perform secondary characterization of the system's operating status, autonomously identify unknown abnormal patterns and potential risks, generate corresponding self-healing strategies and execute them, without the need for external security patches and preset rule bases.

10. The scheduling method for an autogenous cognitive operating system based on a hyperbolic mode space with geodesic boundaries according to claim 1, characterized in that, The method also includes resource recycling and rule sedimentation steps: after the task is completed, the topological resources and hardware quotas of the corresponding sub-module space are recycled, and the stable attractors and cognitive rules that have been verified by the task are sedimented into the global rule base; the system clusters and summarizes the sedimented rules, autonomously extracts general scheduling, security and memory management strategies, updates the system cognitive baseline, and realizes the endogenous growth of kernel control capabilities.

11. The scheduling method for an autogenic cognitive operating system based on a hyperbolic mode space with geodesic boundaries according to claim 1, characterized in that, The method also includes an endogenous directed evolution step: the system autonomously assesses its own capacity bottlenecks through a self-pointing operator, and combines the global free energy accumulation and topological complexity. When the phase transition triggering condition is met, a system-level topological phase transition is triggered, which directionally improves the effective functional genus and computational capacity of the global module space, thereby realizing an endogenous version upgrade of the operating system without the need for external patch injection.

12. The scheduling method for an autogenic cognitive operating system based on a hyperbolic mode space with geodesic boundaries according to claim 11, characterized in that, The system's endogenous evolution employs a triple phase transition determination and a bidirectional reversible mechanism: the determination conditions simultaneously satisfy three conditions: the cumulative integral of the global free energy within a preset sliding window exceeds 5 times the steady-state value; the net increment of the global effective functional genus is greater than 0; the joint closed-loop Lipschitz constant is stably less than 1 for 100 consecutive periods, and the closed-loop convergence meets the standard; the system supports reverse topology pruning and degradation to adapt to low-load scenarios; after the phase transition is completed, the global metric parameters and rule base are automatically updated, and all compatible applications can enjoy the corresponding level of cognitive ability without modification, achieving imperceptible endogenous adjustment.

13. The scheduling method for an autogenous cognitive operating system based on a hyperbolic mode space with geodesic boundaries according to claim 1, characterized in that, The method also includes a two-way elastic topology control step: the system autonomously predicts the global load trend and environmental non-stationarity. When the system is in a low-load steady state for a long time, it reclaims hardware resources by pruning redundant subspaces and reducing global genus. When the load surges and the free energy continues to exceed the steady-state upper threshold, it activates higher-order topology resources by expanding the subspace and increasing the task coupling strength, thereby achieving dynamic resource adaptation at the topology level and satisfying the operational identity constraint throughout the process.

14. The scheduling method for an autogenic cognitive operating system based on a hyperbolic mode space with geodesic boundaries according to claim 1, characterized in that, The method also includes a distributed multi-system self-generated collaboration step: multiple intelligent agents running this operating system construct a joint module space through boundary coupling, calculate the joint Chern number and effective functional genus dual topological invariants of the joint module space in real time, determine essential emergence when both indicators meet the emergence threshold, and endogenously generate collaborative scheduling rules and global steady-state attractors; it supports communication-free implicit collaboration based on shared physical fields, where each node extracts collaborative excitation by sensing disturbances in the common physical field, and can spontaneously converge to the joint steady state without explicit data packet interaction, without the need for central node control, and the joint system also satisfies the EMC ternary self-generated closed loop.

15. A self-generating cognitive operating system system based on hyperbolic mode space with geodesic boundaries, characterized in that, Deployed on a heterogeneous computing hardware platform including FPGA, NPU, and CPU, the system adopts a three-layer native self-generated architecture, consisting of a geometric kernel layer, an endogenous cognitive service layer, and an application interface layer from bottom to top. The operating system itself constitutes a complete EMC ternary self-generated cognitive entity. The geometric kernel layer includes a module space initialization module, a WP metric calculation module, a Riemann Langevin geodetic flow engine, and an identity verification kernel, providing a unified geometric computing foundation for the entire system and supporting a two-layer separation architecture for topological meta-parameters and metric representations. The endogenous cognitive service layer includes an E-dimensional energy management module, an M-dimensional parallel scheduling module, and a C-dimensional rule self-regulation module. The generation module, along with two other modules, forms a kernel-level self-generating closed loop: the C-dimensional module generates control rules and optimization strategies, the E-dimensional module executes dynamic allocation of energy and computing power, and the M-dimensional module drives parallel evolution of multiple tasks. Operational data is fed back to the C-dimensional module to achieve rule iteration. The application interface layer includes standardized system call APIs, supporting various cognitive reasoning, motion control, and decision-making planning AI applications. Each module communicates via an internal kernel bus. The geometric kernel layer outputs basic geometric computing capabilities, the endogenous cognitive service layer calls the geometric kernel capabilities to achieve self-control and external services, and the application interface layer encapsulates system capabilities to provide standard calls.

16. The self-generating cognitive operating system system based on hyperbolic mode space with geodesic boundaries according to claim 15, characterized in that, The endogenous cognitive service layer also includes a cognitive memory management module and a rule storage management module; The cognitive memory management module adopts the WP volume-weighted paging mechanism, which supports endogenous predictive allocation, defragmentation, and boundary isolation access control. The rule storage management module uses steady-state attractors as storage units to realize the persistent storage, retrieval, and reuse of cognitive rules, providing data support for the self-generation of C-dimensional rules.

17. The self-generating cognitive operating system system based on hyperbolic mode space with geodesic boundaries according to claim 15, characterized in that, The intrinsic cognitive service layer also includes a system evolution management and control module and a hardware abstraction scheduling module. The system evolution management and control module supports system-level intrinsic topology phase transitions, can autonomously identify capability bottlenecks and trigger targeted upgrades, and complete system version upgrades without interrupting business. The hardware abstraction scheduling module abstracts heterogeneous resources such as FPGA, NPU, and CPU into a unified geometric operator execution pool, automatically matches the optimal hardware according to operator type and load prediction, and supports dynamic scheduling of computing power and priority preemption.

18. The self-generating cognitive operating system system based on hyperbolic mode space with geodesic boundaries according to claim 15, characterized in that, The intrinsic cognitive service layer also includes a distributed collaboration module and an elastic topology control module. The distributed collaboration module supports the construction of a joint module space by multi-system boundary coupling, realizing intrinsic collaborative scheduling and consistency maintenance without a central node, and supporting communication-free implicit collaboration based on a shared physical field. The elastic topology control module autonomously predicts system load and environmental non-stationarity, performs dynamic pruning and expansion of topology-level resources, and adapts to load fluctuation scenarios.

19. A self-generating cognitive AI chip, characterized in that, The chip integrates a self-generated cognitive underlying operating system as described in any one of claims 15 to 18. The chip has a built-in geometric kernel hardware acceleration circuit, which includes a geodesic distance pipeline, a free energy gradient acceleration core, and a dedicated identity verification core. The three types of operators, namely geodesic distance calculation, free energy gradient calculation, and identity verification, are all implemented using a fully pipelined parallel architecture.

20. A self-generating intelligent agent device, characterized in that, The system includes a sensing unit, a self-generating cognitive AI chip as described in claim 19, and an execution unit. The signal output terminal of the sensing unit is connected to the sensing input interface signal of the self-generating cognitive AI chip, and is used to collect external environmental information and device operating status information and transmit them to the self-generating cognitive AI chip. The control input terminal of the execution unit is connected to the control output interface signal of the self-generating cognitive AI chip, and is used to receive and execute the control commands output by the self-generating cognitive AI chip. The self-generating cognitive AI chip, through its built-in self-generating cognitive underlying operating system, relies on a two-layer architecture with a separation of topology meta-parameter layer and metric representation layer and a kernel-level EMC three-element self-generating closed loop to achieve device steady-state autonomous maintenance, endogenous evolution of system capabilities, and parallel scheduling of multiple cognitive tasks.

21. A computer-readable storage medium having an operating system program stored thereon, characterized in that, When the operating system program is executed by the processor, it implements the self-generating cognitive operating system scheduling method based on hyperbolic mode space with geodesic boundary as described in any one of claims 1 to 14.

Images