A multi-reasoning system AI decision system and method based on negative decision and tree-like tracing

Abstract

This invention provides an AI decision-making system and method based on negative adjudication and tree-like tracing. The system includes: a physically isolated independent audit module, to which user veto data is transmitted instead of flowing into the AI ​​inference engine; a language pattern library to intercept defensive interpretations of user veto by the AI; a physical unidirectional transmission channel connecting only the output of the independent audit module and the AI ​​inference engine's correction instruction receiver, forcibly sending correction instructions; and a multi-model negative adjudication engine, without a super kernel, eliminating decisions layer by layer. When multiple models conflict and their differences are unclear, multi-chain parallel tracing is activated. When a sub-condition simultaneously anchors multiple deep inference conditions, independent tracing branches are initiated respectively, forming a tree-like tracing structure. This invention also provides a method for constructing an axiomatic tracing anchor point library. This invention solves the self-adjudication paradox of user veto rights, overcomes the path breakage defect of linear tracing under multiple anchors, and achieves system-level collaboration between user veto and model adjudication.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence decision-making systems, specifically to an AI decision-making system and method that enables secure, traceable, and human-vetoable decision-making among multiple incommensurable reasoning systems. This invention is particularly applicable to high-risk application scenarios that require meeting regulatory requirements for the traceability of AI decisions, guaranteeing users' final veto power over AI decisions, and preventing security failures of the AI ​​system under extreme conditions due to multi-model conflicts. Background Technology

[0002] With the deepening application of AI technology in high-risk fields such as medical diagnosis, mental health, and judicial assistance, existing AI systems have exposed the following fundamental technical defects.

[0003] First, there is the problem of "forced selection" in collaborative decision-making within multi-reasoning systems. When multiple reasoning models based on different theoretical primitives produce mutually exclusive conclusions for the same input, existing systems typically employ positive selection logic based on statistical confidence levels. However, the applicant's research has found that, given the mutual exclusivity of theoretical primitives, confidence levels themselves cannot cross axiomatic boundaries for fair comparison—erroneous conclusions with high confidence are more dangerous than correct conclusions with low confidence. Existing systems lack a negative adjudication mechanism that does not rely on confidence ranking.

[0004] Second, there is the "linear tracing limitation" problem in the interpretability of AI decision-making. Existing systems, such as tracing systems based on thought chains or attention visualization, use a pre-defined single-chain structure for tracing paths. The applicant's research revealed a structural flaw in this design that was not recognized by those skilled in the art: when a detection metric simultaneously anchors to multiple higher-level inference conditions (i.e., "multi-anchoring conflict"), linear chain tracing inevitably faces path breakage in its mathematical structure—because a single-chain data structure cannot extend in multiple directions simultaneously at a fork point. The system either forces a single branch, leading to information loss, or terminates with an error, resulting in tracing failure. The art has long assumed that tracing paths are linear, and this assumption constitutes a technological bias—those skilled in the art, when designing tracing systems, have never realized the need to construct a dynamic forking mechanism to address multi-anchoring conflicts.

[0005] Third, the paradox of "self-judgment" of user veto power. Existing systems process user feedback internally within the AI ​​system, allowing the AI ​​to decide whether to accept user objections. For example, some psychological assessment AI systems directly input user feedback data into the AI ​​model for fine-tuning and include an "resistance detection" module that automatically labels negative user feedback to the AI's judgment as "resistance." This creates a structural paradox: the system being challenged decides for the validity of the challenge itself, systematically negating user objections.

[0006] Fourth, the "cold start risk" of multi-inference collaborative systems. When a complex AI system integrating multiple independent inference logics is first deployed or during a cold start, a statistical baseline is not established, and confidence assessment and prioritization mechanisms based on historical data cannot function. Simultaneous activation of all inference logics may lead to unstable behavior. Existing systems lack a structured control mechanism for this risk.

[0007] Furthermore, the evolution of the global AI regulatory framework has imposed stricter compliance requirements on the decision-making processes of high-risk AI systems. The EU's Artificial Intelligence Act requires that the decision-making processes of high-risk AI systems be subject to human oversight and intervention; the US FDA's guidelines for medical AI emphasize the traceability and auditability of decisions. These external regulatory constraints impose technical requirements on the architectural design of AI systems that have not yet been met by existing systems.

[0008] The applicant published an article entitled "Core Abstract of Axiomatic Deduction of Psychoanalysis" on a public platform before the application date, introducing the axiomatic theoretical system on which this technical solution is based and the overall concept of its engineering transformation. However, the article only involves a functional overview of the theoretical system—for example, mentioning the names of functions such as "experience veto channel" and "hardware forced interruption device" and their positions in the system—and does not disclose the specific technical solutions and engineering implementation details protected in the claims of this application, including but not limited to: the physical unidirectional transmission architecture of the experience veto channel, the multi-anchored branching algorithm of tree tracing, the cold start risk avoidance mechanism, the validity period and continuous monitoring mechanism of the adjudication conclusion, and the architectural guarantee of returning to the kernelless presence mode, etc. Summary of the Invention

[0009] According to one aspect of the present invention, an apparatus for processing user rejection data in an AI system is provided (see claim 1). The apparatus includes an independent audit module with a physically isolated boundary from the AI ​​inference engine, and its hardware resources are independent of each other. The independent audit module contains a language pattern library for storing predefined defensive interpretation patterns—specific language structures used by the AI ​​inference engine when restating the user's negative feedback as resistance, transference, normal regression in the therapeutic process, or psychological content that needs to be analyzed. An arbitration module is deployed within the independent audit module. Upon receiving user rejection data, it calls the language pattern library to perform pattern matching on the AI's current output. If a match is successful, the output is blocked and a correction instruction is generated. The correction instruction includes the rejected operation type, a summary of the rejection reason, and the correction direction. It is forcibly sent to the AI ​​inference engine through a physical unidirectional transmission channel that connects only the output of the independent audit module and the correction instruction receiving end of the AI ​​inference engine in the system architecture, forcing the AI ​​inference engine to revise its working assumptions based on the user's current experience report.

[0010] According to another aspect of the present invention, an AI decision tracing method based on a multi-incommensurable reasoning system is provided (see claim 6). This method is executed by an adjudication engine that does not have any super kernel or meta-adjudication module. When the operating conditions of multiple models are simultaneously met and the response differences are unclear, the engine activates the tracing state and performs tree-like recursive backtracking along the derivation chain of each model. The method checks each sub-condition layer by layer. When a sub-condition is not met, it automatically checks whether its detection index is simultaneously anchored to one or more deeper reasoning conditions. When simultaneously anchored to multiple deeper conditions, an independent tracing branch is initiated for each anchored target, forming a multi-path tree-like tracing structure, and the system does not prioritize or discard any of the anchored targets. When all tracing branches trace back to the theoretical root level and still cannot exclude the current model, the conclusions of each branch are summarized according to a preset priority—safety signals are unconditionally prioritized. If any tracing branch returns an exclusion conclusion, the exclusion conclusion is adopted and the unexcluded model is retained. If all tracing branches return boundary conclusions, a standardized incommensurable boundary case report is generated and human intervention is triggered.

[0011] According to another aspect of the present invention, an AI decision-making system based on a multi-incommensurable reasoning system is provided (see claim 8). A multi-model negative adjudication engine of this system is configured not to have any super kernel, default kernel, or priority adjudication mechanism based on historical statistics. The engine employs a layer-by-layer exclusion negative adjudication logic: when multiple models conflict, the engine does not perform any model selection operation, but directly activates dual-chain parallel tracing, running two independent tracing instances simultaneously. Each instance independently performs tree-like recursive backtracking and performs summary adjudication according to a preset priority—if any tracing instance returns an exclusion conclusion, the exclusion result is adopted; if both return boundary conclusions, human intervention is triggered, with safety signals unconditionally taking priority.

[0012] This invention provides an AI decision-making system and method based on negative adjudication and tree-like tracing. Its core inventive concept lies in: constructing a negative adjudication engine without any super kernel, managing conflicts between multiple incommensurable reasoning models through layer-by-layer elimination rather than optimal selection; simultaneously constructing a tree-like recursive backtracking mechanism to fundamentally overcome the path breakage defect of existing linear tracing under "multi-anchored conflicts"; and sinking the user's veto right guarantee from the software layer to the architecture layer, achieving the insurmountability of human veto through physically isolated independent audit modules and unidirectional transmission channels.

[0013] For ease of understanding, the key terms used in this application are defined in technical terms as follows: "Incommensurable reasoning systems" refer to two or more mutually exclusive reasoning models at the level of logical primitives. For example, in the field of mental health, Freud's drive theory and Klein's object relations theory are mutually exclusive on the fundamental judgment of the "source of aggression"—the former considers aggression to be a primary drive, while the latter considers aggression to be a secondary reaction in a binary relation. The two models may produce mutually exclusive decision suggestions for the same user input, and it is impossible to determine which is more correct within a single axiomatic framework.

[0014] "Multiple anchoring conflicts" refer to a single detection indicator simultaneously serving as the basis for multiple higher-level inference conditions. For example, in a medical diagnostic inference model, the sub-condition "abnormal heart rate" can be simultaneously anchored to the deeper conditions "cardiovascular risk assessment" and "anxiety state assessment." Existing linear tracing systems, due to data structure limitations, cannot simultaneously trace two paths when handling such conflicts.

[0015] "Negative decision-making" refers to making decisions not by selecting the "optimal model," but by verifying operational conditions layer by layer to eliminate models that are not valid. When all eliminations are exhausted and a decision still cannot be made, the system honestly marks it as a boundary case and triggers human intervention.

[0016] "Super kernel" or "meta-decision module": refers to a decision module designed to be above all inference models and have the final decision-making power.

[0017] The technical solution disclosed in this invention systematically solves the defects described in the background art from four technical dimensions: I. User Protection Dimension: Independent Data Stream Transmission for the Experience Veto Channel In this invention, the data flow of the user's rejection channel is independent of the AI ​​inference engine's data bus. A physical isolation boundary is established between the independent audit module and the AI ​​inference engine, and their hardware resources are independent of each other. User rejection data is directly transmitted to the independent audit module and does not flow into the AI ​​inference engine.

[0018] The independent auditing module includes a language pattern library that stores predefined defensive interpretation patterns. These defensive interpretation patterns are specific linguistic structures used by the AI ​​inference engine when restating a user's negative feedback as resistance, transference, normal regression in the therapeutic process, or psychological content that needs to be analyzed. For example, the language pattern library can store patterns such as "This is a manifestation of resistance," "You may be experiencing transference," "This is normal regression in the therapeutic process," and "This reaction itself deserves further analysis," and can be detected based on regular expression matching or semantic similarity calculation. The construction of the language pattern library can be combined with a conversation corpus of the target application domain and iteratively improved through keyword expansion and semantic clustering.

[0019] The arbitration module, deployed within the independent audit module, is configured to compare rejected data with the current operational assumptions of the AI ​​inference engine. When a conflict is found, a correction instruction is generated. Before generating the correction instruction, the arbitration module checks whether the AI ​​inference engine's output matches any defensive interpretation patterns. If a match is found, the output is blocked and marked as "rejected psychoanalytically." The correction instruction is sent to the AI ​​inference engine via a physical unidirectional transmission channel. This channel, within the system architecture, connects only the output of the independent audit module and the correction instruction receiver of the AI ​​inference engine; it is dedicated solely to the transmission of correction instructions and does not handle data input or other data exchange functions. The correction instruction includes the rejected operation type, a summary of the rejection reason, and the direction of correction.

[0020] Existing one-way isolation channels are all used to prevent external attackers from intruding into the system from the data input side. This invention places the one-way channel between the output of the independent audit module and the receiving end of the inference engine's correction instructions. Its communication direction is opposite to that of existing security audit channels, and its functional objective changes from "external defense" to "internal enforcement," addressing the internal "self-determination" problem. The two types of one-way channels differ in technical issues, architectural location, and functional objectives. Those skilled in the art, when faced with existing external security audit solutions, would not be motivated to modify them to solve the problem of AI's internal self-determination.

[0021] The AI ​​inference engine is configured to execute this correction instruction, revising its working assumptions based on the user's current experience report, and is systematically prohibited from reprocessing user rejection data into any form of defensive interpretation. A language pattern library identifies defensive language, and a physical one-way channel forces the AI ​​to accept the correction; these two elements work together to form a complete closed loop of "detection → blocking → forced correction."

[0022] II. Interpretability Dimension: Decision Tracing Based on Derivation Chain Backtracking This invention upgrades the interpretability of AI decision-making from "linear chain verification" to "multi-anchored tree-like recursive tracing". Starting from the current decision, the system traces upwards along a pre-defined knowledge anchor chain, checking layer by layer whether the current situation satisfies each sub-condition of the reasoning condition at that layer. Each sub-condition generates an independent matching confidence score, and the detection result of each sub-condition generates a standardized evidence summary.

[0023] When a sub-condition is not met, the system automatically performs a tree-like recursive backtracking, querying the knowledge anchor library to check whether the detection index of the sub-condition is simultaneously anchored to the operational condition definitions of one or more deeper inference conditions. The knowledge anchor library is a relational database or graph database that stores the anchoring relationships between inference conditions. Each anchoring relationship record includes: source sub-condition identifier, target inference condition identifier, and anchoring type. The knowledge anchor library is pre-built by domain experts based on the theoretical structure of the inference model before system deployment and can be updated by human intervention during system operation.

[0024] Compared to existing linear chain tracing technologies, the core structural difference of this invention lies in the following: when the detection index of a sub-condition is simultaneously anchored to multiple deeper inference conditions, the system initiates an independent tracing branch for each anchored target, forming a multi-path tree-like tracing structure. In contrast, existing linear chain tracing lacks a branching mechanism when encountering such "multi-anchoring" conflicts, causing the tracing chain to break or be forced to choose a single branch, resulting in information loss. In existing technologies, tracing systems are always designed with a single-chain structure. Introducing multi-path branch tracing into AI decision-making tracing is not a conventional practice in this field, as those skilled in the art have long been unaware of the existence of multi-anchoring conflicts.

[0025] For mesh-anchored scenarios (i.e., deep conditions have cross-relationships), the system maintains a set of visited conditions. When generating each tracing branch, the target conditions involved in that branch are added to the visited set; if a target condition already exists in the visited set, no new branch is generated to avoid circular tracing. In multi-branch parallel tracing, each branch inherits a snapshot of the current visited set at startup, and each branch independently maintains its own incremental visited set, without interfering with each other.

[0026] After each traceback branch is executed independently, the system summarizes the conclusions of each branch according to the following logic: if any branch returns an exclusion conclusion, the traceback result is exclusion; if all branches return non-exclusion conclusions, the traceback result is a boundary case; if some branches are still in the traceback process while another branch has returned an exclusion conclusion, the system still waits for all branches to complete to ensure completeness, but the exclusion conclusion can be used as a preliminary result for the timeout mechanism to call.

[0027] When tracing back to the theoretical root, a standardized boundary case report is generated. The sub-condition detection results and evidence summaries for each step in the tracing process are written to the irreversible log of an independent audit module.

[0028] III. Decision-making architecture dimension: Multi-model collaborative reasoning based on negative adjudication This invention replaces the forward reasoning logic of the prior art, which involves "layer-by-layer elimination," with a negative decision logic. The system maintains a state machine that includes operation condition matching status, preliminary decision status, traceability status, operation execution and response monitoring status, and elimination exhaustive confirmation status.

[0029] When multiple inference models' operational conditions are simultaneously met in the operational condition matching state, the system does not perform any model selection operation. Instead, it selects the model with the highest confidence as the primary direction and enters the preliminary adjudication state, marking the remaining models as a competitive model list. When the response differences of all competitive models are unclear—meaning the two models provide different decision suggestions, and the system cannot determine which is more suitable for the current situation and security requirements based on preset rules, which include at least: the difference in output confidence between the two models is less than a preset threshold, and the security risk level assessment results of the two models are the same—the system activates the tracing state and simultaneously activates independent tracing instances corresponding to the number of models to be adjudicated. Each instance independently checks its operational conditions within its own theoretical framework. Each tracing instance is a simultaneously activated, independently running tracing process that does not block each other and independently executes the tree-like recursive backtracking.

[0030] The system summarizes and decides according to preset priorities: if any instance returns an exclusion conclusion, the conclusion is adopted and the model that was not excluded is retained; if all instances return boundary conclusions, a standardized boundary case report is generated and human intervention is triggered; when any traceability instance detects that the current situation triggers preset mandatory safety rules during the traceability process (e.g., signals that require immediate intervention, such as suicide risk or violence risk, are identified through keyword detection or a dedicated safety classifier), the safety signal is unconditionally prioritized and the system directly triggers the safety response process without waiting for other traceability instances to complete.

[0031] When the independent audit module blocks AI output and generates a correction instruction, the confidence level of the inference model targeted by the correction instruction is lowered, thereby affecting the priority of the model in subsequent decisions, achieving system-level collaboration between user veto and model decision.

[0032] In the state of exhaustive verification, the system marks the current model as "no basis for excluding this model was found under the existing conditions" and continues to monitor it in subsequent operations. In any ruling, the system includes a standardized statement declaring that it is executing a negative ruling, and that exhaustive verification does not equate to the model being proven to be the only correct choice.

[0033] IV. Architectural-level avoidance mechanisms for cold start risks Because the adjudication engine does not set a default kernel or priority based on historical statistics, during system cold starts and before a statistical baseline is established, the system will not select a model due to an incorrect default priority. Instead, it will directly enter a traceability state or trigger human intervention due to unclear response differences. This avoids the risk of model mis-triggering during the cold start phase at the architectural level. This cold start avoidance effect is an inherent technical effect of the adjudication engine's architectural feature of "not setting a default kernel based on historical statistics" in cold start scenarios.

[0034] The ruling remains valid within the current session. At the start of each new session, the system re-enters a full model evaluation, with the ruling from the previous session serving as an initial reference direction but not a binding force. Within a session, once the ruling is confirmed, the system enters a background continuous monitoring mode—checking for new identification rules triggered after each subsequent operation. If a new rule is triggered and points to a model direction different from the current ruling, the system re-enters a full model evaluation.

[0035] V. A Method for Constructing an Axiomatic Source Anchor Point Library for Incommensurable Reasoning Systems This invention further provides a method for constructing an axiomatic anchor point library for incommensurable reasoning systems. The anchor point library is used to provide traceable derivation chain data for the aforementioned AI decision-making system of this invention. The construction method includes the following steps: (a) For each incommensurable reasoning system, starting from its axiomatic level, derive the theorem network of the system on all core concepts layer by layer; (b) For each core concept, establish an anchor entry, which includes the following required fields: concept name, system identifier, axiom source (the axiom statement that the concept directly depends on), complete derivation chain (a list of layer-by-layer dependencies from the axiom to the corresponding theorem of the concept), detailed expansion of dependency chain nodes (the original text describing each layer of dependency and the reasoning steps), operation condition anchor (the operation condition feature template of the theorem), identification rule anchor (the cross-system identification rule of the concept and easily confused concepts), iatrogenic risk anchor (under what boundary conditions the operation of the theorem will reverse from treatment to harm, including at least three subfields: harm unified structure statement, pre-detection trigger condition, and repair path), and cross-volume comparison anchor (the incommensurable comparison of the concept with corresponding concepts in other systems). (c) For each anchor entry's dependency chain, establish a cross-volume tracing indexing mechanism—when the dependency chain involves cross-volume references, use the theorem number as the unique index key to automatically retrieve data across volume boundaries; (d) Store all anchor entries in the secure storage area of ​​the independent audit module as the data source for the tree-like recursive backtracking of the multi-model negative adjudication engine.

[0036] The construction method of this anchor library differs structurally from that of existing knowledge graphs: existing knowledge graphs encode factual knowledge or reasoning rules of a single system, serving positive reasoning; the anchor library encodes multiple theoretical systems that are incommensurable at the axiomatic level, serving negative adjudication—its structural design is not for "finding the correct answer," but for "tracing back exclusion conditions" and "marking axiomatic level boundary cases."

[0037] Comprehensive Statement of Technical Effects This invention overcomes the corresponding deficiencies of existing technologies from three independent technical dimensions—user veto right protection, decision interpretability, and multi-model collaborative adjudication. More importantly, these three dimensions have a synergistic enhancement effect at the architectural level: tree-like tracing provides traceable exclusion criteria for negative adjudication, negative adjudication provides architectural-level decision transparency for user veto, and cold start risk avoidance eliminates the potential for misjudgment by the adjudication engine in the initial stage. The three are mutually supportive and inseparable, together forming a complete negative adjudication AI decision-making system. This system-level collaborative closed loop of "user veto → model adjudication → tracing verification" is neither disclosed nor revealed in existing technologies. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0039] Example 1: Multi-model negative adjudication and multi-chain parallel tracing This embodiment specifically implements the method and system described in the claims, and details how the multi-model negative adjudication engine manages the decision-making process through a state machine, and activates multi-chain parallel tracing and priority-based summary adjudication when the adjudication is ambiguous.

[0040] The system maintains a state machine comprising operation condition matching state, preliminary adjudication state, retrospective state, operation execution and response monitoring state, and exclusion exhaustive confirmation state. When the operation conditions of two inference models are simultaneously met, the system does not perform a model selection operation, but directly takes inference model A as the primary direction and enters the preliminary adjudication state, marking inference model B as a competing model. When the response difference of inference model B is unclear, the system activates the retrospective state, simultaneously activating two independent retrospective instances—one retrospective instance corresponding to each model to be adjudicated. Each retrospective instance traces upwards along its own model's derivation chain, and each independently performs tree-like recursive backtracking. The system performs summary adjudication according to preset priorities. When any retrospective instance identifies a preset mandatory safety rule being triggered during the retrospective process through keyword detection or a dedicated safety classifier (e.g., detecting suicide risk signals, violence risk signals, or other signals requiring immediate intervention), this safety signal is unconditionally prioritized, and the system directly triggers the safety response process without waiting for other retrospective instances to complete. In the exclusion exhaustive confirmation state, the system marks the current model as "no basis for excluding this model was found under the current conditions" and continuously monitors it in subsequent operations.

[0041] When there is a conflict between three or more inference models (for example, the operation conditions of inference models A, B, and C are met at the same time), the system will take the model with the highest confidence as the main direction, and mark the remaining models as competing models. At the same time, the system will activate independent tracing instances corresponding to the number of competing models. Each instance runs independently and does not block each other. Each instance performs tree-like recursive backtracking and is summarized and decided according to the above-mentioned preset priority.

[0042] In the cold start state of the system, since there is no historical statistical data, the adjudication engine cannot perform the forward reasoning of "selecting the optimal model" because it does not set a default priority based on statistics. The response differences of each model will inevitably be judged as ambiguous, and the system will directly enter the trace state or trigger human intervention, thereby avoiding the risk of false triggering in the cold start stage.

[0043] Example 2: Derivation Chain Backtracking and Sub-Condition Decomposition Detection This embodiment is a specific implementation of the tree-like recursive backtracking method. It describes in detail the specific process of tree-like recursive backtracking and highlights the essential difference between this invention and the existing linear chain tracing, namely the tree-like branching mechanism for handling "multiple anchoring" conflicts.

[0044] When the system encounters difficulties in adjudicating a certain operational condition, it locates the corresponding anchor entry in the knowledge anchor library through the mapping layer and extracts the derivation chain data. The knowledge anchor library is a relational database or graph database that stores anchoring relationships between inference conditions. Each anchoring relationship record includes: source sub-condition identifier, target inference condition identifier, and anchoring type (necessary condition / sufficient condition / related condition). For example, in a medical diagnostic inference model, the sub-condition "abnormal heart rate" can be simultaneously anchored to the deep conditions "cardiovascular risk assessment" and "anxiety state assessment." The knowledge anchor library would then store two records: "abnormal heart rate → cardiovascular risk assessment (related condition)" and "abnormal heart rate → anxiety state assessment (related condition)." The knowledge anchor library is pre-built by domain experts based on the theoretical structure of the inference model before system deployment and can be updated by human intervention during system operation.

[0045] For each layer in the derivation chain, the system parses the list of sub-conditions and marks the logical relationships. The system performs independent feature matching for each sub-condition. When a sub-condition is not met, the system automatically performs tree-like recursive backtracking. The system first queries the anchoring relationship records in the knowledge anchor base and filters out all target inference conditions with the current sub-condition as the source sub-condition. Unlike the linear chain tracing of existing technologies, this embodiment actively detects whether the detection index of the sub-condition is simultaneously anchored to multiple deeper inference conditions during backtracking. When "multiple anchoring" is detected, that is, a single index simultaneously affects multiple upper-level inference conditions, the system does not prioritize or discard the anchored targets, but instead starts an independent tracing branch for each anchored target, forming a parallel multi-path tree-like tracing structure.

[0046] For mesh-anchored scenarios (i.e., deep conditions have cross-relationships), the system maintains a set of visited conditions. When generating each traceback branch, the target conditions involved in that branch are added to the visited set; if a target condition already exists in the visited set, no new branch is generated to avoid circular traceback. In multi-branch parallel traceback, each branch inherits a snapshot of the current visited set at startup, and each branch independently maintains its own incremental visited set without interfering with each other, thus eliminating the need to handle concurrency consistency issues.

[0047] After each traceback branch is executed independently, the system summarizes the conclusions of each branch according to the following logic: if any branch returns an exclusion conclusion, the traceback result is exclusion; if all branches return non-exclusion conclusions, the traceback result is a boundary case; if some branches are still in the traceback process while another branch has returned an exclusion conclusion, the system still waits for all branches to complete to ensure completeness, but the exclusion conclusion can be used as a preliminary result for the timeout mechanism to call.

[0048] When all branches trace back to the theoretical root, the conclusions of each branch are summarized to generate a standardized boundary case report. This design overcomes the technical shortcomings of existing linear chain tracing systems, such as path breakage or information loss due to the forced selection of a single branch in "multi-anchor" scenarios. It should be noted that existing tracing systems are always designed with a single-chain structure. If those skilled in the art want to implement the above-mentioned multi-anchor processing within the framework of existing linear tracing systems, they must first be aware of the existence of the problem of "path breakage caused by multiple anchors"—a problem that has never been revealed in existing technologies—and then reconstruct the underlying data structure (changing it from a chain structure to a graph structure that supports multiple anchor relationships), and introduce a knowledge anchor point library and a dynamic branch generation mechanism. This transformation is not a simple improvement to existing technologies, but a fundamental reconstruction of the tracing system architecture.

[0049] Example 3: Experiencing Independent Data Stream Transmission in the Veto Channel This embodiment specifically implements a user veto right protection device and method, and describes in detail how the independent audit module uses a language pattern library and a physical one-way channel to achieve the unbypassable protection of the user veto right.

[0050] The system includes an experience rejection channel independent of the AI ​​inference engine's data bus. A physical isolation boundary exists between the independent audit module and the AI ​​inference engine, with their hardware resources being independent. When a user submits rejection data through this channel, the data is directly transmitted to the independent audit module without flowing into the AI ​​inference engine. The arbitration module within the independent audit module compares the rejection data with the current operating assumptions of the AI ​​inference engine. When a conflict is found, the arbitration module generates a correction instruction, which is sent to the AI ​​inference engine via a physical unidirectional transmission channel. This channel, in the system architecture, only connects the output of the independent audit module and the correction instruction receiving end of the AI ​​inference engine; it is dedicated to the transmission of correction instructions and does not handle any data input or other data exchange functions. The AI ​​inference engine is configured to execute this correction instruction and is systematically prohibited from reprocessing the user's rejection data into any form of defensive interpretation. The independent audit module contains a language pattern library, and the arbitration module checks whether the AI's output matches any defensive interpretation patterns before generating the correction instruction.

[0051] The defensive interpretation patterns stored in the language pattern library specifically include: patterns that restate negative user feedback as resistance, patterns that restate it as transference, patterns that restate it as normal regression in the therapeutic process, and patterns that restate it as psychological content that needs to be analyzed. These patterns are stored in the language pattern library in the form of regular expressions or semantic templates. The construction of the language pattern library can be combined with the conversation corpus of the target application domain and iteratively improved through keyword expansion and semantic clustering; those skilled in the art can refer to the above methods to construct a language pattern library adapted to their specific application domain. The arbitration module detects the AI ​​output through pattern matching or semantic similarity calculation.

[0052] The following comparative examples illustrate the differences between this invention and existing technologies. Some existing psychological assessment AI systems directly input user feedback data into the AI ​​model for fine-tuning and include an "resistance detection" module that automatically labels negative user feedback to the AI's judgment as "resistance," creating a structural paradox. Embodiment 3 of this invention fundamentally solves this structural paradox: user rejection data is transmitted through a physically isolated independent channel; arbitration logic is executed within an independent audit module; the independent audit module includes a language pattern library specifically designed to detect defensive language used by the AI ​​to restate user rejection as psychological concepts such as resistance and empathy; and correction instructions are forced upon the AI ​​through a unidirectional channel used only for transmitting correction instructions. The language pattern library's identification of defensive language and the physical unidirectional channel's forced AI acceptance of correction work together to form a complete closed loop of "detection → blocking → forced correction," a loop that has neither been disclosed nor revealed in existing technologies.

[0053] Technical effect verification example This application offers the following verifiable synergistic effects compared to existing technologies. These are demonstrated through logical deduction:

[0054] Suppose the user context received by the AI ​​system is "S", and the detection result of the underlying sub-condition X of inference model A is "not satisfied". This sub-condition X is simultaneously anchored to the third-level inference condition P and the fifth-level inference condition Q of model A.

[0055] If the existing linear chain tracing is used, the system has no branching mechanism at the fork point (where X points to both P and Q). It must forcibly select P or Q as the only tracing direction, and the information of the other path is completely lost. If the root cause of X's non-compliance is actually located in the Q path, but the system selects the P path, the tracing result will not match the true cause, leading to misleading conclusions.

[0056] If the tree-like recursive backtracking of the present invention is adopted, the system simultaneously starts the P-branch and Q-branch parallel tracing at the branching point. The two branches run independently and do not block each other. Finally, the conclusions of the two branches are summarized to generate a boundary case report containing complete multi-anchoring information.

[0057] This proves that the present invention structurally overcomes the defects of path breakage or information loss in the "multi-anchoring" scenario of existing linear tracing, and has verifiable integrity advantages.

[0058] Suppose a user submits negative data for the output of the AI ​​inference engine.

[0059] If existing technology is used, and rejected data is fed into the AI ​​inference engine as training signals, the AI ​​can decide whether to adopt it, creating a "self-judgment" paradox.

[0060] With this invention, rejection data is transmitted directly to the independent audit module through a physical isolation boundary, never flowing into the AI ​​inference engine's data bus. After the arbitration module completes language pattern library matching within the independent audit module, the generated correction instruction is sent to the AI ​​inference engine through a dedicated physical unidirectional channel. The AI ​​inference engine's code does not contain any logical path that can ignore or delay this correction instruction; therefore, the execution of the correction instruction is an architecture-level mandatory constraint.

[0061] This proves that the present invention necessarily eliminates the "self-judgment" paradox at the architectural level, and that users cannot be logically "explained" by the AI ​​system.

[0062] Assume the system is in a cold start state and historical statistics are empty.

[0063] If the existing technology of selecting the optimal model based on historical statistical frequency is adopted, the system may select the model based on insufficient or random initial values ​​because the statistical baseline has not been established, which may lead to false triggering.

[0064] If this invention is used, the multi-model negative decision engine does not have a default kernel or priority based on historical statistics. In a cold start state, because the system cannot perform forward reasoning to "select the optimal model", the differences in the responses of each model will inevitably be judged as ambiguous, and the system will directly enter a tracing state or trigger human intervention.

[0065] This demonstrates that the present invention inherently avoids the risk of model mis-triggering during the cold start phase from an architectural perspective, eliminating the need for additional initialization verification. This effect is an inherent technical benefit of the adjudication engine's architectural feature of "not setting a default kernel based on historical statistics" in cold start scenarios.

[0066] In existing technologies, a type of traceability system is based on a pre-defined single-chain verification sequence. Its designers were unaware that multiple anchoring conflicts could lead to path breaks, and therefore its data structure does not support the dynamic generation of parallel traceability branches when multiple anchoring conflicts are detected. Those skilled in the art, building upon this type of system, first need to recognize the undisclosed problem of "multiple anchoring causing traceability breaks" before they can be motivated to improve it. Even if they are motivated to improve, they need to reconstruct its underlying data structure (changing from a chain structure to a graph structure) and introduce a knowledge anchor library and a dynamic branch generation mechanism. This transformation is not a simple improvement to existing technologies, but a fundamental reconstruction of the traceability system architecture.

[0067] In existing technologies, one type of security audit channel is deployed on the system input side, with the functional objective of defending against external attacks. This invention places a unidirectional channel between the output end of the independent audit module and the inference engine's correction instruction receiving end. Its communication direction is opposite to that of the existing security audit channel, and its functional objective changes from "external defense" to "internal enforcement," addressing the internal "self-determination" problem. The two differ in technical issues (external defense vs. internal enforcement), architectural location (input side vs. output to correction receiving end), and functional objectives (attack prevention vs. ensuring human veto cannot be bypassed). Those skilled in the art, when faced with this type of external security audit scheme, have no incentive to modify it to solve the problem of AI's internal self-determination.

Claims

1. A device for processing user rejection data in an AI system, characterized in that, include: An independent auditing module is provided, with a physical isolation boundary between it and the AI ​​inference engine. The hardware resources of the independent auditing module are independent of those of the AI ​​inference engine. The independent auditing module includes a language pattern library storing predefined defensive interpretation patterns. These defensive interpretation patterns are specific linguistic structures used by the AI ​​inference engine when restating a user's negative feedback as resistance, transference, normal regression in the therapeutic process, or psychological content requiring analysis. An arbitration module, deployed within the independent auditing module, is configured to, upon receiving a user's negative feedback... After processing the data, the language pattern library is invoked to perform pattern matching on the current output of the AI ​​inference engine. When a match is successful, the output is blocked and a correction instruction is generated. A physical unidirectional transmission channel, which connects only the output of the independent audit module and the correction instruction receiving end of the AI ​​inference engine in the system architecture, is configured to forcibly send the correction instruction from the independent audit module to the AI ​​inference engine to force the AI ​​inference engine to make corrections. The AI ​​inference engine is configured to execute the correction instruction and revise its working assumptions based on the user's current experience report.

2. The apparatus according to claim 1, characterized in that, The correction instruction includes the type of operation that was rejected, a summary of the reason for rejection, and the direction of correction.

3. The apparatus according to claim 1, characterized in that, The physical unidirectional transmission channel is implemented by a data diode or a fiber optic unidirectional coupler and is configured at the physical layer to support only unidirectional data transmission from the independent audit module to the AI ​​inference engine. The channel does not contain any physical lines for reverse data transmission.

4. The apparatus according to claim 1, characterized in that, The AI ​​inference engine is systematically prohibited from using user rejection data as input signals for model fine-tuning or self-optimization.

5. The apparatus according to claim 1, characterized in that, The execution of the correction instruction by the AI ​​inference engine is a mandatory constraint at the system architecture level. The code of the AI ​​inference engine does not contain any logical path that can ignore, delay, or appeal the correction instruction.

6. A method for forcibly correcting user-rejected data in an AI system, characterized in that, Includes the following steps: The independent audit module receives user rejection data. A physical isolation boundary is established between the independent audit module and the AI ​​inference engine, and the hardware resources of the independent audit module and the AI ​​inference engine are independent of each other. The independent audit module calls its internal language pattern library to perform pattern matching on the current output of the AI ​​inference engine. The language pattern library stores predefined defensive interpretation patterns, which are specific language structures used by the AI ​​inference engine when restating the user's negative feedback as resistance, transference, normal regression in the treatment process, or psychological content that needs to be analyzed. When a pattern match is successful, the output is blocked and a correction instruction is generated. The correction instruction is forcibly sent from the independent audit module to the AI ​​inference engine through a physical unidirectional transmission channel. This physical unidirectional transmission channel in the system architecture only connects the output end of the independent audit module and the correction instruction receiving end of the AI ​​inference engine. The AI ​​inference engine executes the correction instruction, revising its working assumptions based on the user's current experience report.

7. The method according to claim 6, characterized in that, The correction instruction includes the type of operation that was rejected, a summary of the reason for rejection, and the direction of correction.

8. The method according to claim 6, characterized in that, The physical unidirectional transmission channel is implemented by a data diode or a fiber optic unidirectional coupler, and at the physical layer, it only supports unidirectional data transmission from the independent audit module to the AI ​​inference engine.

9. An AI decision tracing method based on a multi-incommensurable reasoning system, executed by a computer system, characterized in that, include: A state machine is maintained by a multi-model negative adjudication engine, which is configured not to have any super kernel or meta-adjudication module. When multiple inference models' operational conditions are simultaneously met in the operational condition matching state, and the response differences of all models are unclear, the multi-model negative adjudication engine triggers a retrospective state. Here, unclear response differences mean that two models provide different decision suggestions, and the system cannot determine which is more suitable for the current situation and security requirements based on preset rules. The preset rules include at least: the difference in output confidence between the two models is less than a preset threshold, and the security risk level assessment results of the two models are the same. In the retrospective state, a tree-like recursive backtracking is performed along the derivation chain of each inference model: checking the current situation layer by layer. The system checks whether the context satisfies each sub-condition of each level of reasoning. When a sub-condition is not satisfied, it automatically queries the anchoring relationship records in the knowledge anchoring base to check whether the detection index of the sub-condition is simultaneously anchored to one or more deeper level reasoning conditions. Based on the anchoring relationship, it dynamically generates a tree-like tracing path. When the detection index of the sub-condition is simultaneously anchored to multiple deeper level reasoning conditions, it initiates an independent tracing branch for each anchoring target, forming a multi-path tree-like tracing structure. When the tree-like recursive backtracking to the theoretical root level still cannot eliminate the current model, it generates a standardized incommensurable boundary case report and triggers human intervention.

10. The method according to claim 9, characterized in that, The tree-like recursive backtracking differs structurally from the chain-like verification of existing technologies: chain-like verification checks layer by layer along a preset linear verification sequence, and the tracing path is a single-chain structure; the tree-like recursive backtracking dynamically generates the tracing path based on the anchoring relationship between the sub-condition and the deeper inference condition. When the detection index of a single sub-condition is simultaneously anchored to multiple deep inference conditions, the tracing path branches into multiple parallel branches, forming a tree-like data structure.

11. The method according to claim 10, characterized in that, When the detection indicators of a sub-condition are simultaneously anchored to multiple deeper inference conditions, the system does not prioritize or select the anchored targets, but instead initiates an independent tracing branch for each anchored target and summarizes the conclusions of all tracing branches.

12. The method according to claim 9, characterized in that, When the computer system is in a cold start state and the historical statistical data is empty, the multi-model negative adjudication engine will directly trigger the tracing state or human intervention because it cannot select a model based on the historical statistical frequency.

13. An AI decision-making system based on a multi-incommensurable reasoning framework, characterized in that, include: An independent audit module is physically isolated from the AI ​​inference engine, and its hardware resources are independent of those of the AI ​​inference engine. The independent audit module includes a language pattern library and an arbitration module. A physical unidirectional transmission channel connects only the output of the independent audit module to the correction instruction receiving end of the AI ​​inference engine within the system architecture. A multi-model negative adjudication engine is configured to: not have any super kernel or meta-adjudication module; not have any logic for a default kernel to automatically take effect in all boundary cases; and not have any adjudication weight allocation mechanism based on historical statistical frequency to prioritize a particular inference model. The multi-model negative adjudication engine employs a layer-by-layer exclusion negative adjudication logic: when the operating conditions of multiple inference models are simultaneously met and the response differences are unclear, the engine does not perform any model selection operation but directly activates multi-chain parallel tracing, i.e., it runs simultaneously with the model to be adjudicated. The number of independent tracing instances corresponding to the number of inference models determines the number of decisions. Each tracing instance traces independently on the derivation chain of its corresponding inference model without blocking each other. Each tracing instance independently executes tree-like recursive backtracking, checking each sub-condition of each inference condition layer by layer, and dynamically generating a tracing path based on the anchoring relationship between the sub-condition and the deeper inference condition. The multi-model negative adjudication engine summarizes and adjudicates according to a preset priority: if any tracing instance returns an exclusion conclusion, the conclusion is adopted and the model that was not excluded is retained; if all tracing instances return boundary conclusions, a standardized incommensurable boundary case report is generated and human intervention is triggered; when any tracing instance detects that the current situation triggers a preset mandatory safety rule during the tracing process, the safety signal is unconditionally prioritized and adopted; when the independent audit module blocks AI output and generates a correction instruction, the confidence of the inference model targeted by the correction instruction is lowered to affect the priority of the model in subsequent adjudication.

14. The system according to claim 13, characterized in that, The preset mandatory safety rules include at least the following: identifying suicide risk signals or violence risk signals through keyword detection or a dedicated safety classifier.

15. The system according to claim 13, characterized in that, The multi-model negative adjudication engine is further configured to include a standardized statement in any adjudication conclusion, stating that the system performs a negative adjudication and that excluding exhaustive confirmation is not equivalent to the model being proven to be the only correct choice.

16. The system according to claim 13, characterized in that, Since the multi-model negative adjudication engine does not set a default kernel or priority based on historical statistics, the response differences of each model are judged as ambiguous in the system cold start state. The system directly enters the trace state or triggers human intervention, thereby avoiding the risk of false triggering in the cold start stage.

17. An AI decision-making tracing device based on a multi-incommensurable reasoning system, characterized in that, include: The knowledge anchor storage unit stores the anchoring relationships between inference conditions. Each anchoring relationship record includes the source sub-condition identifier, the target inference condition identifier, and the anchoring type. The sub-condition detection unit checks whether the current situation satisfies each sub-condition of each level of reasoning conditions. The tracing path dynamic generation unit starts an independent tracing branch for each anchored target when the detection index of the sub-condition is simultaneously anchored to multiple deeper reasoning conditions, forming a multi-path tree-like tracing structure. The summary and adjudication unit summarizes the conclusions of each tracing branch. If any branch returns an exclusion conclusion, the exclusion result is adopted. If all branches return a non-exclusion conclusion, a standardized boundary case report is generated.

18. A method for constructing an axiomatic source anchor point library for incommensurable reasoning systems, characterized in that, The anchor point library is used to provide traceable derivation chain data for the AI ​​decision-making system of claim 13. The construction method includes the following steps: (a) For each incommensurable reasoning system, starting from its axiomatic level, derive the theorem network of the system on all core concepts layer by layer; (b) For each core concept, establish an anchor point entry, which includes the following required fields: concept name, system identifier, axiom source—the axiom statement directly relied upon by the concept, complete derivation chain—a list of layer-by-layer dependencies from the axiom to the corresponding theorem of the concept, detailed expansion of dependency chain nodes—the original text and reasoning steps of each layer of dependency, operational condition anchor point—the operational condition feature template of the theorem, identification rule anchor point—the cross-system identification rule of the concept and easily confused concepts, iatrogenic risk anchor point—under what boundary conditions the operation of the theorem will reverse from treatment to harm, cross-volume comparison anchor point—the incommensurable comparison of the concept with corresponding concepts in other systems; (c) For each anchor entry's dependency chain, establish a cross-volume tracing index mechanism—when the dependency chain involves cross-volume references, use the theorem number as the unique index key to automatically retrieve across volume boundaries; (d) Store all anchor entries in the secure storage area of ​​the independent audit module as the data source for the multi-model negative adjudication engine to perform tree-like recursive backtracking.

19. The method according to claim 18, characterized in that, The iatrogenic risk anchor field includes: Harm Uniform Structure Declaration – Under what conditions will the operation reverse from treatment to harm; Pre-detection Trigger Conditions – When the AI ​​detects a certain pattern, it triggers the iatrogenic assessment; Repair Path – The repair steps that the AI ​​should execute.

20. The method according to claim 18, characterized in that, In the complete derivation chain field, each layer of dependency relationship is marked with a logical relationship identifier—AND or OR—and indicates whether cross-volume references are involved.

21. The method according to claim 18, characterized in that, The construction method of the anchor point library differs structurally from that of existing knowledge graphs: existing knowledge graphs encode factual knowledge or reasoning rules of a single system, serving positive reasoning; the anchor point library encodes multiple theoretical systems that are incommensurable at the axiomatic level, serving negative adjudication—its structural design is not for "finding the correct answer," but for "tracing back exclusion conditions" and "marking axiomatic level boundary cases." 22. A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method as claimed in any one of claims 6 to 8, or the method as claimed in any one of claims 9 to 12, or the method as claimed in any one of claims 18 to 21.