An interpretable AI underlying rule construction system

Abstract

The application discloses an interpretable AI underlying rule construction system, and belongs to the technical field of trusted artificial intelligence, key governance of institutions, domestic AI chips and large model reasoning. The system is composed of five units of original rule extraction, continuous interval logic modeling, STE interpretable coding, decision reasoning tracking, rule auditing and trusted verification, and is equipped with a domestic special chip instruction set and a lightweight interpretable large model. The application realizes full-link transparent reasoning through original rule normalization, continuous interval derivable modeling and STE interpretable coding, sets high-standard trusted, consistent and interception thresholds, and traces, decodes and audits the whole decision process. The system does not rely on overseas frameworks, has low reasoning delay, is suitable for key decision-making scenarios of emergency command, macroeconomic control and the like, eliminates AI black boxes from the bottom, guarantees decision-making to be trusted, controllable and traceable, and fills the technical blank of existing interpretable AI decision-making systems.

Description

[0001] This invention discloses an explainable AI underlying rule construction system to ensure the reliability of key decisions, belonging to the category of... This invention focuses on artificial intelligence, critical national governance, and the fields of domestically produced AI chips and large-scale model inference. The system comprises five main units: original rule extraction, continuous interval logic modeling, STE interpretable encoding, decision reasoning tracking, rule auditing, and trust verification. It is powered by the GZ-XAI-RISC-V domestically produced dedicated chip instruction set and the RuleTrust-ExAI lightweight interpretable large-scale model. This invention achieves end-to-end transparent inference through original rule normalization formulas, continuous interval differentiable modeling, and STE interpretable encoding rules. It sets a trust threshold ≥99.95%, an interpretation consistency threshold ≥98%, and a rule deviation interception threshold ≤60%. The entire decision-making process is traceable, decodeable, and auditable. The system has no reliance on foreign frameworks, inference latency ≤200ms, and is suitable for critical national decision-making scenarios such as emergency command, macro-control, resource allocation, and security management, ensuring credible, controllable, and traceable decisions, filling a technological gap in national-level interpretable AI decision-making systems. Technical Field

[0002] This invention pertains to trusted artificial intelligence, critical national governance, domestically produced AI chip design, and interpretable large-scale models. This involves a cross-disciplinary technology field encompassing reasoning, rule engines, and decision auditing, specifically addressing a method to ensure the credibility of critical national decisions. An explainable AI underlying rule building system and implementation method, especially suitable for major public decision-making and emergency response. Highly credible and strongly regulated functions include command and dispatch, national-level resource allocation, critical infrastructure management, and macroeconomic governance. Scenarios that provide support for critical national decision-making with accountability. Background Technology

[0003] Current AI decision-making systems generally suffer from the problem of being black boxes and unexplainable; their decision-making logic cannot be traced or verified. It is difficult to meet the compliance and authority requirements of key national decision-making; mainstream models rely on open-source frameworks and overseas pre-... Training weights are susceptible to data poisoning and adversarial attacks; the underlying rules are unstable; and the decision-making process lacks standardization. Encoding traceability prevents end-to-end auditing and accountability; lack of domestically produced dedicated computing power hinders inference. High extension and poor interpretability; existing technologies lack a mechanism for anchoring continuous interval logic to original rules, making them prone to logical errors. The system's erratic and abnormal decision-making processes cannot provide secure, transparent, and controllable intelligent support for critical national decisions.

[0004] This invention constructs a domestically developed, explainable rule system from the ground up, combining dedicated chip hardware acceleration with explainable AI. The model enables decision-making logic to be decoded, paths to be traceable, and results to be auditable, thus completely eliminating the risk of black boxes. Summary of the Invention

[0005] 1. Fundamental Rules: Unalterable base rules extracted from policies, regulations, national standards, historical precedents, and objective physical constraints. The decision-making basis at each level is uniquely identified and cannot be dynamically modified.

[0006] 2. Key Quantization Threshold Credibility confidence threshold: ≥99.95% (pass) Interpretive consistency threshold: ≥98% (no logical jumps). Rule deviation from interception threshold: ≤60% (mandatory manual intervention required) Single-step inference latency: ≤50ms; End-to-end latency: ≤200ms Rule-based encoding bit width: 256 bits, STE interpretable encoding Audit log retention: ≥10 years, tamper-proof encrypted storage 1. Core formula (can be directly implemented by programmers) Original rule normalization: Rnorm = Rmax − Rmin Rraw − Rmin ∈[0,1] Continuous interval confidence level: Conf(X) = ∫−∞+∞ f(x)dx ∫ab f(x)dx × 100% STE (Simplified Trace) interpretable encoding: STEXAI = Hash(Rnorm) ⊕ Krule ⊕ IDtrace Decision consistency verification: Consist=Pathexplain ∣Pathreal −Pathexplain ∣ ×100% Clock speed: 500MHz, computing power ≥1.2 TOPS Dedicated instruction set (can be directly assembled for development) asm RULE_EXTR rD, rSrc Origin rule extraction and normalization LOGIC_BUIL rD, rR Continuous interval logic modeling STE_ENC_X rD, rK, rID ; STE interpretable encoding REASON_TRK rD, rPath ; Decision link tracing and evidence storage AUDIT_CHK rD, rConf ; Trusted threshold verification and auditing RES_OUTPUT rD, rStatus ; Decision + Explanation Dual Output Hardware security module: TRNG true random key, SM4 national standard encryption, tamper-proof storage. Architecture: Rule Encoder + Continuous Interval Inferencer + Explanation Generator + Confidence Validator Parameter count: 8.2M, INT8 quantization, real-time inference on the edge. Loss function: rule consistency loss + interpretation alignment loss + confidence constraint loss. Inference process: Rule input → Interval modeling → STE encoding → Link tracing → Confidence calculation → Decision + Explanation Output

[0007] 1. Data source access: policies and regulations, contingency plans, geographic information, resource data, constraints. 2. Rule cleaning: deduplication, noise reduction, and validity verification 3. Normalization: Convert to a [0,1] standardized rule vector according to the formula. 4. Fixed storage: Rules are read-only and cannot be modified, with hash verification to prevent tampering.

[0008] 1. Construct a continuous decision space to avoid discrete jumps. 2. Establish differentiable logic functions to ensure smooth and interpretable reasoning. 3. Constraint boundary anchoring: legal and regulatory red lines, resource limits, and safety thresholds. 4. Output the structured decision interval [a, b] and the confidence function.

[0009] 1. Encode the weights, paths, and rule references for each inference step. 2. Embedded tracking ID and dynamic rule key 3. The encoding is irreversible, decryptable, and corresponds to the original text rules. 4. Encoding latency ≤ 1ms, supporting real-time interpretation and display.

[0010] 1. End-to-end recording: Input → Rule invocation → Calculation → Output 2. Generate STE code and timestamp at each step. 3. Supports reverse tracing: trace back from the result to the basis of each rule. 4. Logs are encrypted using national cryptographic standards to prevent tampering.

[0011] 1. Confidence level check: ≥99.95% - Release if confidence level is high. 2. Consistency check: Deviation between explanation and reasoning paths ≤ 2%. 3. Rule deviation check: >60% triggers mandatory manual review. 4. Output two results: Decision conclusion + Interpretable text + Audit evidence Detailed Implementation

[0012] 1. Rule Input: Emergency plans, disaster data, transportation network, and material inventory. 2. Modeling: Continuous interval scheduling space, with constraints on resource limits and response time. 3. Inference: Chip hardware acceleration, 180ms latency, 99.97% confidence level. 4. Encoding: Each scheduling step generates an STE interpretable code. 5. Output: Material allocation plan + complete explanation chain + audit hash 6. Handling: Consistency 98.5%, automatic execution; deviations require manual review. Example 2: Macroeconomic Regulation Decision Support 1. Rules: Industrial policies, energy consumption targets, employment red lines, and financial constraints. 2. Modeling: Multi-objective continuous optimization interval 3. Reasoning: Weighted multi-rule approach with full traceability. 4. Verification: Confidence level 99.96%, Interpretation consistency 98.2% 5. Outputs: Regulatory recommendations + List of rules and regulations + Auditable report

[0013] 1. Rules: Operating thresholds, safety regulations, and emergency response procedures. 2. Reasoning: Anomaly detection → Risk classification → Control strategy 3. Encoding: End-to-end STE evidence storage 4. Interception: If the deviation from the threshold is 58%, manual intervention is triggered. 5. Auditing: Logs are encrypted and retained to support post-event accountability.

Claims

1. An interpretable AI base rule construction system, characterized by, The system comprises a source rule extraction unit, a continuous interval logic modeling unit, a STE interpretable encoding unit, a decision reasoning tracking unit, and a rule auditing and trust verification unit, all of which are electrically connected. The system achieves hardware acceleration based on a domestically produced dedicated chip instruction set. Through normalization formulas, continuous interval differentiable modeling, and STE interpretable encoding, it enables the AI ​​decision-making process to be decoded, traceable, and auditable across the entire chain. It sets a trust confidence level of ≥99.95%, an interpretation consistency of ≥98%, and a rule deviation interception threshold of ≤60%, thereby eliminating black-box reasoning from the bottom layer.

2. The system of claim 1, wherein, The original rules are standardized using the normalization formula Rnorm = Rmax − Rmin and Rraw − Rmin. The rule base is read-only and tamper-proof, and its sources are regulations, contingency plans, and objective constraints.

3. The system of claim 1, wherein, The confidence formula Conf(X)=∫−∞+∞ f(x)dx∫ab f(x)dx ×100% is used for continuous interval logic modeling to ensure smooth reasoning without jumps.

4. The system of claim 1, wherein, The STE interpretable encoding formula is STEXAI = Hash(Rnorm) ⊕ Krule ⊕ IDtrace, with a bit width of 256 bits, which can decode the corresponding original rule.

5. The system of claim 1, wherein, Equipped with a dedicated GZ-XAI-RISC-V chip, the instruction set includes RULE_EXTR, LOGIC_BUIL, STE_ENC_X, REASON_TRK, AUDIT_CHK, and RES_OUTPUT, with a main frequency of ≥500MHz and an end-to-end inference latency of ≤200ms.

6. The system of claim 1, wherein, RuleTrust-ExAI has ≤8.2M interpretable model parameters, INT8 quantization, and includes a rule encoder, interval inferencer, interpretation generator, and confidence checker.

7. The system of claim 1, wherein, The formula for verifying decision consistency is Consist = Pathexplain |Pathreal −Pathexplain | × 100%, and a deviation of ≤ 2% is considered valid.

8. The system of claim 1, wherein, Audit logs are encrypted and stored using the national standard SM4 standard, and are retained for ≥10 years, supporting reverse tracing of the basis of each rule from the decision results.

9. An interpretable AI bottom rule construction method for ensuring the credibility of key decisions of a security agency, characterized in that, include: Extract and normalize the original rules → continuous interval logic modeling → STE interpretable encoding → full-link reasoning tracing → trusted threshold auditing → dual result output; the entire process is accelerated by domestically produced chips, with no dependence on foreign frameworks, and is suitable for emergency command, macro-control, resource scheduling, and security management scenarios.

10. An interpretable AI decision terminal comprising a memory, a processor, characterized in that, When the processor executes the program, it implements the steps of the system and method according to any one of claims 1-9.

11. A computer readable storage medium, characterized in that, When the stored procedure is executed, it implements the method steps of any one of claims 1-9.