11 results about "Validator" patented technology

A system, method, and computer-readable medium enforce compliance, policy governance, and jurisdiction-aware control of digital artifacts generated by AI models. A deterministic digest of each artifact is stored in a registry, and a compliance token is generated encoding policy predicates, consent scope, jurisdictional lineage, monetization parameters, and a zero-knowledge proof component including a proof statement ID, cryptographic commitment, and verification key. A gateway verifier receives the artifact and token, verifies the signature, performs zero-knowledge proof validation without revealing regulated data, evaluates encoded policies, and renders a permit-or-deny decision. Verification results and compliance decisions are recorded via a dual-ledger atomic commit engine that updates both compliance and financial ledgers. A settlement engine calculates royalty values based on validated performance metrics and appends the results to the commit entry. Upon jurisdictional transitions, a transition engine detects boundary crossings and issues a forked compliance token with updated predicates. The system ensures verifiable, policy-aware, and jurisdictionally adaptive governance.

This invention provides a method for generating complex functions on a low-code platform based on large model code interpretation and generation. By capturing and standardizing user input, a structured intent graph containing business entities, operational relationships, parameter constraints, and control flow topology is constructed as a unified semantic anchor. Combined with a pre-defined domain rule base, a semantic alignment validator is designed to achieve multi-dimensional semantic comparison and conflict detection between code and requirements. Code snippets are generated in parallel by multiple models and verified in real time, and semantic conflicts are resolved through local regeneration. By fusing the results of multiple models to construct a joint abstract syntax tree, lightweight symbolic execution is used to simulate the correctness of cross-model code, locate and correct core logical conflicts. This invention significantly improves the consistency, accuracy, and automatic correction capabilities of low-code platform code generation.

This application provides a method and apparatus for training a natural language to SQL model based on reinforcement learning. The method includes: calling an external tool based on natural language instructions and the structural information of a target database to generate a response containing the tool call trajectory and SQL execution results; the external tool includes at least one of a search tool, a calculation tool, and an SQL execution tool; based on the response containing the tool call trajectory and SQL execution results, performing inference verification on the SQL execution results to obtain verification results, the inference verification including comparing the difference between the result table of the SQL generated by the model and the real table through an SQL validator; and adjusting the generation strategy of the large language model according to the verification results to generate executable SQL statements. An end-to-end natural language to SQL training framework is constructed through reinforcement learning algorithms, combining tool fusion inference and interleaved thinking mechanisms to dynamically optimize the model's decision-making capabilities throughout the entire process of data table retrieval, logical calculation, and SQL generation.

Disclosed is a fail-closed conformance control layer for machine-assisted decision artifacts. The system generates replay-verifiable receipts that commit to a canonical artifact digest, evidence and snapshot identifiers, assumptions, toolchain identifiers, and a policy pack identifier and policy version digest. A deterministic validator evaluates a receipt under the versioned policy pack to output PASS, FAIL, or HOLD with reason codes from a reason-code registry, including time-stamped modeled-variability codes. Downstream side effects are gated in a control path and permitted only on PASS; otherwise they are withheld or routed to remediation or human review. The system may emit portable conformance packs and revalidate receipts over time to detect drift and manage a certification lifecycle. Conformance suites with deterministic test vectors and expected validator outputs may certify validator implementations by issuing signed conformance certificates and publishing a certificate or certificate digest to a verifiable certificate registry for independent verification and procurement acceptance testing.

This invention discloses a hallucination detection method and system based on atomic statement decomposition and multi-source evidence fusion. The hallucination detection method includes: responding to a user query input by the user, acquiring the answer text generated by the retrieval enhancement generation system and multiple retrieval evidence documents; decomposing the answer text into multiple atomic statements; classifying each atomic statement to obtain the statement type corresponding to the atomic statement; using a statement validator to verify the atomic statements based on the retrieval evidence documents to obtain a verification score; determining the fusion weight according to the statement type of the atomic statement, and calculating the comprehensive score of the atomic statement based on the fusion weight and the verification score; marking the verification status of the atomic statements according to the comprehensive score, and marking atomic statements with abnormal verification status in the answer text. This invention solves the problems of excessively coarse granularity of hallucination detection and inaccurate error localization in the RAG system through a fine-grained atomic statement and verification mechanism, thereby improving the interpretability of the answer content.

This invention discloses an improved Proof-of-Stake (PoS) consensus method based on proof-of-use, relating to the field of blockchain technology. The method includes the following steps: pre-setting multiple quantifiable on-chain behaviors and assigning a base weight and upper limit score to each type of on-chain behavior; in a consensus cycle, performing the decay and accumulation of behavior scores, and at the end of the consensus cycle, selecting all addresses with behavior scores greater than zero as candidates for the next round of consensus; selecting validators from the candidate set using a weighted random sampling algorithm by obtaining a global seed and the weights of each candidate address; publishing and validating the validator set, using the selected validators to produce new blocks in a pre-set order, and iterating through all transactions within the new block to update scores. This invention can adjust the weight coefficients according to different application scenarios, adapting to diverse needs and easily integrating into existing blockchain systems.

This invention belongs to the field of artificial intelligence security technology, specifically relating to a corpus security detection method, system, and computer-readable storage medium. The corpus security detection method includes: S1, preprocessing the original corpus to obtain paragraph-level corpus text; S2, inputting the paragraph-level corpus text into a five-dimensional detector to obtain a five-dimensional risk score; wherein, the five-dimensional detector includes a harmful information classifier, a sensitive entity identifier, a low-quality scorer, a false statement verifier, and a confidential fingerprint comparer; S3, weighting and fusing the five-dimensional risk scores to obtain a comprehensive risk score; S4, determining whether the comprehensive risk score is greater than a blocking threshold; if so, blocking is performed. This invention can perform integrated real-time detection of harmful, sensitive, low-quality, false, and confidential information in large-scale pre-training corpora before training, achieving real-time blocking before streaming into the database, and is applicable to corpus cleaning at various stages such as large model pre-training, fine-tuning, and continuous pre-training.

This invention relates to an English translation training method, apparatus, device, and medium based on a multimodal corpus. The method encodes source language text and paired images to generate basic multimodal features; these features are then processed by a teacher model and a student model respectively to obtain a teacher translation sequence and an initial student translation sequence; subsequently, a dedicated semantic validator evaluates the consistency between the two sets of translations and the original multimodal input, generating sample weight signals for dynamically adjusting knowledge distillation weights and bias signals for optimizing the validator itself; finally, these signals are used to calculate a multi-task loss, thereby updating the parameters of the student model and the semantic validator. This invention can improve model efficiency while suppressing error propagation at the source, and significantly improve translation accuracy by strengthening cross-modal semantic alignment.

A core data services (CDS) explain agent receives a CDS explain request to generate an explanation with respect to a CDS data model. A CDS explain process is orchestrated for the CDS data model using a CDS explain handler of the CDS explain agent. A validator of the CDS explain agent and advanced business application programming (ABAP) dictionary metadata is used to validate if a CDS entity exists and if the CDS entity can be explained. Using a prompt assembler of the CDS explain agent, a large language model (LLM) prompt is assembled by combining multiple prompt snippets that satisfy the explanation with respect to a CDS data model. Using a CDS explain request processor of the CDS explain agent, the LLM prompt is transmitted to an LLM. Using a post-processor of the CDS explain agent, relevant information of a response from the LLM is extracted.

This invention discloses an AI document content editing system and method for ensuring format fidelity. It employs a three-layer collaborative mechanism to guarantee format fidelity: the first layer is structural specification-level write isolation, where the format fidelity layer field does not appear in the AI ​​tool function input parameter schema, and the structural specification validator silently discards any format fidelity layer field passed in by the AI, physically preventing format damage during writing at the architectural level; the second layer preserves markup language skeleton anchors, storing the corresponding original markup language fragment as a skeleton anchor for each content block, replacing only the text-bearing nodes during export, eliminating format conversion losses during export; the third layer is a tool loop checkpoint chain, automatically capturing content block sequence snapshots before and after each tool call, supporting atomic rollback at any checkpoint. This three-layer collaboration produces an emergent technical effect: after the user rolls back the document to any checkpoint and exports it, all format attributes of the unmodified areas are consistent with the original document, eliminating the need for format repair steps.

Systems and methods provide an integrated technical control architecture for agentic systems, including large language model (LLM)-based agents, tool-calling assistants, and workflow orchestrators, in which delegated authority, connector invocation, and verification evidence share a common cryptographic lineage. An agent authority rail verifies an identity credential bound to an agent-held key, canonicalizes selected request fields, computes a request digest, derives a least-privilege capability scope with temporal constraints, and issues a capability token cryptographically bound to the verified identity credential, the request digest, the capability scope, and the temporal constraints. A connector trust mediation system verifies a connector descriptor, derives an invocation scope, isolates tool output from privileged instruction channels via a context firewall, and emits linked receipts. An append-only transparency log returns proofs and checkpoints. An offline verifier reconstructs lineage and validates proofs, witness requirements, and verification profiles without online access to a receipt producer.