ATLASky-AI: An autonomous framework for physics-based trustworthy verification of LLM-generated spatiotemporal knowledge

• A novel Five-module (Agents) architecture detects hallucinations missed by single-method systems. • Physics constraints catch spatiotemporal errors invisible to semantic checks. • Achieves high precision verifying LLM outputs across aerospace and healthcare. • Autonomous operation without labeled ground truth for each verification decision. • Reduces false positives 39–57%, with faster graph creation and 3.5-month ROI in aerospace. Large Language Models (LLMs) promise to revolutionize knowledge extraction from unstructured data in systems that track physical entities through spacetime, but their tendency to hallucinate (generating plausible but false information) renders them too risky for deployment in safety-critical domains. Prior verification methods are caught in a circular dependency, requiring the ground-truth labels they aim to produce, validating spatial and temporal dimensions independently rather than concurrently, and suffering catastrophic accuracy degradation under distribution shift. To address these limitations, we introduce ATLASky-AI, an autonomous trustworthy AI framework for verifying LLM-generated facts in 4D Spatiotemporal Knowledge Graphs used across safety-critical domains including digital twins, healthcare tracking systems, manufacturing, and logistics. The framework makes real-time verification decisions autonomously using intrinsic quality metrics, eliminating the need for instance-level labeled ground truth during runtime inference. An adaptive monitoring system leverages sparse expert feedback (0.3% sampling) to maintain high accuracy over time under distribution drift. The core architecture orchestrates five specialized verification modules targeting distinct error patterns: ontology validation, standards compliance, physics-based spatiotemporal constraint checking, external source corroboration, and statistical anomaly detection. Evaluated across aerospace and healthcare datasets, ATLASky-AI achieves an average of 94% precision and 93% recall, with false positive rates of just 2.6–4.1%. This represents a 39–57% reduction compared to the best baseline system. A six-month pilot deployment at an aerospace manufacturing company, AddQual Ltd., validates its practical impact, demonstrating a 99.8% reduction in knowledge graph creation time (from 48 hours to 4.2 minutes) and a 3.5-month return on investment. These results establish ATLASky-AI as an effective trustworthy AI solution, enabling safe integration of LLMs with mission-critical systems requiring spatiotemporal knowledge validation.