From Business Events to Auditable Decisions: Ontology-Governed Graph Simulation for Enterprise AI
Summary: arXiv:2604.08603v1 Announce Type: new
Abstract
In the realm of artificial intelligence, existing large language model (LLM)-based agent systems face a critical architectural challenge. These systems often operate from an unrestricted knowledge base, failing to simulate the impact of active business scenarios on that knowledge space. As a result, while their responses may seem fluent, they lack grounding and do not provide an audit trail for decision-making. To address this issue, we introduce LOM-action, a novel framework that integrates event-driven ontology simulation into enterprise AI.
The LOM-action Framework
LOM-action enhances enterprise AI capabilities by utilizing business events to trigger scenario conditions that are encoded within the enterprise ontology (EO). This process facilitates deterministic graph mutations within an isolated sandbox environment, allowing a working copy of the subgraph to evolve into a scenario-valid simulation graph, denoted as Gsim. Importantly, all decisions made are derived exclusively from this evolved graph.
Core Pipeline: Event to Decision
The core operational pipeline of LOM-action can be summarized in three stages: event → simulation → decision. This pipeline is supported by a dual-mode architecture that comprises two operational modes:
- Skill Mode: This mode allows for the execution of predefined skills based on the current state of the simulation graph.
- Reasoning Mode: In this mode, the system applies reasoning capabilities to derive insights and decisions from the simulation graph.
Audit Trails for Decision-Making
One of the standout features of LOM-action is its ability to generate a fully traceable audit log for every decision made. This ensures that each decision can be reviewed and validated, thereby enhancing the overall trustworthiness of the enterprise AI system.
Performance Metrics
LOM-action has demonstrated impressive performance metrics, achieving an accuracy rate of 93.82% and a tool-chain F1 score of 98.74% when benchmarked against leading systems such as Doubao-1.8 and DeepSeek-V3.2. In contrast, these frontier baselines only reached F1 scores between 24% and 36%, despite reporting accuracy levels around 80%. This discrepancy highlights the illusive accuracy phenomenon prevalent in current AI models.
Conclusion
The significant four-fold advantage in F1 score confirms that ontology-governed, event-driven simulation is a critical architectural requirement for developing trustworthy enterprise decision intelligence. As organizations increasingly rely on AI for decision-making, frameworks like LOM-action will be essential for ensuring that these decisions are both grounded and auditable.