From Business Events to Auditable Decisions: Ontology-Governed Graph Simulation for Enterprise AI
New architecture exposes 'illusive accuracy' in models like Doubao-1.8, achieving a four-fold F1 advantage.
A team of eight researchers has introduced LOM-action, a novel architecture designed to solve a fundamental flaw in current LLM-based agent systems for enterprise use. The paper, "From Business Events to Auditable Decisions: Ontology-Governed Graph Simulation for Enterprise AI," argues that existing agents operate in an unrestricted knowledge space, leading to fluent but ungrounded decisions with no audit trail. LOM-action replaces this with an 'event-driven ontology simulation' pipeline. When a business event occurs, it triggers conditions encoded in an enterprise ontology, which then drives deterministic mutations in an isolated sandbox graph. All decisions are derived exclusively from this evolved, scenario-specific simulation graph (G_sim), ensuring they are contextually grounded.
The core innovation is the shift from direct LLM generation to a structured 'event → simulation → decision' workflow, realized through a dual-mode architecture with 'skill' and 'reasoning' modes. Every step produces a fully traceable audit log. In benchmark tests against leading models Doubao-1.8 and DeepSeek-V3.2, LOM-action achieved a 93.82% accuracy rate and a critical 98.74% F1 score for tool-chain execution. The competing models, while showing ~80% accuracy, scored only 24-36% F1, a gap the researchers term 'illusive accuracy'—where models appear correct but fail at reliable, structured action. This four-fold F1 advantage demonstrates that ontology-governed simulation, not simply model scale, is the architectural prerequisite for trustworthy enterprise decision intelligence.
- LOM-action uses event-driven ontology simulation to create deterministic, auditable decision graphs (G_sim) from business events.
- It achieved a 98.74% tool-chain F1 score, exposing the 'illusive accuracy' of models like Doubao-1.8 which scored only 24-36% F1.
- The architecture guarantees a fully traceable audit log for every decision, moving beyond fluent but ungrounded LLM outputs.
Why It Matters
Provides a blueprint for trustworthy, auditable AI in high-stakes business processes where traceability and reliability are non-negotiable.