An Explicit World Model Based on Data-First Ontology: DaoQL Multimodal Storage Validation and Counterfactual Reasoning Evaluation

Large language models (LLMs) encode world knowledge implicitly within their neural weights, leading to significant challenges in high-precision domains, including hallucinations, difficulty in updating knowledge, and a lack of explainability. The authors investigate whether moving deterministic knowledge into an explicit, multimodal database—a 'data-first ontology'—can provide the architectural guarantees necessary for reliable, auditable, and counterfactual reasoning that current implicit models lack.

The authors propose DaoQL, a multimodal database that integrates graph, column, vector, and full-text engines into a single process. The architecture is built on six primitives (Being, Def, Type, Relation, Contract, and Version) that allow for addressable, versioned, and audit-ready data. Theoretically, the authors formalize the difference between implicit and explicit models, proving that explicit models provide a sufficient condition for 'composable counterfactual decomposability'—the ability to reliably predict how a system's conclusions change when facts are altered. Empirically, the authors evaluate the storage layer using microbenchmarks, standard datasets like LDBC SNB and ANN-Benchmarks, and a five-domain counterfactual experiment.

DaoQL demonstrates significant improvements in reasoning reliability. In a counterfactual experiment involving 1,250 queries, the DaoQL+GPT-4o combination achieved 94% composable counterfactual decomposability, a 49-percentage-point improvement over GPT-4o alone. Performance measurements on an embedded setup show sub-millisecond latency for graph BFS (1.20 ms) and hybrid queries (105.8 μs). The authors argue that by unifying storage primitives and using direct mmap reads, they eliminate the cross-process overheads typical of existing RAG (Retrieval-Augmented Generation) frameworks.