Federated Human–AI
Intelligence

From LLM to Coordination Harness

The AI industry has moved through a clear progression. Each stage solves the previous stage’s limitation and reveals a new one. The newest turn is a return of symbolic structure: agent loops are unbounded by construction, and the industry is converging on formal semantics (ontologies shared across the mesh, rather than wired into each agent) as the way to bound them.

Three layers, kept distinct. The ontology is the formal vocabulary: the classes, typed properties and rules that define what can be said. The published corpus at narrativegoldmine.com is that ontology in readable form: pure TBox, every page a class, zero individuals by design. The knowledge graph is the ontology populated with live instances at runtime: VisionClaw’s running graph, agents’ working graphs, the personal graphs written into Solid pods. And the grounding layer, the Ontology Loom, serves the checked graph into a model’s context at query time, so answers restate verified facts rather than guesses. Holding it honest is a machine check at two points: the pipeline’s EL-profile reasoner gates the published closure at build, and a Whelk EL++ reasoner classifies the shared runtime graph and rejects contradictions before they enter it. When these pages say “reasoned”, they mean machine-checked, not an LLM thinking hard.

The Ontology Loom: measured grounding behind a model-swappable façade

We tested whether a formal ontology improves an LLM’s factual recall, objectively, with gold answers derived from the knowledge graph itself. Across every model we have put behind the façade (a frontier model at one end, a small local one at the other), grounding raised recall and cut hallucination in every case, roughly halving hallucination on the set (full report). The result is about the architecture, not the model: probabilistic agents bounded by machine-checkable semantics, with contradictions rejected at the Whelk gate before they enter the graph. We re-run the benchmark on each new model we deploy; the shape holds, and it keeps improving.

A held-out graph-derived benchmark (objective gold, bootstrapped intervals) on the current model, Qwen3.8-27B, across four grounding modes:

The static structured scaffold is the whole win: a bare model near random on the corpus answers at graph parity once the scaffold is injected, prose adds nothing, and letting the model walk the graph with tools tends to make it worse. The scaffold, not the model, does the work, which is exactly why the model can be a swappable URL. (The interference from injecting on weak matches is a known effect in the retrieval literature; the Loom gates for it, below.)

The Loom turns that result into infrastructure

The Ontology Loom is the node that serves this grounding as infrastructure. It consumes the corpus generations that knowledgeGraph publishes, weaves them into a reasoned ontology it holds as canonical, and serves that grounding to any LLM behind a stable, OpenAI-compatible façade. The model is a URL: Qwen3.8-27B today, swap for whatever benchmarks best next, with zero consumer change. VisionClaw supplies the GPU engine and the OWL 2 EL reasoner; knowledgeGraph builds the corpus; the Loom serves the grounding.

Beyond the benchmark, the node has hardened into infrastructure: