One-Directional Derivation: Why the Fact Graph Never Reads from Perception
The Perception Graph is derived from the Fact Graph. Never the reverse. This one-directional constraint is the architectural guarantee of data integrity across the entire SiteNexis graph model.
One-directional derivation is the invariant that prevents the SiteNexis graph model from collapsing into circular data dependency. The Fact Graph contains DOM-observable facts with confidence 1.0. The Perception Graph contains AI-inferred relationships with confidence 0–1. The direction of information flow is strictly: Fact Graph → Perception Graph. Never the reverse.
What Happens If the Invariant Is Violated
If the Fact Graph were allowed to read from the Perception Graph, the following chain of failures becomes possible: an AI inference about an entity relationship updates the Perception Graph; a downstream system reads the Perception Graph and uses it to update the Fact Graph; the Fact Graph now contains an inferred relationship marked as a DOM fact with confidence 1.0; future Perception Graph derivations build on this corrupted fact; the corruption compounds with each audit cycle.
The Derivation Pipeline
The correct pipeline: (1) Crawl writes to fact_nodes and fact_edges, confidence always 1.0. (2) Perception Graph construction reads fact_nodes and fact_edges, plus page text and schema. (3) AI inference produces perception_nodes and perception_edges with confidence 0–1. (4) PCE is computed from the resulting Perception Graph. Nothing in steps 2–4 modifies fact_nodes or fact_edges.
●The derivedFrom field on every perception_node stores the set of fact_node IDs and page URLs from which the perception node was derived. This field makes the derivation auditable: for any entity in the Perception Graph, you can trace exactly which DOM facts led to that inference.
Why This Matters for Machine Trust
Machine trust analysis depends on comparing what is claimed (in schema and body text — Fact Graph) with what AI systems infer (Perception Graph). If these two are allowed to pollute each other, the comparison loses its diagnostic value. The Gap between fact and perception is where machine trust failures hide: entity attributes that are present in schema but never mentioned in body text, schema claims that contradict the page's actual content, AI inferences that diverge from what the DOM actually says. One-directional derivation keeps the gap visible and measurable.
Perception vs Fact Layer — Part 7 of 10