Agentic query layer proposed on top of a real, loaded supply chain graph. The foundation carries parts, vendors, weekly positions, ontology, provenance and caveats; the query layer turns that into source-cited answers.
Eight stages of answering one question: intake and intent classification, entity resolution, hybrid query planning, graph traversal, provenance and caveat assembly, answer composition, analyst sign-off, and an answer record.
Keep source records, relationships and caveats visible through the decision path.
Narrow model judgment to validated entities, allowed paths and deterministic flags.
High-impact output remains a recommendation or proposal until approved.

Approved source data and existing enterprise systems.
Known entities, relationships and deterministic business rules.
Bounded inference over verified context.
Human sign-off where confidence or impact requires it.
Evidence, caveats, decision state and outcome retained.