| Context graph |
The shared, graph-based semantic memory agents reason over — the canonical model of the enterprise’s entities and relationships. |
| Graph traversal |
Retrieving context by navigating the graph’s real relationships out from the entities relevant to a query, rather than flat nearest-neighbor text search. |
| Neighborhood retrieval |
Returning only the connected subgraph relevant to the current query rather than the whole knowledge base. |
| Ontology |
The self-organizing schema of entity and relationship types discovered from ingested data, with human review of low-confidence mappings. |
| Policy-as-code |
Governance rules (built on open OPA and Cedar standards) evaluated at the tool-calling boundary before any action runs; verdicts are allow, deny, or allow-with-gate. |
| MCP |
Model Context Protocol — the open standard by which agents discover and call tools. Fabric is both an MCP server and an MCP client. |
| A2A |
Agent-to-agent interoperability — federating external peer agents as governed skills. |
| HITL approval gate |
A human-in-the-loop checkpoint that pauses a high-risk action for explicit, single-use approval. |
| Dry-run / simulation |
Executing an action against shadow state to preview its before/after effect with zero impact on production. |
| Lineage |
The provenance chain answering “which data influenced this action” — source to ingestion to entity to agent to decision. |
| WORM audit |
Write-once-read-many, hash-chained, tamper-evident audit storage; audit events are never sampled. |
| Sovereignty routing |
Automatic routing of classified data classes to local or private models so regulated data stays in-boundary. |
| Deployment posture |
The per-organization choice of managed cloud, private cloud / on-premises, or air-gapped operation. |
| Workspace |
The handle for a project (ws_<project>); the scope an agent’s credentials are pinned to. |
lifecycle_id |
The identifier returned on every call, tying together its trace, policy decision, approval, and audit entry. |
