Skip to content
Intelligence plane

Ground decisions in traceable evidence

The Intelligence plane specifies stable identity and source-backed evidence. Begin with the records your workflow needs; add ontology and graph retrieval when joins, conflicting sources, or repeated retrieval failures justify them. Evidence must remain current, in scope, and traceable to its source before the Context plane admits it. The actual reader, processor, purpose, tenant and retention rules apply again at use; a historical snapshot or digest does not restore expired access or establish factual truth.

Abstract ContextOS knowledge grounding visual showing source records, identity handles, a knowledge graph, evidence references, validation, and a pinned bundle.
Evidence-bound knowledge

Ontology, CEID/SID identity, graph edges, evidence references, and eligible snapshot pins support scoped audit and replay.

OntologyCEID/SIDKnowledge GraphEvidence refsSnapshot pin
Grounding path
  1. 01Ontology

    Versioned definitions for entities and relationships.

  2. 02CEID/SID

    Stable handles for audit and compact ML identity.

  3. 03Knowledge graph

    Connects facts, entities, and relationships.

  4. 04Evidence refs

    Every retrieved hop points back to source evidence.

  5. 05Snapshot pin

    Declared replay checks require available, eligible retained versions.

CEIDs and SIDs

CEIDs are stable, human-readable handles for audit and prompts. SIDs are compact discrete-token sequences for ML feature use (retrieval, ranking, recommendation). Different jobs, different identities.

Failure to prevent

Split-brain definitions: two teams defining "active customer" differently. The ontology records canonical definitions per version; implementation validation and governed evolution keep uses aligned.

Metrics that matter

Evidence coverage, retrieval precision/recall on golden sets, conflict rate by relationship type, replay agreement on declared properties, stale-access denial and unresolved-conflict handling.

Evidence-bound retrieval

GraphRAG is one retrieval option. The typed EvidenceBundle below is a constructed example, not a live retrieval or verified truth claim. The reference compiler consumes caller-supplied evidence; it does not fetch or authenticate a graph. Source references and snapshot/ontology versions support reconstruction when eligible artifacts remain available. Deterministic reference properties and model-dependent judgments need separate replay checks.

Enforce classification and disclosure for actual recipients across collection, inference, tool input/output, telemetry, cache, memory, export and delegation. Correcting or deleting a source requires an owned copy inventory and target-specific propagation receipts; unavailable or unknown copies remain visible. Missing or conflicting evidence blocks affected decisions until the required resolution is verified. See data and memory controls and evidence-scoped conformance.

Constructed EvidenceBundle (not a live retrieval)
{
  "snapshot_version": "kg_2026_05_03_T0930",
  "ontology_version": "ont.support@4.2.0",
  "seed": { "ceid": "order:ord_881", "entity_type": "Order" },
  "hops": [
    { "hop": 1, "edge": "order_belongs_to_customer",
      "evidence_ref": "oms:db:orders/881#row_v3",
      "to": { "ceid": "customer:cus_77" }, "confidence": 0.99 },
    { "hop": 2, "edge": "customer_has_segment",
      "evidence_ref": "crm:exports/segments_2026_05_01.parquet#offset_4112",
      "to": { "value": "vip" }, "confidence": 0.96 }
  ],
  "conflicts": [],
  "truncated": false
}