Skip to content
Press / to search

Enterprise Data Stewardship

Agentic data quality, lineage, schema drift, and semantic drift management with steward review and replayable decisions.

Use Case PlaybookLast reviewed:
At a glance

Purpose

Keep enterprise data usable for agents, analytics, and operations as schemas, business meanings, and ownership change. A conforming runtime can turn data stewardship into an agentic workflow: detect drift, plan remediation, validate lineage, propose fixes, and route policy-designated changes through accountable stewards.

This is an illustrative deployment playbook. Admit contextos.enterprise@1 and actual data/network/memory features; durable or delegated jobs add their required controls/profile. Roles can be logical responsibilities in one service. Production supplies the actual stores, sandbox/data enforcers and current evidence; the reference does not run pipeline mutations or certify lineage.

Why this is agentic-first

Data quality work is continuous and contextual. A pipeline break can be caused by schema drift, stale ownership, a changed business definition, missing lineage, or an upstream vendor change. IBM describes agentic data management as intent-driven orchestration where agents generate a plan, execute and validate workflows, adapt to schema or semantic changes, and keep guardrails active.

For ContextOS, this use case is foundational: agents cannot make trustworthy decisions if the knowledge substrate is stale, semantically inconsistent, or unaudited.

Context Pack

LayerRequired entries
decision_layer.decision_specs[]data.drift.classify, data.quality.remediate, data.semantic_change.approve, data.lineage.certify
policy_layer.policy_bundles[]Data classification, PII handling, domain ownership, lineage, retention, metric-definition policy
policy_layer.approval_gates[]GATE_DATA_STEWARD, GATE_PRIVACY_REVIEW, GATE_METRIC_OWNER
tooling_layer.adapter_registry[]adp_catalog.query, adp_lineage.trace, adp_quality.profile, adp_pipeline.patch, adp_semantic_layer.update
memory_layer.memory_policy.write_classes_allowedevidence_link, correction, decision_outcome; schema/semantic pattern names are domain candidate content, not new write-class enums
evaluation_layer.eval_targets[]P/U/L/S/E targets; drift precision, data-quality SLA, steward overrides and lineage completeness belong to the linked domain evaluation suite

Agent roles

AgentResponsibilityBoundary
Drift AgentDetects schema, freshness, distribution, and semantic drift.Read-only until remediation is approved.
Lineage AgentMaps upstream and downstream impact.Cannot certify lineage alone.
Quality AgentProfiles data and proposes remediation.Can write only to sandbox datasets.
Semantic AgentCompares metric and entity definitions across systems.Cannot change canonical definitions without owner gate.
Steward AgentChecks owner, privacy, lineage, and rollback requirements.Can block or escalate.

Execution flow

  1. invokeAgent arrives with intent=data.steward, dataset ID, anomaly type, and business domain.
  2. Trusted admission resolves/pins catalog metadata, lineage graph, ontology, privacy policy and owner map; the pure compiler consumes supplied inputs rather than querying or authenticating those services.
  3. Drift Agent classifies the anomaly as schema, freshness, distribution, lineage, or semantic drift.
  4. Lineage Agent builds impact radius across dashboards, features, workflows, and agent tools.
  5. Quality Agent proposes remediation and tests it in a sandbox.
  6. Semantic Agent compares business definitions and flags meaning changes not visible in schema.
  7. Steward Agent verifies ownership, privacy policy, and rollback.
  8. Before any patch, the Gateway checks the exact stored action/approval binding, current owner/tenant authority and revocation, observed sandbox/egress/data route and original aggregate reservation; then claims the stable operation under current store revision/generation.
  9. The implementing runtime emits a DecisionRecord with decision_key="data.quality.remediate".

Verify the actual schema/data/lineage postconditions before reporting remediation. A lost mutation acknowledgement remains unknown until source-system reconciliation; a key or checkpoint cannot justify blind reapplication. Resume keeps exact interpretation and current ownership/fencing, authority and original limits. A changed ontology/policy/target definition needs impact review, explicit migration and reauthorization.

Apply actual purpose/source-label/reader/processor/residency policy across profiling extracts, model inference, telemetry, caches, indexes and exports. Minimize sampled production records in sandboxes; a sandbox dataset label does not permit copying sensitive rows. Correction/deletion completion requires the scoped complete copy inventory and verified propagation receipts, with external unknown copies disclosed. See Execution Contract.

Decision gates

GateTriggerRequired evidence
GATE_DATA_STEWARDAny pipeline, catalog, or lineage change affecting certified data.owner, lineage impact, quality profile, rollback
GATE_PRIVACY_REVIEWPII class, retention, access, or masking change.privacy class, affected users, legal basis
GATE_METRIC_OWNERSemantic-layer or KPI-definition change.old/new definition, downstream impact, approval owner

Failure modes

  • Silent semantic drift - schema stays stable but business meaning changes; Semantic Agent must compare definitions and usage.
  • Pipeline patch without impact analysis - Tool Gateway blocks writes unless lineage impact is attached.
  • Privacy class downgrade - reject self-authorized relabeling; protected independent review and verified permitted transformation/source-result binding are required. An owner click cannot waive mandatory data or applicable legal requirements.
  • Bad remediation loop - sandbox validation must pass before production patch.
  • Memory poisoning - proposed schema or semantic learnings enter review, not automatic promotion.

Metrics

  • Drift detection precision and recall.
  • Mean time to classify and remediate data incidents.
  • Certified lineage coverage.
  • Steward approval latency.
  • Privacy gate violations blocked.
  • Downstream incident recurrence after remediation.

Research signals