Structured Data Lineage

The Lineage module records directed relationships between datasets, AI systems, and use cases — implementing ISO/IEC 42001:2023 Annex A.7.6 (Data quality and lineage). It is opened from the Lineage button on each row of the Dataset Registry (/datasets) and is available to Enterprise tier when the per-organization ISO 42001 toggle is enabled.

Edge kinds

Each edge is directed (source → target) and tagged with one of four kinds:

feeds — dataset supplies data to a downstream dataset or system.
derives_from — dataset is produced by transforming another dataset.
trains — dataset trains a model / AI system.
evaluates — dataset is used to evaluate or benchmark a system.

Endpoints can be datasets, AI systems, or use cases. Uniqueness is enforced per (org, source kind+id, target kind+id, edge kind).

Graph & adjacency views

The lineage sheet renders two complementary views of the same edge set:

Graph view — interactive React Flow canvas with layered columns (datasets / systems / use cases). Lazy-loaded (~120 KB) only when opened.
Adjacency table — always visible, screenshot- and print-ready, and the only view on small viewports.

Cycle detection

A cycle in lineage usually indicates a modelling error (e.g., a dataset that "derives from" something that depends on it). The graph helper runs a Kahn's-algorithm topological sort on every render; any nodes participating in a cycle are highlighted in the graph view and a warning banner appears above the form.

Auditing & access

All edge changes write immutable entries to the org and Platform Activity Log:

LINEAGE_EDGE_CREATED, _UPDATED, _DELETED
LINEAGE_GRAPH_EXPORTED — reserved for the future PNG export action
LINEAGE_CYCLE_DETECTED — reserved for cycle telemetry

Read access is org-wide; admin / approver / reviewer roles can manage edges. Multi-org lineage (cross-tenant edges) is forbidden by the tenant-isolation model.