Data Lineage

Lineage records the full ancestry of every feature — which channels, calibrations, and transforms produced it — so any twin decision can be traced to raw bytes.

From decision back to bytes

When the twin forecasts a disruption or the failsafe context is questioned, the answer must be traceable: exactly which samples, calibrations, imputations, and feature versions produced the number that drove the decision. Lineage is the record of that ancestry, maintained for every engineered value in the fabric.

What lineage captures

• The raw channels and samples that fed a feature.
• The calibration and drift-correction versions applied.
• Any imputation of dead channels and the graph model version used.
• The feature definition version and normalization coordinate frame.
• The downstream models and decisions that consumed the feature.

A directed graph

Lineage is naturally a directed acyclic graph from raw samples up through validation, normalization, feature engineering, and into model inputs. The fabric stores this graph so any node can be traced up to its consumers or down to its raw sources. Combined with the immutable raw archive, this means any past feature can be recomputed and verified.

Why a fusion program needs it

Lineage is what turns the twin from a black box into an auditable instrument. It supports forensic analysis of a breeder disruption or a burner plug event, MLOps investigation of model drift, and the decision-audit requirements of a safety-relevant control system (see provenance audit trail). It is enforced by metadata governance and applies to both machines. This is a design and operational discipline for machines not yet built.