Data Provenance

Explanation

Data Provenance checks that every dataset used to train, fine-tune, or ground the AI system has documented origin, lineage, and usage rights, so the institution knows where its data came from and can defend its legitimacy — essential in BFSI where training a model on data of unknown or improper origin can introduce bias, poisoning, privacy violations, or licensing and regulatory breaches into a lending or fraud model. It is measured by Training-Data Provenance Coverage (metric and calcMethod identically named): the share of training data for which provenance is documented. Implement it by maintaining a data catalogue or lineage record that captures each dataset's source, collection method, licensing/consent basis, transformations, and approval status, and by gating data ingestion so that undocumented data cannot enter the training or retrieval corpus; log provenance records and link them to the datasets actually used so coverage can be audited. The threshold requires 100% of data to have documented provenance: any dataset lacking a complete provenance record is a breach that must be flagged and resolved before that data is used, reflecting the control's Immediate priority and its role in supply-chain integrity.