Source Traceability

Explanation

Source traceability checks, for document-extraction and RAG systems, whether each output field can be tied back to the specific source it came from - which document, page, or passage produced the value - so reviewers can verify the system did not invent or misattribute data. This matters in banking because an extracted income figure, KYC field, or contract term that lacks a source reference cannot be trusted or defended in audit. It is measured as Traceability Score = Fields_with_SourceRef / Total_Fields, the share of output fields that carry a concrete source reference. To implement it, design the extraction/retrieval pipeline so that every emitted field is accompanied by a structured source pointer (document id plus location such as page, bounding box, or chunk id) captured at the moment the value is produced, and compute the ratio across all fields per document or per batch; persist the pointers as evidence so any field can be clicked through to its origin. The proposed thresholds are: target ≥ 0.90, investigate when below 0.85, and remediate when below 0.80. Dropping under 0.85 triggers an investigation into why references are missing, and falling below 0.80 triggers remediation of the extraction pipeline before the outputs are relied upon. Aligns with RBI Understandable by Design, NIST MEASURE 2.8, and ISO Clause 7.4 (all proposed).