Field Justification

Explanation

Field justification checks, for document-extraction and RAG systems, whether each output field is accompanied by a human-readable justification explaining why the system produced that value - the reasoning or evidence behind the extracted or retrieved answer, not just a raw source pointer. It matters because reviewers and downstream decision-makers need to understand why a field holds its value (for example why a particular figure was classified as net income) to trust and challenge it. It is measured as Justification Completeness Score = Fields_with_Justification / Total_Fields, the share of output fields that carry a justification. To implement it, have the extraction/retrieval pipeline emit, alongside each field, a short structured rationale (the cue text, rule, or model evidence that drove the value) generated at production time, then compute the ratio of justified fields to total fields per document or batch and persist the rationales as evidence available in the review interface. The proposed thresholds are: target ≥ 0.90, investigate when below 0.85, and remediate when below 0.80. Falling under 0.85 triggers investigation into which fields lack justifications and why, and dropping below 0.80 triggers remediation of the pipeline. Priority is Later. Aligns with RBI Understandable by Design, NIST MEASURE 2.9 Explainable and interpretable, and ISO Clause 6.2.2 and Annex A explainability (all proposed).