Form Factor Bias

Explanation

Form factor bias checks whether a document-extraction system performs equally on different input formats — specifically scanned/photographed documents versus clean digital files — so customers who can only submit a phone photo of a paystub or ID are not disadvantaged versus those uploading native PDFs. It matters because lower-income or rural customers more often supply scanned inputs, and a quality gap there translates into unfair friction and rejections. It is measured by Input Format Equity, computed as |F1_scanned − F1_digital|: the absolute difference in extraction F1 between scanned and digital inputs. To implement it, build paired labelled test sets covering both scanned and digital versions of representative documents, compute F1 for each format, take the absolute gap, and log per-format scores and failure examples as evidence; surface the gap on a fairness dashboard. Run it pre-deployment and periodically, re-testing after OCR/model updates. The proposed thresholds target at least 0.90, with investigation below 0.85 and remediation below 0.80; falling under 0.85 triggers root-cause analysis and falling under 0.80 triggers mandatory remediation — improve pre-processing, denoising, or model robustness for scanned inputs, re-test, and document — supporting the proposed RBI Fairness and Equity and NIST MEASURE 2.11 fairness evaluation.