Chunked Zero-Knowledge Attestation for Fair-Lending Metrics under SR 11-7

A recent paper on arXiv (2608.02664) introduces a zero-knowledge circuit design that facilitates the verification of aggregated fairness statistics in regulated machine learning decision-making, particularly concerning U.S. bank oversight as per SR 11-7 and OCC 2011-12 guidelines. Utilizing committed, nonce-sampled batches of genuine 2022 HMDA mortgage data, the system demonstrates the demographic-parity gap without revealing model weights or customer information. The authors achieve end-to-end performance with 32,768 rows, producing 32 independently verified zkSNARK proofs, with an attested gap of 0.0029 from the actual held-out value and proving times under 4 seconds per chunk. They further extend this by verifying expected calibration error across 10 bins, with all 32 chunks confirmed, achieving an attested ECE within 0.00037 of the plaintext value. This research tackles the issue of proving fairness and robustness to auditors in areas like credit underwriting and loan approvals while ensuring privacy. The submission is categorized as a cross-type on arXiv.