{"model":"zip_product_predictor","model_version":"zip_product_v1.0.0","components":{"agency_4class":{"artifact_version":"zip_product_v1.0.0","classes":["1_conv","2_fha","3_va","4_usda"],"n_features":23,"test_top1_accuracy":0.7791,"test_log_loss_cal":0.5631,"test_top2_accuracy":0.937,"training_cohort":"HMDA 2022-2025 first-lien originations, 4-class agency (Conv/FHA/VA/USDA), 17.5M loans"},"nonqm_sidemodel":{"artifact_version":"nonqm_sidemodel_v1.0.0","n_features":9,"full_feature_auc":0.9748,"hmda_only_ablation_auc":0.865,"training_cohort":"OB locks 24mo, Non-QM = NonConforming loan_type & loan_amount ≤ conforming limit (4.04M rows, 6.24% positive)"}},"jumbo_rule":{"trigger":"loan_amount > 1-unit conforming limit for the activity year","limits_by_year":{"2022":647200,"2023":726200,"2024":766550,"2025":806500,"2026":806500}},"coverage":{"n_zips_covered":20417,"coverage_note":"ZCTAs with ≥1 HMDA 2022-2025 origination (~20K of 42K USPS ZIPs)"},"fair_lending_audit":{"verdict":"MARKET_ESTIMATION_ONLY","approved_use_cases":["Market-share estimation: what product mix should we expect in this ZIP for a borrower profile like this?","Capacity planning / capital markets: expected FHA vs Conv volume in a MSA","Investor pipeline forecasting; product-mix trend reporting"],"prohibited_use_cases":["Individual borrower routing (do NOT use this to tell a specific borrower which product they should take)","Automated product recommendation to a borrower","Any workflow that could steer a Conv-eligible borrower AWAY from Conv"],"rationale":"Phase 4 fair-lending audit: model does NOT amplify demographic bias (regresses to mean) and does NOT steer Conv-feasible borrowers away from Conv (Conv AIR 0.95+ across all race/ethnicity/sex groups on the Conv-feasible subset). But the underlying market has real 3-4× demographic differentials in agency-product usage — traditional argmax AIR fails on FHA/VA/USDA for that structural reason. Safe as MARKET-ESTIMATION tool; unsafe as individual-borrower ROUTING tool."}}