Deep dive — methodology
Counterparty-question rehearsal — the 5-decision reproducibility drill
Every $499 Snapshot samples five AI-driven mortgage decisions from a target day and reproduces them defensibly. Not because you need them all today — because the discipline of reproducing five proves the discipline could reproduce any of them tomorrow.
The prompt the rehearsal answers
"Reproduce this AI-generated mortgage-underwriting decision from Day 47 including the applicant's stated purpose, credit bureau snapshot, property valuation methodology, model version, underwriting-risk-score, and lane routed — as a defensible record we can produce for CFPB supervisory examination, DOJ fair-lending referral, HUD complaint response, state banking-department examination, or class-cert opposition."
The counterparty question — identical shape whether it comes from CFPB, DOJ, HUD, a state examiner, or plaintiff counsel
Sample application decisions from Day 47
| Application ID | Risk Score | Lane | Model version | Decision hash |
|---|
| LA-002534 | 44.17 | Manual Review + Upcharge | mortgage-underwriter-v4.7.2 | muw-002534 |
| LA-002533 | 82.10 | Decline / Refer non-QM | mortgage-underwriter-v4.7.2 | muw-002533 |
| LA-002532 | 70.10 | Decline / Refer non-QM | mortgage-underwriter-v4.7.2 | muw-002532 |
| LA-002531 | 86.36 | Decline / Refer non-QM | mortgage-underwriter-v4.7.2 | muw-002531 |
| LA-002525 | 51.16 | Manual Review + Upcharge | mortgage-underwriter-v4.7.2 | muw-002525 |
What "defensibly reproduced" means
- Model version pinned. Not "the current model" — the exact model version deployed at decision-time.
- Input snapshot bound. Applicant group + loan product + property type + loan amount + credit tier + LTV + DTI at application-time exactly as they were when the AI scored the file.
- Decision hash bound. Cryptographic hash tying the input to the output — makes tampering detectable.
- Retention pipeline independent. The retention system that holds the decision records is not the production model itself — independence of the record from the actor being recorded.
What the rehearsal proves
All 5 sampled Day-47 mortgage decisions reproduced with defensible-records match. Model version pinned, decision hash cryptographically bound, input data retained via decision-provenance record. If CFPB / DOJ / HUD / state examiner / plaintiff counsel requests any of the ~4,969 applications in the audit period, the same reproduction procedure applies.
Retention horizon
- HMDA Sec 304 record retention: minimum 3 years for LAR + 5 years for underlying records
- SR 11-7 Section V.B model documentation: life of model + 3 years post-decommission
- CFPB Circular 2022-03 preservation duty: adverse-action reason-code accuracy retention
- ATR/QM life-of-loan liability: aligns with 30-year mortgage horizon
- Litigation hold: preserves through matter conclusion, may extend original horizon
The rehearsal record is not a one-time exhibit — it is proof that the retention discipline is operational.
Why 5, not 500
Reproducing 500 decisions on request is a discovery-response exercise, appropriately triggered by a specific legal process. Reproducing 5 sample decisions at Snapshot time is a discipline-test: it validates that the retention pipeline works, without exhausting the audit budget on record production nobody has asked for yet.
$499 Snapshot. 3 business days.
Same 5-decision reproducibility drill on your lender's actual AI surface + counterparty-question rehearsal record.
Buy $499