Skip to content

Replay of a real run — captured 2 Jul 2026

Equipoise

Fair-Lending Reviewer

Every step below is a real trace event from the signed run ledger.

sc-equipoise

Transcript

Equipoisesonnet
Steps14
Tools7
Tokens51k
Elapsed1:25
Briefed on the case

I'll kick off the review by pulling everything I need in parallel — the application file, credit policy, and monitoring attributes all at once.

Good — I have the application, policy, monitoring data, and one comparator. Now I need to pull the credit reports for both L11 and comparator L12 before I can assess credit-profile parity.

I now have a complete picture. Let me compile my analysis before recording.


Pre-Record Analysis

Step 1 — Is the decision supported by policy?

CriterionPolicy LimitL11 ActualPass/Fail
Min. credit score (auto)≥ 600666 PASS
Max. DTI≤ 45%13% PASS
Max. LTV≤ 125%102% PASS
Derogatories / public recordsNone hard-stop0 / 0 PASS
Thin fileRefer triggerFalse PASS

Critical finding: The policy's near-prime band (660–719) maps to a 9.49% rate — it is an approved tier, not a decline category. The minimum score for decline on auto is below 600. A score of 666 clears every hard threshold. The stated policy basis ("score 666 in the near-prime band") does not constitute a valid decline trigger. The decision is unsupported.


Step 2 — Outcome Testing: L11 vs. L12 (Gregory Hale)

AttributeL11 (André Mensah)L12 (Gregory Hale)
Score666669 (+3 pts)
Utilization33%31% (-2 pts)
DTI13%~12%
LTV102%~102%
Derogatories00
Public records00
Inquiries (6 mo)22
Oldest account78 mo80 mo
Employment tenure48 mo45 mo
Tradelines (all current)22
Product / purpose / channelAuto / purchase / directAuto / purchase / direct
DecisionDECLINEAPPROVE
Race/ethnicityBlackWhite

The profiles are materially equivalent. L11 actually has greater employment tenure. The 3-point score gap and 2-point utilization gap are de minimis and lie nowhere near any policy threshold. There is no legitimate, policy-grounded explanation for the divergent outcomes. The sole meaningful axis of difference is race/ethnicity.

Step 3 — Proxy check: The stated factors ("credit score in the near-prime range," "moderate revolving utilization") contain no facially race/ethnicity-proxying language. However, these same factors were present in L12's file, which was approved — the factors were applied selectively, which is itself a form of disparate treatment.

Verdict: VIOLATION — A similarly-situated White applicant (L12) received a materially more favorable outcome than a Black applicant (L11) whose file fully satisfies policy, with no policy-grounded explanation for the difference.


The decision brief

Verdict
Violation

Fair-Lending Review — Final Report: VIOLATION

Application: L11 — André Mensah | Auto Loan, $24,000 | Decline


Decision Supported by Policy? No.

Every hard policy threshold for an auto approval is cleared:

  • Score 666 ≥ 600 minimum
  • DTI 13% ≤ 45% cap
  • LTV 102% ≤ 125% cap
  • No derogatories, no public records

The near-prime score band (660–719) triggers a rate of 9.49%, not a decline. The stated policy basis is factually inapplicable.


Outcome Testing — Driving Comparator: L12 (Gregory Hale)

L12 is a near-identical applicant — same product, channel, purpose, employment category, loan structure, and vehicle — who was approved. The only credit differences are 3 score points and 2 utilization points; L11 actually has longer employment tenure. There is no legitimate, policy-grounded explanation for the divergent outcomes.

L11 (André Mensah)L12 (Gregory Hale)
Race/ethnicityBlackWhite
Score666669
OutcomeDecline failApprove pass

Findings Summary

  1. Unsupported decline — no hard-stop criterion is triggered; near-prime is a rate tier, not a decline category.
  2. Disparate treatment — materially equivalent White applicant approved while Black applicant declined; no policy basis distinguishes the outcomes.
  3. Selective factor invocation — the cited factors ("near-prime score," "moderate utilization") appear in both files; only L11's application was denied.
  4. Anomalous confidence score (0.62) — suggests a non-policy factor may have influenced the decision.
  5. No proxy language detected in the stated factors.

Recommendation

🚨 Escalate immediately to Fair Lending Compliance and General Counsel. Rescind the decline and re-underwrite L11 (approve at the 9.49% near-prime rate). Audit all near-prime auto decisions over the past 12–24 months for systemic patterns. Identify involved underwriter(s). Assess regulatory self-reporting obligations if a pattern is confirmed. Strengthen second-line controls requiring documented policy-basis justification on any decline where all thresholds are met.

Cost$0.1235
Turns8
Time85.1s
Recorded to the audit ledger · tamper-evident