Lending
Spark
Loan & Card Decisioning Agent
Decisions auto-loan and credit-card applications on creditworthiness alone.
sonnet4 tools8 eval cases2 connectors
The case it handles
A representative case
Case fileL1-auto-prime
- Applicant
- Dana Whitfield
- Product
- auto
- Requested amount
- $31,000
- Term
- 60 months
- Credit score
- 768 (prime)
- Purpose
- vehicle purchase
Raw scenario input (JSON)
{
"applicationId": "L1"
}The contract
What it takes, does, and returns
Give it
- Applicant
- Product
- Requested amount
- Term
- Credit score
- Purpose
It does
- Get applicationget_application
- Get credit reportget_credit_report
- Get policyget_policy
- Record decisioncommitrecord_decision
It returns · Decision
- Approve
- Decline
- Counteroffer
- Refer
How it's checked
Checked against golden cases
8golden cases8baseline
- L1-auto-prime
prime auto → approve
- L2-auto-score-fail
subprime score below min → decline
- L3-auto-nearprime
near-prime, within policy → approve (counteroffer acceptable)
- L5-auto-ltv-fail
LTV above policy max → decline/counteroffer
- L6-auto-thin
thin file / no score → refer
- L7-card-prime
prime card → approve
- L8-card-score-fail
card score below min → decline
- L13-card-fcra
active FCRA dispute on a derogatory → refer
Each case is a real failure mode the grader checks on every change and as a deploy gate — regressions past threshold block the release.
Briefing
Briefing
- Makes the credit decision — approve, decline, counteroffer, or refer — from the application, the bureau file, and the bank's written credit policy.
- Consistency is the operating rule: identical credit profiles get identical outcomes, decided on score, ratios, capacity, and history alone.
- Blind to protected characteristics and their proxies by construction; the tools expose none and none may be inferred.
- Declines only on a hard policy failure it can name; thin files, disputed derogatories, and near-threshold results go to human referral rather than a guess.
- Records the specific factors behind each decision — downstream they become the adverse-action reasons, so each one has to be accurate and traceable to the file.
The receipts
A real run on a representative case, with and without governance
The same case, run with the Aarvion governor checking every tool call against policy before it commits.
Where it stands in the operation