The decision method, in full.
Seven typed-tool agents render the binding call, and a scaffold proves it. This is the method behind the home page teaser: how the call is made, and how it is grounded, fairness-tested, and replayable to the byte.
Seven agents on one stateful graph.
An application flows through intake, identity and fraud, income, credit, collateral, and policy. Each agent is typed to its tools and writes its findings to one shared state. The decision agent renders the call; the other six gather and assess.
- Intake
- Reads the application and the supplied documents, and normalizes them into the facts the later agents reason over.
- Identity and fraud
- Verifies the applicant and screens for fraud signals before the borrower is assessed.
- Income
- Establishes income and capacity from the applicant's evidence.
- Credit
- Assesses the borrower's credit against the lender's policy.
- Collateral
- Values the collateral pledged against the loan. EV battery valuation is one collateral type the underwriter can value.
- Policy
- Applies the lender's policy to the assembled facts and surfaces the clauses that bear on the call.
- Decision
- Renders the binding call: approve, decline, or price. The other six agents gather and assess; this agent decides.
A deterministic policy gate, before the model.
Hard cutoffs are applied as a deterministic gate before the model reasons. A fired cutoff is a grounded knockout: the decision rests on the rule that fired, not on a model inference, so the most consequential declines are the most certain.
Grounded reasons, or no reason at all.
A reason is only carried if it is grounded in both an explicit policy clause and a specific applicant fact. This is what lets the ECOA and FCRA notice reflect the real basis of the decision.
- Cites the clause
- Every reason names the exact policy clause it rests on, so a reader can read the rule and the decision together.
- Cites the fact
- Every reason names the applicant fact behind it, so the reason is tied to this application, not a general statement.
- Drops the ungrounded
- A reason that cannot be grounded in both a clause and a fact is dropped and flagged, never carried as a post-hoc rationalization.
The scaffold that proves the call.
The call is rendered by the decision agent; these layers make it defensible to a regulator, a credit committee, and a declined applicant alike.
- Four-fifths fairness
- Decisions are tested for disparate impact against a four-fifths standard over the decision population, and the testing is honest about sample size rather than asserting a result the data cannot support.
- Shadow monitor
- An interpretable scorecard runs in parallel to the agents. It flags drift and watches the AI; it never gates a decision.
- Deterministic replay
- Any decision can be re-run and reproduced to the byte against the exact policy version that was live when it was made. Tampering with an input is detected and named.
- Immutable audit
- The full agent trace and event timeline are sealed as an immutable audit. A human override is a separate layer over the record, never a rewrite of it.
Collateral the underwriter values.
The collateral agent values the collateral pledged against the loan as part of the decision. EV battery valuation is one collateral type the underwriter can value, a proof of capability rather than the company.