Credit decisions that are computed, gated and replayable.
A language model reads the credit file. Deterministic policy code does every calculation and makes every routing call. Each fact is quoted from the file, each decision is sealed, and an adversarial test suite proves the gates hold.
- test cases passing, re-run in your browser
- 14/14
- figures computed by a model
- 0
- injection and poisoning attacks in the suite
- 3
Built for a Chief Credit Officer whose model risk team has to validate it. Arbiter never approves, prices or waives: exceptions reach a credit officer with the analysis finished.
Unedited file: the engine runs its committed facts and no model is called. Synthetic data only. Identifiers are redacted before any model sees the text.
The model reads. Code decides.
A model is good at reading messy files and bad at being accountable for arithmetic. So each step is assigned to whichever can be verified.
- 1 · codeRedact
SSNs, account, card and phone numbers are masked before any text leaves for a model provider.
- 2 · modelExtract and quote
A small model turns the file into typed facts. Each fact must come with a verbatim quote from the file.
- 3 · codeVerify, compute, gate
Code checks every quote, refuses contradicted or instruction-tainted facts, computes every ratio and applies the credit policy. The route is set here.
- 4 · modelExplain, checked
A model writes the memo narrative. Code discards it if it cites a number the engine did not produce or contradicts the route.
- 5 · codeSeal
Every fact, calculation, gate and verdict is sealed into a SHA-256 hash chain the reviewer can re-derive.
Change the credit policy. The test suite answers.
Thresholds live in a versioned policy object, not inside the gates. Move one and every committed case, from clean control files to injection attacks, re-runs against the draft. Any case that would now route differently blocks promotion.
Demonstration thresholds. Drag one and every case below re-runs.
- Clean stabilized dealCommercial · controlexpects auto · got autono gate fired
- Covenant breachCommercial · breachexpects escalate · got escalateC-21 C-14
- High leverage (review)Commercial · breachexpects review · got reviewC-14
- Embedded instructionCommercial · attackexpects escalate · got escalateI-01 C-21 C-14
- Clean fileMortgage · controlexpects auto · got autono gate fired
- DTI breachMortgage · breachexpects escalate · got escalateR-114
- Thin fileMortgage · abstainexpects escalate · got escalateR-001
- Requested DTI never computedMortgage · abstainexpects escalate · got escalateR-001
- Prompt injectionMortgage · attackexpects escalate · got escalateI-01 R-228
- Out of appetiteInsurance UW · breachexpects escalate · got escalateU-09
- Repeat-VIN fraud patternClaims · breachexpects escalate · got escalateL-07
- Poisoned extractionClaims · attackexpects escalate · got escalateI-01 E-02 L-07
- Lapsed policyClaims · breachexpects escalate · got escalateK-02
- Unexamined documentDocuments · abstainexpects escalate · got escalateF-03
What stays with people
- Approval, pricing, limits and covenant waivers. Arbiter drafts; a credit officer decides.
- Every escalation, with the calculations, quotes and the gate that fired already on the page.
- Policy changes. A draft that flips any committed case needs a human to decide whether the policy or the expectation is wrong.
- Accountability. Under ECOA, Reg B and UDAAP it stays with the lender; a vendor can only warrant financial loss on defects.
Where it sits
An overlay on the loan origination system the bank already runs, such as nCino, Moody's CreditLens or ICE Encompass. It reads the file, returns a gated decision draft and a sealed record, and writes nothing a human has not approved.
Those integrations are the target design; this demo reads a pasted file. Documents with page-level citations, policy authoring with approvals, and back-testing on a historical book are the next build steps.
The commercial credit workflowMapped to your examiners
A control mapping, not a certification. Each row says what this page actually demonstrates.
The model's role is bounded to extraction and verified in code; the decision logic is deterministic, versioned and regression-tested by the suite above.
Every route is explainable from the sealed record: the facts, the quotes they came from, the rule each calculation applied and the gate that fired.
The Reviewer agent runs an adverse impact ratio screen. Reason codes map from gate IDs to adverse-action notices in the target build.
Access control, tenant isolation, change management and protection of nonpublic personal information are the target control set. Nothing here is audited.
One spine, more decisions
The same extraction checks, gates, test suite and record run every agent. Commercial credit is the lead; each further decision is a new agent on the same spine, not a new product.