Verified adverse-action notices
When credit is declined, the applicant must be told the specific reasons. This system makes those reasons checkable.
A model scores. Code, not AI, decides the band, derives the reasons and writes the notice. A separate code verifier checks it, and a human sends it. An AI analyst sits beside the product to explain, never to decide.
How it works, and who owns each step
- 1Model
Score
A calibrated logistic regression estimates the chance of a bad outcome.
- 2Code
Explain
Code applies the cost policy and derives up to four principal reasons from the coefficients.
- 3Code
Write and verify
Code fills the notice template; eight independent code checks verify it.
- 4Human
Decide
A reviewer approves, or sends the verified notice. Overrides need a reason.
- 5AI (read-only)
Ask
An analyst agent answers questions and drafts reports. Every number it states is checked.
Verifier under attack A deliberately broken copy of a real notice (e.g. a reason dropped or a number changed). The verifier must reject every one.
Loading…
Cost per applicant Average cost per applicant under the 5-to-1 cost matrix. Lower is better.
Loading…
Finding we disclose
Loading…
What this is, and is not
A method demonstration on the 1994 UCI German Credit benchmark (1,000 applicants, amounts in Deutsche Marks). The review queue holds 60 synthetic applicants derived from it. It is not a lender, not compliance advice, and the reasons are model attributions, not causes. Read the honest limits.
Try it
- Open the review queue and pick an application.
- Score it. Watch the reasons, the eight verifier checks and the drafted notice appear.
- Decide as the reviewer: send the notice, change its wording, or override with a reason.
- Ask the analyst why, or have it draft a queue report.