
ECOA requires the specific reasons for adverse action. The CFPB says a complex model is no excuse. Here is a decision format that produces the reasons at decision time, by construction.
System notes
ECOA at 15 U.S.C. § 1691(d) entitles a rejected credit applicant to a statement of specific reasons, and the statute itself defines sufficiency: the statement must contain the specific reasons for the action taken.
Regulation B at 12 C.F.R. § 1002.9 requires notification within 30 days and a statement of the specific principal reasons; the official interpretations reject vague or general statements as insufficient.
CFPB Circular 2022-03 states that the adverse-action requirements apply regardless of the technology used, and that a creditor may not lawfully use a complex algorithm when it cannot identify and state the specific reasons for the adverse action.
CFPB Circular 2023-03 states that checking the closest entry on the Regulation B sample checklist is not compliant when it does not reflect the actual principal reasons for the adverse action.
In the governed decision format, each model seat must emit a per-clause derivation — whether each clause's condition fired, whether it supports or defeats the action, and on which evidence records — plus the records that were absent and the finding that would flip the verdict, all against a rule set pinned to a content hash.
The derivation-agreement gate refuses to authorise when independent seats reach the same verdict through different derivations; a unanimous verdict has been refused on the record for exactly this.
Every sealed decision leaves a permanent public receipt carrying the hashes, contract, and lineage, with the complete request and response credentialed behind it — so the reasons on the notice can be checked against the reasons in the record, at any later date.
In the first calibration study — 30 oracle-labelled synthetic cases through the production gate, three seats across two model families — glm-5.2 scored 30/30, kimi-k2.7 29/30, and the gate produced zero wrongful authorisations across all 30 cases.
Where a lender's underlying scorer is a separate machine-learning model, this format governs rule-application decisions — policy overlays, exception handling, verification-driven denials — and does not produce the reasons of the scoring model itself; Regulation B requires the actual principal reasons from whatever actually scored the applicant.
No conformance analysis against Regulation B's sample notification forms or its specific notice-content requirements has been performed; the calibration fixtures are synthetic and determinate; the running exhibits use three seats across two model families, not three.
Evidence ledger 10 · tier-ranked · API
5 more ranked claims
Ask this article · 8 suggested prompts
Text the build (+14245134626) or WhatsApp — slug|question creates a question node. Paste evidence with ingest slug|q:NODE_ID|your paste.