Skip to main content
Decision Engine

The decision engine that produces decisions you can explain to an examiner.

Hard policy rules run first. ML augmentation fills the grey zone. Both outputs are written to the same auditable decision record.

Request Demo
Architecture

How rules and ML coexist in one execution layer

The decision engine does not choose between rules and ML — it runs both in the same execution context. Hard policy rules execute as filters first. Applications that pass all hard rules enter the ML scoring zone. The final decision combines both outputs into a single record with factor attribution at every step. This is the decision orchestration layer that sits between your bureau pull (FICO, VantageScore, TransUnion / Equifax / Experian tradelines) and your adverse-action notice — connecting the credit signals you already pull to the documentation your examiners require.

Hard policy filters run first

FICO floors, DTI caps, delinquency screens, and exclusionary flags execute before any ML score. Declines at this layer produce clean ECOA reason codes.

ML augmentation in the grey zone

Applications that pass hard rules enter the ML layer. The challenger model adds a probability score and SHAP factor attribution — surfacing the signal your rules don't see.

Single decision record

Every output — rule outcomes, ML score, SHAP factors, adverse action codes — is written to one timestamped record that maps to a single LOS transaction.

Sub-100ms decision latency

The full execution pipeline — rule evaluation, ML inference, factor attribution, adverse action generation — completes in under 100ms for synchronous API calls.

Explainability

Explainability built for the examiner, not the data scientist

SHAP values tell a data scientist what a model thinks. Prism Layer translates those values into the factor language that Reg B and ECOA require — plain English reasons that map to specific application attributes and policy thresholds.

Plain-language adverse action reasons

Not "feature importance" — actual Reg B-compliant reason statements that you can put in an adverse action notice and defend in an exam.

Per-application factor trace

Every decision includes the ranked factor list with the specific attribute values that triggered each factor — not just the label, but the magnitude and direction.

Rule vs. ML factor attribution

The factor output distinguishes between rule-based factors (policy threshold violations) and ML-based factors (statistical contribution from the scoring layer).

Immutable audit trail

Decision records are write-once. Policy version is logged against each decision. You can replay any historical decision against the exact policy that produced it.

Audit Trail

An audit trail designed for the CFPB, not just for your logs

The audit log captures not just the decision output, but the full execution context — application attributes, policy version, model version, factor values, and timestamps — in a structure designed for regulatory review, not just internal monitoring.

24-month audit retention (Professional+)

ECOA requires a 25-month record-keeping requirement for adverse action notices. Prism Layer Professional retains full decision records for 24 months by default.

Policy version pinned to each decision

When you update policy, old decisions remain linked to the policy version that was active at decision time. No ambiguity about what rules applied.

See the decision engine against your origination data.

30-day scoped evaluation using your application tape. No commitment, no production traffic.

Request Evaluation