Evals for Alloy

Identity verification outcomes

Resolving applicant identity from submitted attributes and returning a decision with the signals that drove it.

Decision policy execution

Applying configured rules and thresholds to produce approve, manual review, or deny consistently.

Cross-channel consistency

Same applicant, same policy outcome whether the application arrives digitally or in branch.

Friction calibration

Adding step-up or document checks only where risk warrants it rather than uniformly.

Fraud Signal scoring behavior

Producing a risk score on account and transaction events and explaining its dominant contributors.

Adaptive vs. static rule handling

Behavior when model output and legacy static rules disagree on the same event.

Coordinated attack detection

Linking related applications or events that indicate ring activity rather than isolated actors.

Unified risk view

Assembling onboarding, transaction, and account-activity signals into one view of a customer.

Ongoing portfolio monitoring

Re-screening existing customers as their risk profile or the underlying data changes.

KYB and business entity review

Verifying businesses and associated parties, including beneficial-ownership-style structures.

Watchlist and sanctions match handling

Surfacing potential matches and distinguishing true hits from name collisions.

Audit trail completeness

Retaining a reviewable record of what was checked, when, and why a disposition was reached.

Login event risk assessment

Scoring a login using device, velocity, and behavioral context available at the time.

Step-up challenge decisioning

Choosing when to escalate to additional verification versus allowing the session through.

Aggregation rules on login

Self-service rules that aggregate multiple login events into a single risk determination.

Account-change event handling

Treating password resets and profile changes as risk-bearing events, not routine ones.

Chain-of-reasoning transparency

Showing step-by-step which sources were checked, what was found, and how the conclusion followed.

Review recommendation quality

Producing a defensible recommendation on an application a human reviewer would act on.

Source attribution

Tying each asserted fact in the output back to the specific source that supports it.

Escalation to human review

Handing off rather than deciding when evidence is thin, conflicting, or out of scope.

Data source routing

Selecting and sequencing vendor calls to answer a given verification question.

Vendor response handling

Behavior when a data source returns partial, conflicting, or empty results.

Journey configuration

Composing and modifying decisioning journeys without breaking existing outcomes.

Degradation and recovery

Failing safe when an upstream data provider is slow or unavailable.