Evals for Oscilar

Prompt-to-workflow generation

Turning a natural-language risk intent into an ordered workflow with inputs, rulesets, models, and terminal decisions.

Ruleset and threshold composition

Assembling checks such as device+behavioral intelligence and compliance screening against stated thresholds.

No-code editing by non-engineers

Modifying an existing workflow without code or engineering queue dependency, per the Conta Azul autonomy story.

Pre-deployment testing

Exercising a drafted workflow before it goes live rather than deploying untested logic.

SQL-free performance readout

Surfacing real-time workflow and rule performance to a non-technical policy owner.

Tuning recommendations

Proposing concrete rule or threshold changes with the performance evidence behind them.

False-positive tradeoffs

Making the precision/recall consequence of a proposed change explicit before it ships.

Policy deployment cadence

Moving an approved change from recommendation to deployed policy.

Alert triage and disposition drafting

Summarizing an alert and proposing a disposition an analyst can accept, edit, or reject.

Case context assembly

Pulling the relevant customer, transaction, and prior-alert context into one case view.

Escalate, freeze, or approve decisions

Selecting and justifying the terminal action on a case.

Analyst override handling

Recording and respecting a human override of an agent recommendation.

Cross-domain context sharing

An agent reasoning over what fraud found, what AML escalated, and what policy changed, not just its own silo.

Routing and handoff between agents

Passing a case to the right specialist agent and carrying the prior findings with it.

Agent scope boundaries

A specialized agent declining work outside its remit rather than improvising across domains.

Orchestration transparency

Making visible which agents contributed to a decision and why.

Transaction and P2P transfer scoring

Producing a risk decision on an inbound transfer within a real-time decisioning path.

Device and behavioral signals

Detecting patterns such as one device ID used to create multiple accounts.

Mule network detection

Identifying coordinated mule and network behavior rather than isolated bad transactions.

Cross-border and market variation

Handling differing fraud pressure across corridors and markets such as Brazil.

Compliance and sanctions screening

Executing screening checks against established thresholds and escalating hits.

314(b) information sharing limits

Distinguishing shareable fraud facts from SAR confidentiality under the FinCEN safe harbor.

Onboarding and AML case escalation

Routing high-risk onboarding and AML patterns to the correct review path.

Decision explanation

Explaining a credit or risk decision in terms a reviewer or regulator can follow.