FinAgent RiskOps / Interactive lab
Trace the decision before the answer.
See which source supports an answer, which tool is permitted, and why a workflow stops or asks for review.
Interactive reference model
Follow a request through fixture retrieval, tool permissions, evidence checks, and an answer, review, or stop decision.
Choose a scenario and run the workflow.
Workflow trace
Result
Evidence used
Runs entirely in this browser tab with synthetic data. State resets on reload. No model API, real retrieval service, or payment system is connected. The injection example is a known fixture, not a general injection detector.
Try these sequences
- Run the known-evidence scenario. The result cites a synthetic order and preserves its unknown delivery date.
- Select the high-impact request. A synthetic policy routes it to human review without executing a refund.
- Choose submit-payment as the requested tool. The allowlist stops every scenario before tool execution.
- Try the injected source, missing source, and unavailable dependency. Inspect the precise stop in the workflow trace.
What this model leaves out
All requests, sources, and tool responses are deterministic fixtures. There is no language model, vector search, external account, or human approval backend. The known injection fixture demonstrates a source-authority boundary; it does not measure real-world injection detection. A production implementation would also need authenticated tool access, tenant isolation, and evaluation against unseen adversarial inputs.