Overview
Where the graph disagrees with the score
The case pack
Assessed probability
The agent's own assessment after investigating, not the bank's score.
Patterns identified
Agent activity
Queue
Investigation queue
Every alert in the case pack, ranked by the agent's assessed probability. Open a row to read the evidence behind it.
| Case | Trigger | Card | Bank score | Agent p(fraud) | Confidence | Verdict | Pattern | Exposure | Next action | Approval | Updated |
|---|
Investigate any transaction
Opens a new investigation on a transaction outside the case pack. It joins the queue tagged ad hoc and is kept out of the case-pack figures.
Investigating running
Each line is emitted as it happens: the agent's reasoning steps, and every graph query with the time it really took. Nothing is replayed from the answer file.
Memory
Case memory
What the agent can recall: the investigations shipped with the dataset, and the cases it has closed itself. Memory is context for a new alert, never its verdict.
Model & policy
How a verdict is reached
How the probability is built, how it holds up against closed cases with known outcomes, and the rules that decide what the agent may do on its own.