assay is a fraud investigator that runs on a TigerGraph graph. This page replays every query it ran on each case, in order. The graph grows as results come back, and each finding moves a probability that is calibrated against the bank's 5,565 closed cases. Nothing here is staged: the traces come from rerunning the agent and match the committed answer files.
Every investigation from the exam period. Each dot is assay's final probability, and the brass tick is the bank's risk score for the same transaction. The line between them is the disagreement. Hollow dots are the 60 alerts the agent raised on its own: structuring bursts, device rings, and fraud the bank's score missed. Click a dot to replay that investigation.
The model learns only from the bank's closed cases for July to September. It is then scored on October, which it never saw, and calibrated with cross-fitted isotonic regression.