Preprint reports improved payment fraud detection without added inference cost

Researchers report in a preprint that a payment fraud policy improved dollar-weighted precision without adding inference cost. On a held-out four-month test, it beat a prior agent trained only on labeled transactions by +4.7, +2.4 and +3.0 points at 80%, 85% and 90% recall, respectively. The method uses a language model to write a reward function that evaluates both labeled transactions and previously blocked transactions lacking fraud labels. For the latter, it contrasts a newer risk signal with the scores behind the original blocks. The resulting policy retained the prior agent's architecture and beat the production baseline by +6.8 points on average; the reported results come from a held-out test.