Use case · fraud / payments
Put a fraud task model under the payments agent
Measure dollar-weighted recall on your outcomes — not vibes. Keep the harness; swap the weak frontier call for a small open-weight model trained on your labels.
- PainFrontier miss rate × $ exposureOr API cost on every txn path your agent already reviews.
- BuySame agent, better callRecipe: fraud / payments risk. Eval first — see methodology.
- ProofHeld-out $ metrics + costUse the cost calculator for volume; Pilot for beat-frontier on your labels.
FAQ
Does this replace my fraud agent?
No. Keep the agent and tools. Train a small open-weight task model on your chargebacks or approve/review outcomes and point the high-volume call at it.
What should I measure?
Dollar-weighted recall (and precision/F1) on a held-out set — not vibes. Ship only if the task model beats frontier on the metrics that map to $ exposure.
Where does training run?
Pilot on Shared, Build on your VPC when rows can’t leave, Scale on confidential GPUs when attestation is required.
Start from the fraud recipe
Pilot is free with training credits on Shared.