Recipe · fraud_risk
Fraud / payments risk task model
Replace frontier LLM fraud calls with a small open-weight task model trained on your payment outcomes — cheaper, measurable, callable by your agent.
- Shapedecision_riskPayments fraud task model with dollar-weighted recall — Coinbase-style narrow task pack.
- BaseQwen2.5-1.5B-InstructDefault open-weight base · train mode grpo_rlvr · min 1,000 rows
- Metricsprecision · recall · f1 · dollar_weighted_recall · latency_p50_msEval against your held-out outcomes before you swap the agent call
FAQ
Does this replace my fraud agent?
No. Your agent or harness stays. We train a task model — small open weights on your labeled outcomes — that the agent calls for the high-volume decision.
What labels do I need?
Historical approve/review/block (or equivalent) outcomes with features. Dollar amounts help when you care about dollar-weighted recall.
Where does training run?
Pilot on Shared cloud, Build on your VPC (BYOC), or Scale on confidential GPUs with attestation when data path is the blocker.
Put a task model under the agent
Pilot is free to start. Prove beat-frontier on your labels.