Recipe · decision_risk
Decision / risk classifier task model
Train a scoped decision/risk task model on your outcomes and put it under your agent — beat frontier cost and flakiness on one labeled job.
- Shapedecision_riskBinary or multiclass decisions with historical ground-truth outcomes. Core task-model loop.
- BaseQwen2.5-1.5B-InstructDefault open-weight base · train mode lora_sft · min 500 rows
- Metricsprecision · recall · f1 · latency_p50_msEval against your held-out outcomes before you swap the agent call
FAQ
What jobs fit this recipe?
Binary or multiclass decisions with ground-truth labels — lending risk, marketplace review, insurance triage, ops routing.
How do I know it beats frontier?
Run an eval on held-out outcomes before and after training. Pilot credits exist to prove the gap on your labels.
Is inference basically free?
Relative to frontier API prices on high-volume loops, yes — you serve a small open-weight model you control. Use the cost calculator for your call volume.
Put a task model under the agent
Pilot is free to start. Prove beat-frontier on your labels.