A fraud agent trained
to make better calls.
A focused fraud workload shows how post-training an open-weight model can improve detection and reduce response time.
Higher F1 score
Compared with Opus 4.5 on the reported fraud benchmark.
Lower latency
Reported reduction in online end-to-end inference latency.
Model parameters
A post-trained Qwen3.5-9B open-weight model.
Fraud decisions
need domain context.
General model rankings do not settle the question of performance on a specific fraud task.
The published research describes an Onramp fraud agent, domain-specific evaluation and post-training of an open-weight model. It reports improved F1 and lower online latency.
The lesson for a new engagement: establish verified outcomes and a reproducible benchmark, then measure the specialist model against the actual operational baseline.
| Workflow | Onramp fraud detection |
|---|---|
| Training target | Domain-specific fraud decisions |
| Model | Post-trained open-weight model |
| Evaluation | Verified outcomes and fraud metrics |
| Operational focus | Detection quality and inference latency |
Use the pattern.
Validate the outcome.
We would scope the training data, select a licensed base model, establish held-out evaluations and agree on rollout gates with your experts. Your results would be measured independently.
Discuss a similar workflowStart with the decision
Use a narrow risk workflow with clear escalation boundaries.
Score what matters
Measure missed fraud, false positives and the cost of human review.
Keep experts in the loop
Route uncertain or high-impact decisions to accountable reviewers.
Metrics are the original authors’ reported, workload-specific findings. The baselines differ across metrics; they are not a universal comparison of specialist and general models. Names are redacted in this presentation while original sources remain available for verification.
Find your
specialist advantage.
Explore the economics for your company.
Calculate your value