Graph Evolve Cookbooks

End-to-end examples for training and serving graphs with Graph Evolve.

These cookbooks show how to use Graph Evolve on real tasks. Each example demonstrates a different graph type and use case.

Featured Examples

• Banking77 Classification: Customer messages in, intent labels out. Train a policy graph for 77-class classification without hand-authored rules. See the Banking77 cookbook.
• Crafter Verifier: Agent traces in, calibrated scores out. Train a verifier graph to evaluate agent performance using V3 traces. See the Crafter Verifier cookbook.

All Included Examples

Cookbook	Graph Type	Task
Banking77 Classification	Policy	Intent classification (77 classes)
Crafter Verifier	Verifier	Custom verifier for agent traces

Policy vs Verifier

Policy graphs solve tasks - they take inputs and produce outputs:

• Banking77: {customer_message} → {intent}

Verifier graphs evaluate results - they take traces and produce scores:

• Crafter: {trace} → {score, reasoning}

Each Cookbook Includes

• A GraphEvolveTaskSet JSON dataset
• Runnable training code
• Inference examples
• Notes on verifier configuration

Getting Started

1. Pick the cookbook matching your use case
2. Copy the dataset format
3. Adapt to your domain
4. Train with Graph Evolve

Need help? Send us your dataset and we’ll walk you through it.

---

Learn more about the Graphs API in our introductory blog post.