Which evaluation-set design gives the strongest release evidence for a customer-support a…

Which evaluation-set design gives the strongest release evidence for a customer-support assistant?

1. A versioned mix of sampled task categories, costly incidents, boundary cases, adversarial cases, and predefined user slices
2. The ten demonstrations used while editing the system prompt
3. A large synthetic set generated by the same prompt and model under evaluation
4. Only the most common happy-path question, repeated with different order IDs