Ground Truth & Reference Evaluation

Pass an optional ground_truth (string or structured object) on evaluate. The engine injects it before scoring so the critic can compare the assistant response against a reference answer or expected state.

result = client.evaluate(
    messages=[...],
    criteria=["Reward responses that match the reference refund window."],
    ground_truth="Refunds are available within 30 days of purchase.",
)

In Studio Visual Builder, expand + Ground Truth / Reference. Pair with criteria.ground_truth style criteria when the reference is authoritative.