Evaluate LLM agent responses against rubrics

Write and run ADK-based eval rubrics with LLM-as-Judge patterns

Validating agent responses against business rules without manual test review.

• · Test case (input/expected pair)
• · Rubric definition

• · Eval metric scores 0.0–1.0

• · Google ADK AgentEvaluator
• · gemini-3.1-flash-lite

• · Test data JSON formatted
• · Rubric ID unique

• · Rubric threshold too strict
• · Judge model out of quota

• · Rubric-based criterion with threshold
• · 3-sample judge mode