Module judge 

LLM-as-judge — the shared async evaluation primitive behind the LLM-backed G:: guards and E:: criteria.

Mirrors ADK Python’s final_response_match_v2 / safety-evaluator approach: a judge model is prompted to render a structured verdict, and the verdict label is parsed back out of the model’s reply (robust to surrounding prose).

Structs

JudgeVerdict

The judge’s verdict (distinct from the flow’s JudgeVerdict on tool calls).

LlmJudge

A configured LLM-as-judge. Cheap to clone (Arc inside).

Functions

parse_verdict

Parse a verdict from the judge model’s reply. Tolerant of extra prose around the JSON: it scans for the violation field’s boolean and the reason string, falling back to common labels (invalid, unsafe).

render_contents

Render conversation history into a plain-text block for a judge prompt, keeping role labels so the judge can reason about grounding.