judgment-base-agent

The README opens with

Calibrated System One Judgment (Choice, Score, Noul) for Google ADK 2.0 Workflows & Composite Agents

When building AI agents in Google ADK 2.0, standard LLM prompts struggle with control flow: 1. Routers always guess: A standard LLM router will pick a department (techsupport vs. billing) even when the user's message is vague or mixes two unrelated problems—sending customers to the wrong team. 2. Safety checks lack calibrated probabilities: Asking an LLM "Is this refund safe?" returns text instead of calibrated probabilities (0.00–1.00) and ordered risk scores (0.00–3.00) that deterministic Python if/else code can enforce. 3. Batch triage is slow (N calls instead of 1): Evaluating 5 reviews or tickets in a for loop makes 5 separate LLM calls...