The Confidence Judge Node is a fast, cheap pre-flight quality gate. It scores the candidate answer on a described rubric using the Jev decision model (TypeSafe System One) and emits one of three verdicts — pass, regenerate, or escalate — using the same edge vocabulary as the LLM Judge Node.
How it works
- The node sends the user question plus candidate answer to Jev as a single Score question over 2-10 described levels.
- Jev returns a probability-weighted score (0..n_levels-1) and a confidence value.
- The score is normalized to 0..1 and compared against your thresholds:
>= pass_threshold→ pass<= regenerate_threshold→ regenerate- anything between, or confidence below
min_confidence, or the service being unavailable → escalate
Parameters
- rubric (string): The scoring question (e.g. "How well does this answer address the user's question, correctly and completely?").
- levels (array, 2-10, ordered low→high): Level descriptions. Describe situations ("Fully addresses the question, correct and complete"), not degrees ("good").
- pass_threshold (number 0-1, default 0.8): Normalized score at or above which the answer passes.
- regenerate_threshold (number 0-1, default 0.4): Normalized score at or below which the answer should be regenerated.
- min_confidence (number 0-1, default 0.35): Jev confidence below which the node escalates instead of deciding.
- evaluate_key (string, default "candidate_answer"): Which state field to judge (falls back to final_answer).
Output Handles
pass— continue downstream.regenerate— loop back to the producer node to try again.route:escalate— wire to an LLM Judge Node for the full rubric audit.
State
Writes evaluation_verdict, evaluation_score (0-1), and evaluation_source — the same contract the LLM Judge Node uses, so the existing retry/pass edges work unchanged.
Use Cases
- Two-tier judging: fast Jev gate in front, LLM Judge only on ambiguous answers — near-zero judge cost on easy turns.
- Reflection loops: actor → judge → regenerate until pass.
- Escalation pressure valve: route low-confidence cases to a human review card.
Tips
- Escalate is the safe failure mode: an unavailable decision service escalates rather than silently passing a bad answer.
- Keep each rubric to one dimension; split complex judgments into multiple Scorer Nodes and combine in a Condition Node.