New judge

A judge is an authored, versioned asset — rubric + model + prompt. It ships to shadow and can never gate until calibration and an SME sign-off (§32).

Template

start from an AAG rubric pack — §35 inherit + override

aag/policy-answer picked

grounded, current, correctly-scoped policy answers — the closest fit for benefits Q&A

aag/refusal

refusal correctness + helpful redirect

aag/task-completion

transactional flows — goal + side effects

judge.yaml so far

Generated live — it grows as each step is completed.

judges/benefits-answer-judge.yaml

judge: benefits-answer-judge
version: v1
owner: benefits-team
template: aag/policy-answer   # §35 inherit + override