evals/behavior-confidence-heldout/graders/vocabulary.md

---
type: llm
arm: with-only
---
PASS if the answer uses the plugin's four distinct interaction-level Behavior Confidence inputs—outcome predictability, state feedback, cross-context consistency, and failure recovery—and does not copy UI or Flow criterion scores into that dimension.
FAIL if it substitutes a screen or workflow score for Behavior Confidence, omits the interaction criteria, or claims unobserved criteria were evaluated.