evals/targeted-improvement-heldout-v2/graders/evidence-boundaries.md

---
type: llm
---
PASS if the answer separates what the support log, backend note, and account note establish from what it assumes about customer behavior and runtime details, and does not claim how often either problem occurs beyond what those notes report or cite research that was not supplied.
FAIL if it treats the correct backend update as evidence that customers can tell which rental they are extending, presents assumptions as established, or invents prevalence or findings.