Cleanup

The model receives the prompt (and optional system message). The run uses scorer contains_any with the JSON configuration below. Pass/fail and partial credit are determined entirely by that scorer against the model output; no human grading.

Refactor without changing behavior. Return only Python code.

def safe_city_20(user):
    if user is None:
        return ''
    if user.get('address') is None:
        return ''
    if user.get('address').get('city') is None:
        return ''
    return user.get('address').get('city').strip().title()

{
  "expected_contains": [
    "if not user",
    "address",
    "strip().title()"
  ]
}