Analyze the given image for these distortion classes: {allowed_distortions}.
If a distortion is present, classify it using one of these severity levels: {allowed_severities}.
Return a JSON object with a "findings" array.
For now, return at most one finding. Each finding must contain:
- "label": one allowed distortion class;
- "severity": one allowed severity level;
- "confidence": a float between 0 and 1.
If no distortion is present, return {{"findings": []}}.
