A VLM reads an invoice total from a low-resolution screenshot before payment. What is the…

A VLM reads an invoice total from a low-resolution screenshot before payment. What is the safest design?

1. Treat the extracted value as uncertain, verify it against a reliable source or dedicated OCR, and require approval before payment
2. Trust the value whenever the VLM returns fluent explanatory text
3. Increase temperature until several readings agree, then pay automatically