Yes, Opus could check the image to see if it matched the prompt, but I adviced the model to stop and ask the human for a better check and a description of what the cause of the corrupted image could be. But the fact it could catch obvious regressions was good.